Manufacturing (Non-Automotive)Readiness & Transformation Roadmap

AI readiness for ESG in manufacturing: the evidence chain behind every disclosed number

AI readiness for ESG in manufacturing is the condition in which a plant's environmental, social and governance data is metered, allocated to production and traceable to source — so a model can join the calculation path without lowering the assurance grade of the disclosed number. It is an evidence problem before it is a modelling one.

Manufacturing (Non-Automotive) · Readiness & Transformation Roadmap

Atomic Loops EngineeringIndustrial AI practiceReviewed by Atomic Loops Solutions ArchitecturePublished 20 August 202663 min read

Key takeaways

An ESG number is worth what its weakest input can survive. Grade every input A to D — metered primary, activity data times a published factor, supplier-specific secondary, or modelled proxy — and the readiness question stops being philosophical: what share of the disclosed figure is grade A, and can you prove it?
AI's legitimate first job in manufacturing ESG is to move data up the grade ladder and to label what stays modelled. Extraction from supplier documents, anomaly detection on meter series, factor matching and bounded gap-filling all do that. Generative drafting of the disclosure narrative does not, and is the stage-2 distraction.
The binding constraint in almost every plant is allocation, not measurement. Sub-meters usually exist; the join between the 15-minute interval series and the MES production order does not, so no figure can be cut by line, order or SKU — which is exactly the resolution CBAM, ESPR and customer PCF requests demand.
Assurance readiness is measurable as trace time: pick a disclosed figure at random and time how long it takes to reach the meter reading or supplier document behind it. Weeks means the number is rebuilt rather than traced. Minutes means you have a lineage record, and limited assurance becomes a sampling exercise.
Freeze the methodology and the factor set per reporting period, and treat any change as a labelled restatement. Silently recomputing history when a factor library updates is the single most common way a manufacturer turns a defensible number into an audit finding.

Abbreviations used on this page

CSRD: Corporate Sustainability Reporting Directive (EU)
ESRS: European Sustainability Reporting Standards (E1 is the climate standard)
CBAM: Carbon Border Adjustment Mechanism (EU import carbon regime)
ESPR: Ecodesign for Sustainable Products Regulation (EU)
DPP: Digital Product Passport (the data carrier ESPR introduces)
PCF: Product carbon footprint — kg CO₂e per unit of product
LCA: Life-cycle assessment
EPD: Environmental product declaration (a verified supplier LCA summary)
MES: Manufacturing execution system — the production-order system of record
EMS: Energy management system (the ISO 50001 construct and its software)
ERP: Enterprise resource planning
BOM: Bill of materials

What AI readiness for ESG means in a manufacturing plant

A definition, the four evidence grades every ESG input falls into, and the chain a number has to survive between a meter and an assurance opinion.

AI readiness for ESG in manufacturing is the condition in which environmental, social and governance data is metered, allocated to production and traceable to source, so that a model can join the calculation path without lowering the assurance grade of the disclosed number. It is deliberately a narrower question than general AI readiness. The models involved are unexceptional — extraction, classification, anomaly detection, gap-filling — and the difficulty sits entirely in what they are being asked to touch: figures that will be published, verified by a third party, quoted to customers and, for some product categories, declared to a customs authority.

The organising idea on this page is the evidence grade. Every input to an ESG figure sits in one of four grades, from a metered primary reading down to a modelled proxy, and a composite figure is worth what its weakest material input can survive. That single framing settles most of the arguments a readiness review runs into. It tells you which data to fix first, what a model may legitimately do to each class of input, and what must never happen — a modelled value presented as if it had been measured. It also converts an abstract readiness question into an arithmetic one: what share of this figure is grade A, and can you prove it?

Grade	What it is	Typical manufacturing source	What it survives	What AI may legitimately do
A · Metered primary	A physical reading tied to a timestamp and an asset	Sub-meters, flow meters, weighbridge and waste tickets, DCS and PLC tags	Limited and reasonable assurance; CBAM verification; a customer's own audit	Detect faults, outliers and drift; reconcile meters against each other — never generate the value
B · Activity × published factor	A measured mass, volume or run time multiplied by a documented factor	MES consumption records, BOM quantities, ERP goods receipts	Limited assurance where the factor is versioned, dated and sourced	Match a material to the closest available factor and keep the register current — a human confirms each match
C · Supplier-specific secondary	A supplier's own declared figure for a purchased item	EPDs, supplier PCF sheets, questionnaire returns, contractual disclosures	Limited assurance for Scope 3 category 1 where the document is retained and current	Extract from PDFs, normalise units, flag boundary mismatches — a human verifies before the value is used
D · Modelled or proxy	An estimate: spend-based factors, sector averages, imputed gaps	Purchase ledger, industry databases, interpolation across missing periods	Disclosure only, and only where it is explicitly labelled as estimated	Fill gaps with an explicit uncertainty band and a label — never present the result as measured

The four evidence grades. Grade is a property of an input, not of a report: a single disclosed figure normally mixes all four, and the honest question is the mix, not the label on the cover. The right-hand column is the boundary of what a model may legitimately do.

Grades are not a maturity score in disguise. A mature manufacturer still discloses grade D figures — a small, immaterial purchased category is not worth a supplier engagement programme — and an immature one occasionally has excellent grade A metering on the one utility that dominates its cost base. What changes with maturity is whether the mix is known, whether it is disclosed, and whether the grade travels with the number when it moves between systems. That last property is what makes the difference between a chain a model can safely join and one it cannot, because a model that cannot see the grade of its inputs will happily average a meter reading with a spend-based estimate and produce something that looks like a measurement.

The ESG evidence chain, and where a model is allowed to touch it

Three lanes, one destination. The top lane is where primary data is produced and — in most plants — where it stops. The middle lane is where the largest share of the footprint comes from and where the grade is lowest. The bottom lane is where the methodology is set and where the whole chain is finally tested. The dashed red path is the failure mode this page exists to prevent.

Data & feeds
System-of-record action
AI / model
Human in the loop
Where value leaks

The process, in words

Site lane: meters and process instruments produce interval readings, a historian or energy management system retains them raw, and the allocation step joins that series to MES production orders using written rules for shared utilities. What comes out is kWh and kg CO₂e per production order — grade A data at the resolution CBAM, product passports and customer requests all ask for. In most plants this lane stops at the historian.
Chain lane: purchase and BOM records say who supplied what; supplier documents — EPDs, PCF sheets, questionnaire returns — arrive as PDFs in a dozen layouts. An extraction model reads them and proposes structured entries, a named human verifies each one, and the result lands in a factor register where every factor carries a grade, a source and a date. Gaps that remain are filled with an explicitly labelled estimate and an uncertainty band.
Methodology lane: the boundary, the allocation rules and the factor set are frozen at the start of the reporting period and constrain both other lanes — the allocation rule governs the site join, the gap-fill bounds govern what the model is permitted to invent. The disclosed figure carries the grade mix forward, and the assurance opinion is formed by sampling it back to raw.
The failure path, dashed in red: a modelled value loses its label somewhere between the model and the disclosure, is aggregated with metered data as if it were equivalent, and is published as measured. Nothing about that figure is wrong until someone asks how it was produced — at which point it cannot be traced, and it is restated.

Step-by-step insights

Meters and instruments — retention is the decision that matters: The metering itself is rarely the constraint; retention is. Many energy management systems are configured to keep raw interval data for months and monthly rollups forever, because that is what a utility-cost use case needs. The moment you want to allocate to production orders, rollups are useless: a production order lasts hours, not months, and no monthly total can be decomposed into one. Check the retention policy before the meter list. A plant with fewer meters and full interval history is closer to grade A allocation than a heavily instrumented plant that throws the detail away.
The allocation join — one rule per shared utility, written down: This is the single highest-leverage piece of work on the page and it is mostly adjudication rather than engineering. Which meter serves which line. How compressed air is apportioned across three lines sharing a compressor — run hours, nameplate, or measured flow. Where changeover and clean-down energy goes, given it belongs to the product but sits in no production order. Whether waste is allocated by mass or by cost. Each rule needs to be decided once, written down, and applied consistently, because an assurer will ask and a customer comparing two suppliers' footprints will notice if the rules differ. The join itself lives at the production-order level, which is ISA-95 level 3 — the layer the MES already owns.
Supplier documents — extraction is where AI genuinely pays: A mid-size manufacturer with a few thousand purchased line items may receive supplier environmental data in twenty formats: EPDs following different product category rules, PCF sheets with different boundaries, questionnaire returns in spreadsheets, and a long tail of PDFs. Extracting these by hand is the reason Scope 3 stays grade D. A model that reads the document, proposes a structured entry and flags a boundary mismatch — cradle-to-gate versus cradle-to-grave, mass versus unit basis — turns weeks of analyst time into a review queue. The control that makes it admissible is the verification step: a named human confirms each entry, and the original document is retained and linked, so the assurer samples the document rather than the model.
Gap-fill — bounded, labelled, and never silently promoted: Every real inventory has holes: a supplier that will not respond, a month of missing sub-meter data, a new material with no factor. Filling them is legitimate and expected — what matters is that the fill is bounded by a rule set in the methodology register, carries an uncertainty band, and keeps its grade D label everywhere downstream. The failure is not the estimate; it is the estimate that arrives at the disclosure indistinguishable from a meter reading. Set a ceiling on the modelled share of any material figure and alert when it is approached, because gap-fill expands quietly: each individually reasonable fill moves the mix one notch, and nobody is watching the aggregate.
The methodology register — freeze the period, log the change: The methodology register holds the boundary, the consolidation approach, the allocation rules and the factor set, and its defining property is that it is frozen for the duration of a reporting period. Factor libraries update; grid emission factors are restated by their publishers; a supplier issues a revised EPD. None of those may change a period already in flight. New values apply prospectively, and if a change is material enough to warrant restating history, that restatement is explicit and logged. Manufacturers who let libraries update in place discover it the following year, when last year's published figure no longer reproduces and no one can explain the delta.
The assurance opinion — a sample, not a rebuild: The whole chain exists to make one interaction cheap. An assurer names a figure and asks how it was produced; either the answer is a lineage record returned in minutes, or it is three weeks of reconstruction. That difference is what separates a stage-4 manufacturer from a stage-3 one, and it is directly measurable as trace time. It is also the reason to build the lineage store while building the pipeline: retrofitting lineage onto a calculation that already runs means reconstructing provenance for figures whose inputs have since changed, which is strictly harder than recording it as you go.

Real today, and in production
Extraction of supplier EPDs and PCF sheets into a structured, human-verified register. Anomaly and drift detection on meter series, which catches a stuck sensor in a day rather than at quarter end. Matching BOM materials to the nearest available factor with a confidence score. Bounded gap-filling with an explicit uncertainty band. Drafting disclosure narrative from a locked, already-computed dataset. Every one of these is ordinary engineering with an ordinary control around it.
Claimed, and true only under conditions
Real-time Scope 3, continuous product footprints and automated supplier engagement are all achievable — where suppliers actually exchange primary data, which for most manufacturers is a minority of spend today. The honest version of the claim is that the machinery works and the inputs do not exist yet. Buy it as a destination, sequence it behind a supplier data programme, and do not let a demonstration on a curated dataset set the expectation for your own long tail.
Not credible, and a live compliance risk
A model that infers your Scope 3 from your purchase ledger and presents the output as a measured inventory. Assurance evidence generated rather than retrieved. A product footprint produced without a stated boundary or allocation rule. Anything that removes the grade label from a value on its way to a disclosure. These are not conservative objections about model quality — they are the specific artefacts that turn a reporting exercise into a misstatement, and they are the reason the register and the labelling exist.

What the evidence ladder releases, stage by stage

Value stays close to flat through the first two stages — the data exists but cannot be cut, so nothing outside the reporting function can use it — and inflects at allocation, when a figure first becomes something a customer, a planner or a designer can act on. This is why programmes measured in meters installed rather than figures allocated report activity without results.

Usable, defensible ESG value released by stage

Stage 1 · Invoice-grade — 22% of operators. ESG numbers are assembled once a year from utility invoices, the purchase ledger and supplier emails, at group or site level, in spreadsheets.
Stage 2 · Metered — 37% of operators. Site-level primary data exists — sub-meters, an EMS, weighbridge tickets — but it lives in the energy team's tools and is reconciled to the disclosure by hand.
Stage 3 · Allocated — 27% of operators. Environmental data is joined to production data, so energy, emissions, water and waste can be cut by line, production order and SKU, and rebuilt from raw.
Stage 4 · Assured — 11% of operators. Every disclosed figure carries its evidence grade, factor version and lineage, so assurance is a sampling exercise — and any model in the calculation path is registered, bounded and disclosed.
Stage 5 · Steering — 3% of operators. The same assured dataset drives operating decisions — scheduling against carbon intensity, setpoints, supplier awards, product design — with models recommending and, inside stated bounds, acting.

Curve shape: logistic, plotted from the stage data above. Distribution: Shape consistent with acatech's Industrie 4.0 Maturity Index stage progression.

The five stages of an ESG evidence chain

For each stage: what it looks like on a real plant, the diagnostic signals a reviewer can check in an afternoon, the anti-pattern that traps manufacturers there, and what leaving costs.

The five stages describe how far an ESG number has travelled from an invoice to an operating decision, and each one is defined by what the number can survive rather than by how much technology sits behind it. The ladder runs invoice-grade, metered, allocated, assured, steering. It is written for practitioners: the hallmarks are observable conditions, the diagnostic signals are checks you can run against your own systems this week, and the anti-pattern is the specific mistake most often made trying to leave that stage.

Stage	Smallest unit available	Typical grade mix	What the figure survives
1 · Invoice-grade	Site, financial year	Mostly D, some B	An internal report; a first-year assurance engagement with findings
2 · Metered	Site, month	A for site energy, D for purchased inputs	An energy-cost conversation; still no answer to a customer's product question
3 · Allocated	Line, shift, production order	A for energy, B–C for materials, D labelled	A customer PCF request, with follow-up questions answered
4 · Assured	Production order, with lineage	Known and disclosed mix	A sampling assurance engagement; a CBAM verification; a restatement handled cleanly
5 · Steering	Batch or decision, near real time	Known mix, reconciled between views	All of the above, plus use as a live constraint without corrupting the disclosed copy

What a figure can survive at each stage. The right-hand column is the practical test: hand the figure to the audience named and see what happens.

Select a stage

Every stage's full detail is present in the page source — the selector only changes which panel is visible, so nothing here depends on JavaScript to exist.

Stage 1

Invoice-grade

22% of operators sit here

ESG numbers are assembled once a year from utility invoices, the purchase ledger and supplier emails, at group or site level, in spreadsheets.

Stage 1 is not a failure of intent. Most manufacturers arrive here because the first reporting obligation was met the way every first obligation is met — by a small team, in a spreadsheet, against a deadline. The numbers are usually defensible in the narrow sense that each one traces to an invoice. What is missing is resolution and repeatability: the figure exists for the group and the year, and reproducing it next year means repeating the same manual assembly rather than re-running anything.

The tell is the unit of account. At stage 1 the atom is a bill. Electricity is what the utility charged, gas is what the meter reader recorded at the boundary, and the emissions attached to purchased materials are the ledger spend multiplied by a sector-average factor. None of those can be divided by a product, a line or a shift, because the underlying record was never associated with production in the first place. Every downstream question — what does this SKU emit, which line is the intensity outlier, what did the efficiency project actually save — is unanswerable by construction.

This is a cheap stage to leave and an expensive one to stay in, and the expense is not the reporting effort. It is that the ESG number never becomes an operational number, so it never earns operational attention, so the metering and allocation work that would fix it never gets funded. The loop closes on itself. Manufacturers who sit here for three reporting cycles usually find that the workbook has grown, the assurance scope has widened, and the underlying data is exactly where it was.

In practice

The February scramble

A building-products manufacturer with eleven plants closes its sustainability data every February. Two analysts email each plant controller for twelve utility invoices, a waste contractor summary and a water bill, retype the figures into a master workbook, and apply a factor set downloaded the previous spring. Scope 3 is the purchase ledger sliced by commodity code and multiplied by spend-based factors. The pack is good, the assurer's findings are manageable, and not one figure in it can be attributed to a production line.

What it looks like

Scope 1 and 2 come from utility invoices; Scope 3 from spend times an average factor
The smallest unit anyone can produce is a site and a financial year
The methodology lives in the workbook and in the head of whoever built it
No model touches the numbers, because there is no dataset for one to touch

Diagnostic signals you can check this week

Ask for one plant's figure for one month. If the answer requires the analyst who built the workbook, you are here
Ask which factor set version was used last year and this year. A screenshot or a filename is the stage-1 answer
Count the spreadsheets between a meter and a disclosed figure. More than two is diagnostic
Ask what operating decision would change if the number moved 10%. If none, the number is a reporting artefact

Anti-pattern · Buying an ESG platform to fix a data problem

The instinctive move is to procure a reporting platform, on the theory that the software will supply the rigour. It cannot: the platform ingests exactly the invoices the spreadsheet ingested, and adds a workflow, an audit log of who typed what, and a licence fee. What it does not add is a meter, a production-order join, or a graded factor register. Buy the platform after the evidence chain exists and it will save real effort; buy it before, and you have paid to industrialise the estimate.

What holds you here

There is no primary data and no production context, so every figure is an estimate assembled by hand and nothing can be automated without automating the estimate.

Highest-leverage next move

Pick one site and one utility. Get interval metering into a system that keeps history, and freeze a written methodology for that one number.

Cost of leaving

Effort: 3–6 months
Team: One plant energy engineer and one reporting controller, part-time
Risk: Low — nothing currently disclosed depends on the new work yet
To next stage: 3–6 months

If this is you, the next step is

A two-week review: which meters exist, which are missing, what one line would cost to instrument.

Scope a first metered site

Stage 2

Metered

37% of operators sit here

Site-level primary data exists — sub-meters, an EMS, weighbridge tickets — but it lives in the energy team's tools and is reconciled to the disclosure by hand.

Stage 2 is the most populated stage and the most misread. There is genuine primary data: half-hourly electricity, gas by meter, often steam and compressed air, sometimes water and effluent, retained in an energy management system bought for ISO 50001 or for a utility-cost programme. The plant energy engineer can tell you the baseload, the weekend draw and which compressor is drifting. By any reasonable definition the measurement problem is solved.

What is not solved is the join. The interval series is indexed by meter and timestamp; production is indexed by order, product and shift in the MES. Nobody has written the rule that connects them, so the ESG number continues to be built from monthly totals — the one shape the interval data can be reduced to without needing production context. The energy team's rich dataset and the reporting team's thin dataset coexist for years, and each team assumes the other has already solved the part it can see.

The organisational consequence matters more than the technical one. Because no figure can be cut by product or order, no commercial conversation can use it. Sales cannot answer a customer's product carbon footprint request, procurement cannot price a low-carbon variant, and the plant cannot rank its lines by intensity. The ESG dataset stays inside the sustainability function, which is why it stays underfunded, which is why the join never gets built.

In practice

Two teams, two datasets, one plant

At a speciality chemicals site the energy manager has four years of half-hourly electricity by substation and can show you the exact hour a dryer was left running over a bank holiday. The group reporting team, three floors up, discloses annual site electricity from the utility portal. When a customer asks for the cradle-to-gate footprint of one product family, neither dataset answers: one has no product, the other has no resolution. The work to connect them is two weeks and has never been anyone's objective.

What it looks like

Interval data from electricity, gas, steam, compressed-air and water meters is collected and retained
The energy team can answer questions the reporting team cannot, and vice versa
Reporting still consumes monthly totals, not the interval series
AI, where present, drafts narrative or classifies invoices — nothing in the calculation path

Diagnostic signals you can check this week

Ask the energy team and the reporting team for the same site's annual electricity. Compare the two numbers and, more revealingly, the two methods
Ask whether any interval series has ever been joined to a production order. Not 'could it' — has it
Check whether the EMS retains raw interval data or only monthly rollups. Rollups cap you here permanently
Ask who would notice if a sub-meter flatlined for a fortnight. Silence is the stage-2 answer

Anti-pattern · Adding more meters instead of using the ones you have

When the ESG number still looks coarse, the reflex is to extend metering — more sub-meters, more points, a bigger capital request. It is the wrong order. An unallocated meter adds a series nobody can attribute; a meter joined to production changes what the business can sell and schedule. Do the join on the coverage you already have, discover which three missing points actually block a product-level figure, and buy those. The metering plan written after the first allocation is a fraction of the one written before it.

What holds you here

Metered data is not joined to production, so no figure can be cut by line, order or product — and every obligation that matters now asks for exactly that cut.

Highest-leverage next move

Join one interval meter to one line's MES production orders and agree the allocation rule for shared utilities. One line, not one site.

Cost of leaving

Effort: 4–8 months
Team: One data engineer, the plant energy engineer, an MES or IT contact
Risk: Medium — the first join changes numbers people have already disclosed
To next stage: 4–8 months

If this is you, the next step is

The stage 2→3 move on a single line, typically inside a quarter.

Join one meter to one line

Stage 3

Allocated

27% of operators sit here

Environmental data is joined to production data, so energy, emissions, water and waste can be cut by line, production order and SKU, and rebuilt from raw.

Stage 3 is the first stage where the ESG dataset can answer a commercial question. Once metered consumption is allocated to production orders, cradle-to-gate footprints become arithmetic rather than a project: the BOM supplies the purchased inputs, the factor register supplies their emissions, the allocated energy supplies the conversion step, and a per-unit figure falls out. The same join answers the intensity question by line and shift, which is the number a plant manager can actually act on.

The work is unglamorous and mostly about rules. Which meter serves which line. How compressed air is apportioned when three lines share a compressor — by run hours, by nameplate, by measured flow. What happens to the energy consumed during changeover and cleaning, which belongs to the product but is not in any production order. Whether waste is allocated by mass or by cost. Every one of these is a decision that has to be written down, defended once, and then applied consistently, because an assurer will ask and a customer comparing two suppliers' footprints will care.

This is also where AI first belongs in the chain, and where its role should be narrowly drawn. Supplier EPDs and PCF sheets arrive as PDFs in a dozen layouts; extraction into a structured factor register with a human verifying each entry is a genuine order-of-magnitude saving. Meter series drift, stick and flatline; anomaly detection catches that before a quarterly review does. Materials on a BOM need matching to the closest available factor; a model proposes, a human confirms. What must not happen is a model producing a value that is then disclosed as if it had been measured.

In practice

The first product footprint that survived a customer

An industrial pump manufacturer joined half-hourly electricity from three cells to MES production orders, apportioned compressed air by measured flow, and matched every line on the BOM of one high-runner family to a factor with a stated source. The resulting cradle-to-gate figure went to a customer who came back with two questions — the allocation rule for the shared paint line, and the date of the aluminium factor. Both had answers. That exchange, not the number itself, is what stage 3 buys.

What it looks like

kWh and kg CO₂e are available per production order, not only per site per month
A written allocation rule exists for shared utilities — compressed air, chilled water, HVAC
Emission factors sit in a register with a source and a date, not in a spreadsheet column
Model-assisted work has started where it belongs: extraction, anomaly detection, factor matching

Diagnostic signals you can check this week

Ask for kWh per unit for one SKU last month. A number in minutes is stage 3; a project plan is stage 2
Ask to see the allocation rule for a shared utility in writing. Verbal consensus is not a rule
Check whether the factor register records a source and a date per factor, and who last changed each one
Ask what happens to a disclosed figure when a factor library updates. 'It just changes' is the finding waiting to happen

Anti-pattern · Chasing resolution before grade

Allocation makes finer cuts possible, and the temptation is to push resolution everywhere — per batch, per machine, per minute — while the inputs are still mostly modelled. A per-batch figure built on a spend-based Scope 3 factor is not more accurate than the annual one; it is the same estimate with more decimal places and more surface area for an assurer to sample. Raise the grade of the largest contributors first, then raise the resolution. Precision that outruns evidence is the fastest route to a restatement.

What holds you here

The chain from raw reading to disclosed figure is reproducible but not evidenced: the trace lives in scripts and people's heads, not in a record an assurer can sample.

Highest-leverage next move

Make lineage a stored artefact. Every disclosed figure carries its grade, its factor version and a pointer to the raw record that produced it.

Cost of leaving

Effort: 6–12 months
Team: Data engineer, plant energy engineer, LCA or sustainability analyst, MES owner
Risk: Medium — the first allocated figures will disagree with previously disclosed ones
To next stage: 6–12 months

If this is you, the next step is

A working session on shared utilities, changeover and waste — the three rules that get argued about.

Design your allocation rules

Stage 4

Assured

11% of operators sit here

Every disclosed figure carries its evidence grade, factor version and lineage, so assurance is a sampling exercise — and any model in the calculation path is registered, bounded and disclosed.

Stage 4 changes what assurance costs and what it means. Below it, an assurance engagement is an evidence hunt: the assurer names a figure, the team goes looking, and three weeks disappear into reconstructing a calculation that was never designed to be reconstructed. At stage 4 the same request is a query. The figure carries a lineage record — which raw readings, which factor version, which allocation rule, which reviewer — and the engagement becomes what it is supposed to be, a sample of a controlled process rather than a rebuild of an uncontrolled one.

The discipline that makes this work is period freezing. The methodology, the boundary and the factor set are fixed at the start of a reporting period and cannot change inside it. When a factor library publishes an update, the new values apply prospectively; if the change is material enough to warrant restating history, that restatement is explicit, labelled and logged. Manufacturers who let a factor library update recompute prior periods in place discover the problem the following year, when last year's disclosed figure no longer reproduces and nobody can say why.

This is also the stage where AI has to be governed rather than merely used. Any model contributing to a disclosed number needs a register entry: what it does, which version, what inputs, who reviews its output, what the override rate is, and how its contribution is labelled in the disclosure. That is not a bureaucratic flourish — it is the same control an assurer applies to any other estimation technique, and it is increasingly what the AI-specific frameworks expect. The manufacturers who find this easy are the ones who wrote the register while building the pipeline rather than in the month before the audit.

In practice

The twenty-figure dry run

Before its first assured reporting cycle, a packaging manufacturer asked its internal audit team to pick twenty disclosed figures at random and trace each to source. Fourteen took under five minutes from the lineage store. Four took a day, all of them Scope 3 items where a supplier document existed as an email attachment rather than a register entry. Two could not be traced at all and were restated. The exercise cost a week and removed every material finding from the external engagement that followed.

What it looks like

A sampled figure traces to a meter reading or supplier document in minutes, on demand
The factor set and methodology are frozen per reporting period, with a change log
A written restatement policy exists and has been used at least once
Model-derived fields are flagged, version-pinned and reviewed, with the override log retained

Diagnostic signals you can check this week

Pick a disclosed figure at random and time the trace to raw. Median trace time is the stage-4 metric
Ask whether last year's figure still reproduces exactly from this year's system. If not, factors are updating in place
Ask to see the model register entry for any AI in the reporting chain, including the override log
Ask when the restatement policy was last used. A policy never exercised is a document, not a control

Anti-pattern · Treating assurance as the destination

Reaching assurance-grade data is a large achievement and a natural place to stop, which is exactly the trap. An assured dataset that changes no operating decision is a cost centre with excellent documentation, defended annually against a finance function that can see its price and not its return. The transition out of stage 4 is not more control — it is putting one assured number into a live decision, with a holdout, so somebody outside the sustainability function has a reason to care whether it is right.

What holds you here

The assured dataset is still an output. Nothing operational changes because of it, so the apparatus is a cost centre defended annually rather than a capability anyone fights for.

Highest-leverage next move

Put one assured number into a live operating decision — scheduling, a setpoint, a supplier award — with a holdout so the change is attributable.

Cost of leaving

Effort: 12–18 months
Team: Reporting controller, data engineer, internal audit partner, model owner
Risk: Higher — the control framework, not the pipeline, becomes the binding constraint
To next stage: 12–18 months

If this is you, the next step is

We sample twenty figures against your own systems and report trace time and gaps.

Run an assurance dry run

Stage 5

Steering

3% of operators sit here

The same assured dataset drives operating decisions — scheduling against carbon intensity, setpoints, supplier awards, product design — with models recommending and, inside stated bounds, acting.

Stage 5 is narrower than the phrase suggests and should stay that way. It is not an autonomous sustainability function; it is a small, enumerated set of decisions in which an environmental figure is a live constraint or objective. Scheduling energy-intensive batches against a forecast grid carbon intensity qualifies. Setpoint optimisation on a dryer or a kiln where energy is the dominant cost qualifies. Choosing between two qualified suppliers on a blended price-and-footprint score qualifies. Anything touching product safety, permitted emissions limits or a regulatory declaration stays under human decision indefinitely, and that is the correct answer rather than a stage not yet reached.

The engineering is largely done by the time an operator arrives here; what is new is a control problem with a specific shape. A number that steers operations and is also disclosed has become both a management KPI and a reported figure, and the incentive to shade it now exists inside the plant rather than only at group level. The structural answer is separation with reconciliation: one calculation engine, a steering view that can be provisional and fast, a disclosure view that is frozen and slow, and a scheduled reconciliation whose divergences are investigated rather than explained.

Regression is the standing risk. Grid factors change, product mixes shift, a new line arrives with no allocation rule, and a decision policy tuned on last year's conditions quietly stops being valid. The leading indicator is the same one autonomy programmes use everywhere: the rate at which decisions fall outside their stated bounds and escalate. When it rises, the world has moved outside the policy before an incident says so.

In practice

The batch that moved to Tuesday

A food manufacturer with a large refrigerated drying load schedules its most energy-intensive campaigns against a day-ahead grid carbon-intensity forecast, inside bounds set by the production planner: never at the cost of a customer commitment, never more than a stated shift in the published plan, always reversible by one action. Roughly one campaign in six is moved. The reported saving is computed against a holdout of comparable campaigns left on the original schedule, and the steering figure reconciles monthly to the disclosed one.

What it looks like

At least one scheduled decision carries an environmental constraint or objective, not just a report
Design and procurement see product-level footprints at the point of decision
The steering copy and the disclosed copy of every figure reconcile on a stated cadence
The decision policy — what may execute unattended, within what bounds — is versioned and reviewed

Diagnostic signals you can check this week

Name a decision in the last quarter that changed because of an environmental figure, and the holdout it was measured against
Ask whether the steering figure and the disclosed figure have ever diverged, and what happened when they did
Check whether the decision policy is versioned and reviewed, or edited in a settings screen
Ask what the escalation rate has done over the last six months. Nobody watching it is the regression signal

Anti-pattern · Letting the steering number become the reported number

Once a provisional figure is good enough to schedule against, the temptation is to disclose it — it is fresher, it is already in front of people, and maintaining two views feels like duplication. It is not duplication; it is the control. The steering view is allowed to be fast and revisable, which is precisely what a disclosed figure must not be. Collapse them and the first time a provisional value is restated after publication, the whole dataset's credibility is spent.

What holds you here

Sustaining it is a control problem: a figure that steers operations and is also disclosed is both a target and a measure, and the policy behind it drifts out of validity without anyone noticing.

Highest-leverage next move

Separate the steering copy from the disclosed copy with a scheduled reconciliation, and version the decision policy with the same rigour as the model.

Cost of leaving

Effort: Continuous
Team: Platform team, production planning, plus a standing reporting and controls forum
Risk: Concentrated — low frequency, high consequence, and now visible to a regulator

If this is you, the next step is

We take one decision and test its bounds, its reconciliation and its rollback against a real scenario.

Stress-test a steering decision

Where manufacturers actually sit on the ladder

The distribution across the five stages, and why the metered-to-allocated step is the largest single transition loss.

Most non-automotive manufacturers are at stage 2: they have real primary data on site energy and no way to attribute it to a product. The distribution below is weighted heavily toward that middle, and the gap between stage 2 and stage 3 is the largest single drop on the ladder — not because allocation is technically hard, but because it sits between two functions and belongs to neither. The energy team owns the meters, the MES team owns the production orders, and the join is nobody's objective until a customer or a customs declaration makes it one.

Illustrative distribution of manufacturers across the evidence ladder

Illustrative, not measured: a model-derived distribution synthesised from CDP's disclosure data on primary-data availability, acatech's Industrie 4.0 Maturity Index stage progression and the WEF Global Lighthouse Network's reporting on scaled deployments. Treat the shape as the argument, not the individual percentages.

Share of manufacturers

22% — 1 · Invoice-grade
37% — 2 · Metered (the plateau)
27% — 3 · Allocated
11% — 4 · Assured
3% — 5 · Steering

Source: Illustrative distribution, synthesised from CDP, acatech and World Economic Forum research

The imbalance is structural rather than cultural. CDP reports (opens in a new tab) that supply-chain Scope 3 emissions are on average 26 times a company's direct operational emissions, and that corporates are roughly twice as likely to measure operational emissions as supply-chain ones. For a manufacturer, that ratio maps almost exactly onto the evidence-grade picture: the smaller part of the footprint is grade A and the larger part is grade D. Any readiness programme that starts with the metered part is starting where the data already is rather than where the footprint is — which is defensible as a first step and indefensible as a destination. It is worth reading that alongside the IEA's tracking of industrial energy demand and emissions (opens in a new tab), which is a reminder that the absolute quantities involved are large enough that the resolution argument is not an accounting nicety.

The external anchors for the ladder itself are worth naming. acatech's Industrie 4.0 Maturity Index (opens in a new tab) describes the same shape for production data — computerisation, connectivity, visibility, transparency, predictive capacity, adaptability — and the ESG ladder is essentially that progression applied to environmental data with an assurance constraint bolted on. The World Economic Forum's Global Lighthouse Network (opens in a new tab) documents manufacturers that have scaled digital deployments across multiple sites, and its recurring finding — that scaling is an organisational problem long before it is a technical one — is precisely what the stage 2 plateau looks like from the inside. NIST's manufacturing programmes (opens in a new tab) and MHI's annual industry survey (opens in a new tab) track the same adoption-versus-impact gap across the wider industrial base.

What each obligation actually demands from the plant

CSRD and ESRS, CBAM, ESPR and the product passport, ISSB, ISO 50001 and the social and governance strands — the number each one wants, the grade it demands, and the system it has to come from.

Each reporting obligation resolves, in the plant, to a specific number at a specific resolution from a specific system — and they do not all ask for the same thing. That distinction is the whole planning problem. A manufacturer that reads CSRD, CBAM, ESPR and a customer questionnaire as four versions of 'report your emissions' will build four workbooks. A manufacturer that reads them as four cuts of one allocated dataset will build the allocation once and serve all four from it, which is the difference between a permanent reporting function and a capability.

Obligation or driver	The number it demands	Grade demanded	System of record	Answerable from stage
ESRS E1 under CSRD	Scope 1, 2 and 3 by category, energy mix, targets and transition plan	A for Scope 1–2, B–C for material Scope 3	EMS or historian, ERP, factor register	3
CBAM definitive regime	Embedded emissions per tonne of covered good, per consignment, verified	A for direct, B for precursors	MES, EMS, ERP and the customs declaration	3
ESPR and the digital product passport	Product-level attributes and footprint per item or batch	A–B at product resolution	PLM, MES, factor register	4
Customer PCF requests	Cradle-to-gate kg CO₂e per unit with a stated boundary	B, with grade A energy	PLM and MES, joined to the BOM	3
ISSB IFRS S1 and S2	Financially material sustainability risk, on investor-grade controls	A–B with lineage	Group consolidation, on the same source data	4
ISO 50001 and ISO 14001	Energy performance indicators, significant energy uses, objectives	A	EMS, normalised for output and weather	2
ESRS S1 — own workforce	Headcount, incident and injury rates, training hours, pay gap	A from HR and EHS systems	HRIS and EHS incident system	2
EU AI Act and AI management systems	An inventory of AI systems, risk classification, human oversight evidence	Register, not measurement	Model register and change control	4

The manufacturing ESG obligation map. 'Grade demanded' is the minimum evidence grade the obligation realistically requires for a material line; 'stage' is where on the ladder the requirement first becomes answerable without heroics.

Three rows carry most of the sequencing weight. The CBAM definitive regime (opens in a new tab) has applied since 1 January 2026 across six goods sectors — cement, iron and steel, aluminium, fertilisers, electricity and hydrogen — and requires embedded-emissions data per consignment rather than per year, which is a resolution demand no site-and-year figure can be reshaped into. The Ecodesign for Sustainable Products Regulation (opens in a new tab) introduces the digital product passport, pushing the same requirement down to the individual item or batch. And CSRD reporting under the ESRS (opens in a new tab) brings assurance into the picture, which is what converts a data-quality preference into a control requirement. The scope of who must report has been actively debated and narrowed since the directive came into force; the resolution the underlying obligations demand has not moved.

A digital identity card for products, components, and materials, which will store relevant information to support products' sustainability, promote their circularity and strengthen legal compliance.
European Commission, on the Digital Product Passport

The accounting standards underneath all of this are stable and worth reading once rather than paraphrased forever. The GHG Protocol Corporate Standard (opens in a new tab) defines the organisational inventory and its five accounting principles; the Product Standard (opens in a new tab) and ISO 14067 define a product footprint and, critically, the allocation rules a manufacturer will otherwise argue about internally; and the Scope 3 calculation guidance (opens in a new tab) is explicit that supplier-specific data is preferred over averages, which is the formal statement of the grade ladder. The ISSB standards (opens in a new tab) layer investor-grade control expectations on the same numbers. ISO 14064-1 for organisational inventories, ISO 14064-3 for verification, ISO 50001 for energy management and ISO 14001 for environmental management are all catalogued at the ISO standards catalogue (opens in a new tab); none of them prohibits a model in the calculation path, and all of them require that the path be describable.

The social and governance strands are usually treated as a separate problem and should not be. ESRS S1 asks for workforce data — headcount, incident rates, training hours, pay gaps — which lives in HR and EHS systems that have their own grade problem: incident narratives are free text, contractor headcount is often reconstructed from site access logs, and training records span three systems. The temptation to point a language model at all of it is strong and the constraint is sharp, because worker data carries obligations that energy data does not. The governance strand closes the loop: under the EU AI Act's risk-based framework (opens in a new tab) and management-system standards such as ISO/IEC 42001, the models you deploy in the ESG chain are themselves in scope for inventory, classification and oversight evidence. NIST's AI Risk Management Framework (opens in a new tab) is the most practical free reference for structuring that register, and the human-rights and worker-data side is treated in depth in our manufacturing AI human-rights governance guide.

What the transitions look like in public

Three publicly reported manufacturing programmes, read against the evidence ladder. None is an Atomic Loops engagement — every figure is quoted from the operator's own published material.

Very few manufacturers publish the detail of how AI touches their ESG numbers, and that absence is itself informative: what they do publish is the evidence chain — the unit of account, the resolution, the reporting perimeter — which is exactly the thing that has to exist before any model matters. Read the three below for their data architecture rather than their targets. Each one shows a different rung of the ladder made visible in public disclosure.

Three manufacturers, read against the ladder

Outcomes and figures as published by the operators themselves; verify against the linked source before reusing them, as we have not independently audited them. Card images are generated industry scenes from our own library, not photographs of these companies, and imply no endorsement.

Illustrated scene: engineers reviewing an overlaid data model of an automated assembly line inside a manufacturing hall

SiemensIndustrial technology and electronics · multi-site global estate24

Challenge: Reporting environmental performance consistently across a very large, heterogeneous plant estate, where each site had its own metering history, its own systems and its own definitions of the same quantity.
Approach: Siemens organises its sustainability work under a published framework it calls DEGREE, with targets set to 2030, and consolidates performance into a single Sustainability Statement covering strategy, governance, operationalisation and performance across its material topics.
Reported outcome: Siemens publishes a Sustainability Statement 2025 that it describes as delivering 'clear, comprehensive insights into our ESG strategy, governance frameworks, operationalization, and performance across all material sustainability topics', and states that more than 90% of its business enables customers to achieve a positive sustainability impact.; What it shows about the curveThe visible artefact is a single consolidated statement across a heterogeneous estate. That is a stage-4 signature: it is only producible when every site's figures share definitions, boundaries and a controlled consolidation path, which is a data-governance achievement long before it is a reporting one.

Siemens — sustainability (opens in a new tab)

Illustrated scene: a leadership team in a meeting room reviewing an environmental reporting schematic, with a production hall visible through the glass wall

HolcimBuilding materials · cement, aggregates and ready-mix concrete34

Challenge: In cement and concrete the product is the emission, and the unit of account is the tonne. A site-and-year figure is commercially useless when customers, specifiers and — under CBAM — customs authorities are asking about a specific material at a specific intensity.
Approach: Holcim reports environmental performance in per-tonne intensity terms alongside the commercial share of its lower-carbon product ranges, tying the environmental unit of account directly to the sales unit of account.
Reported outcome: Holcim publishes Scope 1 intensity of 502 kg CO₂ per tonne of cementitious material, down 11% against its 2020 baseline, against a 2030 target of under 400 kg; it reports ECOPlanet cement at a 36% share of cement net sales and ECOPact ready-mix at 31% of ready-mix net sales, with a target of over 50% of net sales from the two ranges by 2030, and freshwater withdrawal of 179 litres per tonne, 25% below its 2020 baseline.; What it shows about the curveWhen the environmental figure and the revenue figure share a denominator, allocation stops being a reporting cost and becomes a commercial instrument. That alignment is what funds the stage 3 to 4 transition in materials businesses, and it is why CBAM-exposed sectors are ahead of the rest of manufacturing on product-level data.

Holcim — sustainability (opens in a new tab)

Illustrated scene: an operator at a machine terminal on a production line, with an overlaid environmental data panel above the cell

HenkelAdhesives, consumer brands and industrial chemicals · global plant network34

Challenge: Moving sustainability data out of a standalone reporting cycle and into the same control environment, calendar and perimeter as financial reporting, across a plant network producing both industrial and consumer goods.
Approach: Henkel publishes under what it calls its 2030+ Sustainability Ambition Framework, covering the three ESG dimensions it names Regenerative Planet, Thriving Communities and Trusted Partner, and reports through a Sustainable Impact Report and a separate Sustainability Indicators set.
Reported outcome: Henkel publishes its Sustainability Statement inside its Annual Report — the report is listed as 'Annual Report 2025 incl. Sustainability Statement' — alongside a Sustainable Impact Report 2025 and Sustainability Indicators 2025.; What it shows about the curvePutting the sustainability statement inside the annual report moves the data into the audited perimeter, on the financial calendar, under the same controls. That is the clearest external marker of stage 4 available to an outsider: the number now has to close when the books close.

Henkel — sustainability (opens in a new tab)

The common thread is that none of these manufacturers is publicly differentiated by its models. What each publishes is a property of its evidence chain: a consolidated statement across a heterogeneous estate, an environmental unit of account matched to the commercial one, or a sustainability statement inside the audited annual report. Those are the conditions under which a model becomes useful and safe. A manufacturer at stage 2 that deploys the same model gets a faster route to a number nobody can trace.

The four dimensions that set your stage

Readiness is not one number. Four dimensions gate each other, and the lowest one caps the grade of everything you publish.

Readiness is scored on four dimensions — evidence quality, allocation and traceability, assurance and control, and decision use — and the lowest of them is the real stage, because each gates the others. Excellent metering with no allocation produces a dataset only the energy team can use. Perfect allocation over spend-based factors produces an auditable pipeline of estimates. Rigorous control over data nobody acts on produces an expensive annual artefact. The dimensions are not a scorecard for its own sake; they are the four ways a chain can fail while looking healthy from every other angle.

Evidence quality
What grade your inputs actually are, and whether the grade is recorded rather than remembered. The binding question is what share of a material figure is grade A or B, and whether anyone could tell from the disclosure. This dimension is usually strong on site energy and weak everywhere else, which mirrors the 26-to-1 ratio CDP reports (opens in a new tab) between supply-chain and operational emissions.
Allocation and traceability
Whether a figure can be cut to the resolution an obligation demands, and whether it can be traced back to raw. This is the dimension that separates stage 2 from stage 3, and it is overwhelmingly the lowest-scoring one in manufacturing — not for technical reasons but because the join between meters and production orders sits between two functions and belongs to neither.
Assurance and control
Whether the methodology and factor set are versioned and frozen per period, whether restatement is a policy rather than an accident, and whether any model in the path is registered, reviewed and labelled. Manufacturers consistently overestimate this dimension because the controls exist as documents; the test is whether the restatement policy has ever been used.
Decision use
Whether an environmental figure changes an operating decision, and whether the change can be attributed. This is the dimension that determines funding. A dataset used only for disclosure is defended annually against its cost; a dataset that schedules a batch, sets a setpoint or wins a supplier award has an operational owner whose targets improve when it is right.

Diagnosing the real constraint

Plot your primary-data share against your allocation and traceability. The quadrant names the next investment — and three of the four common answers are not 'get better data'.

Metered but unallocated

Good primary data trapped inside the energy function
The most common position, and the highest-leverage one
Fix: join one meter to one line's production orders — not more meters

Assurance-ready

Grade and resolution both in place
Constraint has moved to decision use and control
Fix: put one assured figure into a live decision with a holdout

Spreadsheet ESG

Neither grade nor resolution
Normal at stage 1, and cheap to leave
Fix: meter one utility at one site and freeze one written methodology

Precisely wrong

An auditable pipeline built on modelled inputs
The most dangerous quadrant, because it feels assured
Fix: raise the grade of the largest contributors before raising resolution further

Primary-data share — top: Mostly metered inputs, bottom: Mostly modelled inputs

Allocation and traceability — left: Site and year only, right: Order and SKU, traceable to raw

The bottom-right quadrant deserves the warning it gets. A manufacturer that has invested in pipelines, lineage and controls over a Scope 3 inventory built from spend-based factors has produced something that looks exactly like readiness from a governance review and is, in evidence terms, an annual report of estimates with excellent version history. It is also the position most likely to be reached by buying software first, because a platform can supply structure and cannot supply grade. The test is simple and unforgiving: pick your three largest emission lines and ask what physical measurement or supplier document sits at the bottom of each.

The evidence architecture, layer by layer

What actually has to exist for each stage of the ladder — and which layer you can safely defer.

An assurance-grade ESG chain needs six layers, and the order in which they are built decides whether the programme compounds or stalls. The architecture below is deliberately unfashionable: nothing in it names a vendor, every layer is defined by what it must guarantee rather than what product provides it, and each is annotated with the stage that first requires it. Two layers are routinely built too early — a reporting platform before allocation, and a model before a factor register — and one is routinely built too late, which is the lineage store.

Layers required by stage

Read down the stack, not across it. A manufacturer trying to reach stage 4 without the methodology register and the lineage layer is building a stage-3 pipeline with a compliance label on it.

Sensing and metering
Stage 1+
- Utility sub-metersElectricity, gas, steam, compressed air, water — at area or asset level
- Mass and waste recordsWeighbridge tickets, waste contractor manifests, effluent monitoring
- Process instrumentationTags already present in the DCS or PLC layer, retained as interval series
→
Production context
Stage 2+
- Production ordersStart, stop, quantity and product from the MES — the join key for everything
- BOM and routingsPurchased inputs per unit, from PLM or ERP, versioned with the product
- Downtime and scrapSo energy consumed making rejects is attributed rather than lost
→
Methodology and factor register
Stage 3+
- Boundary and consolidation rulesOrganisational and operational boundaries, frozen per reporting period
- Emission factor setEach factor versioned, dated, sourced and graded; no in-place updates
- Allocation rulesShared utilities, changeover, co-products and waste — written and defensible
→
Calculation and allocation engine
Stage 3+
- Reproducible from rawAny figure recomputable from stored inputs and a stated factor version
- Grade carried per fieldEvery value travels with its grade; aggregation propagates the weakest
- Recomputation on changeA factor or rule change produces a labelled restatement, never a silent edit
→
AI assist layer
Stage 3+
- Document extractionEPDs and PCF sheets to structured entries, each verified by a named human
- Anomaly and drift detectionOn meter series: stuck sensors, flat lines, step changes, seasonal drift
- Bounded gap-fillWithin limits set by the methodology register, labelled and given a band
- Model register entryPurpose, version, inputs, reviewer, override rate — one entry per use
→
Disclosure, control and assurance
Stage 4+
- Lineage storeFigure to raw record, on demand, months later — the trace-time asset
- Restatement policy and change logWhat triggers a restatement, who approves it, how it is disclosed
- Assurance sampling interfaceThe assurer picks a figure and gets the trace without a project

Pipeline described

Sensing and metering (stage 1+) — Utility sub-meters: Electricity, gas, steam, compressed air, water — at area or asset level; Mass and waste records: Weighbridge tickets, waste contractor manifests, effluent monitoring; Process instrumentation: Tags already present in the DCS or PLC layer, retained as interval series
Production context (stage 2+) — Production orders: Start, stop, quantity and product from the MES — the join key for everything; BOM and routings: Purchased inputs per unit, from PLM or ERP, versioned with the product; Downtime and scrap: So energy consumed making rejects is attributed rather than lost
Methodology and factor register (stage 3+) — Boundary and consolidation rules: Organisational and operational boundaries, frozen per reporting period; Emission factor set: Each factor versioned, dated, sourced and graded; no in-place updates; Allocation rules: Shared utilities, changeover, co-products and waste — written and defensible
Calculation and allocation engine (stage 3+) — Reproducible from raw: Any figure recomputable from stored inputs and a stated factor version; Grade carried per field: Every value travels with its grade; aggregation propagates the weakest; Recomputation on change: A factor or rule change produces a labelled restatement, never a silent edit
AI assist layer (stage 3+) — Document extraction: EPDs and PCF sheets to structured entries, each verified by a named human; Anomaly and drift detection: On meter series: stuck sensors, flat lines, step changes, seasonal drift; Bounded gap-fill: Within limits set by the methodology register, labelled and given a band; Model register entry: Purpose, version, inputs, reviewer, override rate — one entry per use
Disclosure, control and assurance (stage 4+) — Lineage store: Figure to raw record, on demand, months later — the trace-time asset; Restatement policy and change log: What triggers a restatement, who approves it, how it is disclosed; Assurance sampling interface: The assurer picks a figure and gets the trace without a project

Step-by-step insights

Sensing and metering — coverage matters less than retention and identity: Two properties decide whether a metering estate can support allocation, and neither is the meter count. The first is retention of raw interval data, because a production order lasts hours and cannot be decomposed out of a monthly rollup. The second is asset identity: a meter labelled 'MCC-3' in the EMS and 'Line 2 feeder' on the drawing is a reconciliation problem every single time somebody wants to attribute it. Fixing the naming and the retention policy on the meters you already have is usually a fortnight, and it unblocks more than a capital metering programme does.
Production context — the MES is the join, and it already exists: The allocation join happens at the production-order level, which is ISA-95 level 3 and already the MES's job. This is the layer manufacturers most often try to solve somewhere else — in the reporting platform, in a data warehouse, in a spreadsheet — because the MES team has a full backlog. It does not work: the order start and stop times, the quantity and the product are the join key, and every other source of them is a copy with drift. Include downtime and scrap, or energy consumed making rejects silently improves your per-unit figure while the plant gets worse.
Methodology and factor register — the artefact an assurer actually reads: This layer is a register, not a system, and it can start as a version-controlled document. Its content is the boundary, the consolidation approach, the allocation rules and the factor set with a version and a date per factor. Its defining rule is that it is frozen for the reporting period. Manufacturers who skip it discover the cost the first time two plants compute compressed-air allocation differently and a customer compares the two product footprints. The register is also where the bounds on gap-fill live, which is what makes the AI layer governable rather than merely monitored.
Calculation and allocation engine — propagate the weakest grade: The engineering rule that makes the grade ladder work in practice is propagation: when values are aggregated, the result carries the weakest grade of any material contributor, not an average. That single rule prevents the most common silent failure, in which a grade A energy figure and a grade D purchased-material estimate combine into a total that inherits the credibility of the better half. It also produces a genuinely useful management metric — the modelled share of each disclosed figure — which is the number to put a ceiling on and alert against.
AI assist layer — narrow scope, named reviewer, retained originals: Everything in this layer proposes; nothing disposes. Extraction proposes a structured entry from a supplier PDF and a human confirms it, with the original document retained and linked so the assurer samples the document rather than the model. Anomaly detection proposes that a meter has stuck and a human decides whether to exclude the period. Gap-fill proposes a value inside stated bounds and the result stays labelled grade D forever. The register entry is the control that makes all of it defensible: purpose, version, inputs, reviewer, override rate. Write it while building, not in the month before the audit.
Disclosure and assurance — build lineage while the pipeline is being built: The lineage store is the layer most often deferred and the one that is hardest to retrofit, because reconstructing provenance for figures whose inputs have since changed is strictly harder than recording it as you go. Its payoff is a single measurable property: trace time. A manufacturer that can return the full derivation of a sampled figure in minutes turns an assurance engagement from an evidence hunt into a sampling exercise, and gets a second benefit that nobody plans for — internal disputes about whose number is right end, because the trace settles them.

The layer most often skipped is the methodology register, and its absence is invisible until it is expensive. Everything above it still functions: meters read, orders join, figures compute. What is missing is the statement of what the figures mean, so two plants diverge, a factor library updates in place, and the first restatement arrives as a surprise rather than as a process. It costs a fortnight to write and it is the cheapest insurance on the page. Note also that nothing in this stack is a new layer of the plant: the production context it depends on is the level-3 layer described by the ISA-95 enterprise-control integration standard (opens in a new tab), which your MES already occupies. The evidence architecture is a set of guarantees laid over systems you have, not a parallel estate.

A 90-day plan: a product carbon footprint for one SKU family

The stage 2 to 3 transition made concrete on one manufacturing problem — cradle-to-gate kg CO₂e per unit, from metered energy allocated to production orders. Contains no model development.

Moving one stage takes about 90 days when it is scoped to a single product family and several years when it is scoped to a reporting function. To make that concrete, the plan below runs the transition on the most common trigger in non-automotive manufacturing: a customer, or a CBAM declaration, asking for a cradle-to-gate footprint for a specific product, at a resolution the current reporting chain cannot produce. The metering already exists at most stage-2 plants, so the quarter contains no new instrumentation and no model development — it is allocation, grading and control.

Stage 2 to stage 3 on one SKU family, in one quarter

One plant, one line, one high-runner SKU family, two named owners. If a phase needs longer than its window, narrow the scope — fewer SKUs, one shift pattern — rather than extending the plan.

Days 1–15
Freeze the boundary and name the owners
Pick one high-runner SKU family on one line. Fix the boundary as cradle-to-gate and write the methodology down against the GHG Protocol Product Standard and ISO 14067. Inventory the meters that cover the line and list the gaps rather than filling them. Name two owners jointly: the plant energy engineer, who owns the physical data, and the group reporting controller, who owns the disclosure. One owner alone always fails — the first cannot publish and the second cannot measure.
A frozen boundary, a written method, two named owners, a meter gap list
Days 16–45
Allocate metered energy to production orders
Join the interval series from the historian or EMS to MES production-order start and stop times. Agree and document the allocation rule for every shared utility touching the line — compressed air, chilled water, HVAC, extraction — and decide explicitly where changeover and clean-down energy goes. Compute kWh per unit by line and shift, then reconcile the total against the site's monthly utility invoice. The reconciliation is the test: if allocated energy does not sum to metered energy, the rule is wrong.
kWh and kg CO₂e per production order, reconciled to the invoice
Days 46–70
Grade the inputs and build the factor register
Every line on the BOM gets a factor with a grade, a source and a date. Supplier EPDs and PCF sheets go through extraction into the register with a named human verifying each entry and the original document retained and linked. Materials with no supplier data get the best available published factor at grade B, and genuine gaps get a labelled grade D estimate with an uncertainty band. Record the resulting grade mix for the product — this number is the honest headline, not the footprint itself.
A graded BOM with nothing unlabelled, and a stated grade mix
Days 71–90
Dry-run the assurance and publish with the mix stated
Hand your assurer, or internal audit if you have not appointed one, twenty sampled figures from the calculation and ask them to trace each to a meter reading or a supplier document. Time every trace and record the findings by category. Fix what the sample exposes, then issue the product footprint with its boundary, its allocation rules and its grade mix stated on the face of it. A footprint published without those three is a number; published with them it is evidence.
A footprint a customer and a verifier both accept, and a trace time you can quote

The order matters

Grade before you automate
Automating the production of a grade D figure produces grade D figures faster. Establish what each input actually is, then decide what a model is allowed to do to it. The extraction pipeline built after grading knows which documents matter; the one built before it processes everything at equal priority and saves nothing that counts.
Reconcile to the invoice before you trust the meter
Allocated energy must sum back to metered energy, and metered energy must sum back to the invoice within a stated tolerance. Both reconciliations catch the errors that matter — a missing sub-meter, a double-counted feeder, a rule that quietly drops changeover. Skip them and the first person to check will be an assurer.
One SKU family before one catalogue
The rules argued about on the first product family — shared utilities, changeover, co-products, scrap — are the same rules the next fifty need. Settle them once on a family where the answer is checkable against a customer's own expectations, then scale. Starting with the catalogue means settling every rule simultaneously with no way to test any of them.
Publish the grade mix, not just the number
Stating that a footprint is, say, mostly grade A energy and grade B materials with a small labelled estimate is a stronger commercial position than a bare number, because it is the thing a sophisticated customer will ask for next. It also protects you: a disclosed mix cannot be misread as a claim of measurement.

What makes this a 90-day plan rather than a programme is that every phase has a checkable exit condition and none of them requires new hardware, new vendors or a model. The wider digitisation sequencing this sits inside — which systems get replaced, in what order, and how the plant floor absorbs the change — is covered in our factory digitisation phases guide, and the two plans are designed to run against the same MES and metering work rather than competing for it.

Instrumenting readiness: formula, source, cadence

Where each readiness metric actually comes from — the formula, the system that produces it, and how often to read it. All telemetry, no self-report.

A readiness metric you cannot name a source system for is an opinion, and ESG readiness is unusually prone to opinions because everyone involved has a stake in the answer. Every metric below reduces to counts and timestamps that the historian, the MES, the factor register, the calculation engine or the review log already record — the work is joining them, not creating them. The right-hand column is the honesty test: the stage at which the metric first measures something real rather than something aspirational.

Metric	Formula / read	Source	Cadence	Honest from
Primary-data share	Grade A and B inputs ÷ all material inputs, by contribution not by count	Calculation engine + factor register	Monthly	Stage 2
Allocation coverage	Production orders carrying an allocated energy figure ÷ all orders in scope	MES + calculation engine	Weekly	Stage 3
Reconciliation variance	\|allocated energy − metered energy\| ÷ metered energy	Historian + calculation engine	Monthly	Stage 3
Trace time	Median minutes to trace a sampled disclosed figure to its raw record	Lineage store	Per assurance sample	Stage 3
Factor currency	Factors inside their stated refresh window ÷ all factors in use	Factor register	Quarterly	Stage 3
Model-assisted field share	Fields whose value a model proposed ÷ all fields in the disclosure	Calculation engine + model register	Monthly	Stage 3
Override rate	Reviewer changes ÷ model-proposed entries	Review log	Monthly	Stage 3
Restatement rate	Figures restated after publication ÷ figures published	Disclosure record + change log	Per reporting cycle	Stage 4
Assurance findings	Findings raised per engagement, by category and severity	Assurer's report	Per reporting cycle	Stage 4
Steered decision share	Decisions executed against an environmental constraint ÷ decisions in scope	Scheduler or decision log	Weekly	Stage 5

Instrumentation build sheet for ESG evidence-chain readiness in a manufacturing estate. 'Honest from' is the stage at which the metric starts measuring something real.

Four of those metrics are enough to verify a stage transition on their own, and each has a threshold separating the stage beneath from the stage above. All four are readable from system telemetry rather than from a workshop, which is the point: a readiness claim that can only be established by asking people is a readiness claim that will not survive an assurer asking the same people differently.

Metric	Stage 2	Stage 3	Stage 4	How to read it
Smallest producible unit	Site, month	Production order	Order, with lineage	Ask for one SKU's figure for last month and time the answer
Trace time (median)	Weeks	Hours	Minutes	Sample a disclosed figure at random and time the trace to raw
Grade mix	Unknown	Known internally	Known and disclosed	Whether the modelled share of a material figure can be stated
Factor control	Copied forward	Versioned	Frozen per period, with restatement policy	Whether last year's figure still reproduces exactly today

Verification metrics for each stage transition. All four are readable from telemetry rather than self-report.

Stage 4 assurance-readiness checklist

If you cannot tick all seven, you are still at stage 3 regardless of how good the pipeline is. Tick as you go — this list works without JavaScript.

Every disclosed figure carries an evidence grade, not just a valueAnd aggregation propagates the weakest grade rather than averaging it.
The factor set is versioned, dated and frozen for the reporting periodA library update applies prospectively, never in place.
A sampled figure traces to a meter reading or supplier document in under an hour
Model-derived fields are flagged and have a named human reviewerWith the override log retained and readable months later.
A written restatement policy exists and has been used at least onceA policy never exercised is a document, not a control.
The allocation rule is documented per product familyShared utilities, changeover, co-products and scrap, all written down.
Every AI system touching the ESG chain is in a register with a risk classification

0 of 7 ticked

Tick honestly — an empty list is still a diagnosis

Most stage-3 manufacturers can genuinely tick two or three of these, not zero. If none apply yet, do not start with controls: run the 90-day plan above on one SKU family. Five of these seven items fall out of doing that once, because the plan forces the grading, the allocation rule and the trace.

Scope the 90 days with us

Failure modes that send manufacturers backwards

Readiness is not monotonic. Five regressions account for almost all of it, and four are silent until an assurer finds them.

Readiness regresses, usually without anyone noticing, because the conditions that made a figure defensible quietly stopped holding while the figure kept being produced. That is the specific danger of an ESG chain compared with an operational one: an operational model that degrades gets noticed when the plant does worse, whereas a disclosure that degrades gets noticed when a third party checks it, which is between twelve and eighteen months later. Five regressions account for almost all of it.

Likelihood: highImpact: high

A factor library updates and history silently changes

Grid factors, material factors and supplier EPDs all get revised by their publishers. Where the calculation reads a live factor library rather than a frozen set, prior-period figures move without a change record, and last year's disclosed number stops reproducing. The discovery point is almost always an assurer asking why.

PreventionFreeze the factor set per reporting period; apply updates prospectively and handle material changes as a labelled restatement.

Likelihood: highImpact: high

Gap-fill quietly becomes the number

Each individual estimate is reasonable — one unresponsive supplier, one month of missing meter data, one new material. Nobody watches the aggregate, so the modelled share of a material figure drifts upward over two reporting cycles until most of the total is grade D wearing the credibility of the grade A part.

PreventionSet a ceiling on the modelled share of any material figure, compute it monthly, and alert when it is approached.

Likelihood: highImpact: medium

The meter that stopped reading

A sub-meter sticks, a gateway drops, an integration silently returns the last known value. Consumption appears flat or falls, which reads as an efficiency gain rather than a fault, and it can survive a full reporting cycle because nobody expects good news to be wrong.

PreventionFreshness and plausibility alerting on every meter feeding a disclosed figure — a flat line is a failure state, not a saving.

Likelihood: mediumImpact: high

The steering copy and the disclosed copy diverge

Once a provisional figure is good enough to schedule against, it starts being quoted in meetings, then in customer conversations, then in a slide that reaches a report. Two versions of the same quantity now exist with no reconciliation, and the first person to notice is outside the company.

PreventionOne calculation engine, two views, and a scheduled reconciliation whose divergences are investigated rather than explained away.

Likelihood: mediumImpact: high

The AI use nobody registered

A model arrives inside a purchased tool, or an analyst uses one to fill a supplier-data gap under time pressure, and neither appears in any register. The disclosure then contains model-derived values that are indistinguishable from measured ones and cannot be reviewed, which is a control failure regardless of whether the values were good.

PreventionMake the model register entry a gate in the reporting close rather than an annual survey, and ask vendors what estimation their tool performs.

26×

Supply-chain versus operational emissions, on average, for CDP disclosers

CDP

15%

Share of CDP disclosers with any Scope 3 target set

CDP

CBAM goods sectors requiring per-consignment embedded emissions

European Commission

The common structure across all five is that the failure is invisible from inside the reporting process, because the process continues to produce a number of the expected shape. That is why every prevention listed above is a monitor rather than a policy: a ceiling that alerts, a reconciliation that runs, a plausibility check that fires. Policies describe intent, and intent is not what degrades. A useful test of whether your controls are real is to ask which of the five would be caught by something automatic, and which would depend on someone remembering.

Glossary

Hover a term for its definition — or expand the map full screen. The full definitions are written out below.

Evidence grade: The class of an ESG input, from A (a metered primary reading tied to a timestamp and an asset) through B (measured activity times a published factor) and C (a supplier-specific declared figure) to D (a modelled or proxy estimate). A composite figure is worth what its weakest material input can survive, so grade propagates by the weakest contributor rather than by an average.
Primary data: A value obtained by measuring the thing itself — a sub-meter reading, a weighbridge ticket, a supplier's own measured process data — rather than inferred from spend, mass or sector averages. Under the GHG Protocol's Scope 3 guidance, supplier-specific primary data is the preferred input for material categories.
Activity data: A measured quantity of an activity — kilograms of resin consumed, cubic metres of gas burned, tonne-kilometres shipped — that is multiplied by an emission factor to produce an emissions figure. Grade B in this page's ladder: only as good as the factor applied to it and the version control around that factor.
Emission factor: A published coefficient converting an activity quantity into an emissions figure — kg CO₂e per kWh, per kilogram, per tonne-kilometre. Factors are revised by their publishers, which is why they must be versioned, dated and frozen for the duration of a reporting period rather than read live from a library.
Allocation: The rule that assigns a shared consumption to a specific product, order or line — apportioning compressed air across three lines by measured flow, or assigning changeover energy to the product being changed to. Allocation rules must be written down and applied consistently, because they materially change a product footprint and a customer comparing two suppliers will notice if they differ.
Product carbon footprint (PCF): The greenhouse-gas emissions attributable to one unit of a product across a stated boundary, most often cradle-to-gate for a manufacturer. Defined by the GHG Protocol Product Standard and ISO 14067. A PCF without its boundary and allocation rules stated is not comparable with any other PCF.
Limited assurance: An engagement in which an assurer performs enough work to state that nothing has come to their attention suggesting the information is materially misstated — a negative form of conclusion, and a lower bar than the reasonable assurance applied to financial statements. It is still a sampling exercise, which is why trace time matters.
Restatement: A deliberate, labelled revision of a previously published figure, triggered by a factor revision, a boundary change, an acquisition or a discovered error. The distinction that matters is between a restatement, which is disclosed and logged, and a silent recomputation, which is the same change without the record.
Double materiality: The CSRD principle that a company reports both how sustainability matters affect it financially and how its own activities affect people and the environment. Operationally it widens the dataset a manufacturer must be able to evidence, because impact-side disclosures reach into the supply chain and the workforce, not just the balance sheet.
Factor register: The controlled store of every emission factor in use, each with a source, a publication date, a version and a grade, plus a record of who changed what and when. The register, together with the boundary and the allocation rules, forms the methodology artefact an assurer actually reads.
Scope 3: Indirect emissions in a company's value chain that are not from purchased energy — purchased goods and services, transport, use of sold products, end of life. For manufacturers it is usually the dominant share of the footprint and the lowest-grade part of the inventory, which is the central tension this page addresses.
Trace time: The elapsed time to trace a sampled disclosed figure back to the meter reading or supplier document that produced it. The single most useful readiness metric on this page, because it is measured rather than judged: sample a figure, start a timer, and the answer cannot be self-reported optimistically.

Frequently asked questions

The questions manufacturing reporting teams, plant engineers and sustainability leads ask most often when they start grading their own evidence chain.

What does AI readiness for ESG actually mean in a factory?

It means your environmental, social and governance data is metered, allocated to production and traceable to source, so a model can join the calculation path without lowering the assurance grade of the published number. Concretely: interval energy data joined to MES production orders, emission factors held in a versioned register, every value carrying an evidence grade, and a lineage record that lets a sampled figure be traced back to raw. The models involved are ordinary. The admissibility of what they touch is the hard part.

Where should AI be used in ESG reporting, and where should it not?

Use it to move data up the evidence-grade ladder and to label what stays modelled: extracting supplier EPDs and product footprints into a structured register with human verification, detecting stuck or drifting meters, matching bill-of-materials lines to the nearest factor, and gap-filling inside stated bounds with an uncertainty band. Do not use it to produce a value that is then disclosed as if it were measured, to infer a Scope 3 inventory from the purchase ledger, or to generate assurance evidence. Drafting narrative from a locked, already-computed dataset is fine and is not readiness.

How long does it take to move from metered to allocated data?

About 90 days when scoped to one line and one product family with two named owners, and several years when scoped to a reporting function. The work is a join, a set of allocation rules and a reconciliation, not new instrumentation or model development — most stage-2 plants already have the meters. What extends it is scope: attempting every line, every shared utility and every SKU at once means settling every allocation argument simultaneously with no way to test any of the answers against a real customer expectation.

Do we need a data platform before we can do any of this?

No, and buying one first is the most reliable way to spend a year industrialising an estimate. A reporting platform ingests exactly the invoices your spreadsheet ingests and adds workflow, not evidence. Start by joining one interval meter series to one line's production orders and writing down the allocation rules; the requirements the platform actually has to meet become visible after the first product family, not before it. Buy the platform when you know what grade of data it will be holding.

What is the difference between primary data and a spend-based estimate?

Primary data measures the thing: a sub-meter reading, a weighbridge ticket, a supplier's own measured process data. A spend-based estimate multiplies money by a sector average, so it moves with your purchase prices rather than with your emissions — a supplier cutting its footprint in half changes nothing in your inventory, and a price rise increases your reported emissions. That insensitivity is why the GHG Protocol's Scope 3 guidance prefers supplier-specific data for material categories, and why grade D figures should never be presented as measurements.

How do we handle emission factors that change between reporting years?

Freeze the factor set at the start of each reporting period and apply library updates prospectively. Where a revision is material enough to change a published figure, handle it as an explicit, labelled restatement with a change-log entry naming the trigger, the affected figures and the approver. The failure mode to avoid is a calculation that reads a live factor library: prior-year figures then move without a record, last year's disclosure stops reproducing, and the discovery point is an assurer asking why.

What resolution does CBAM actually require from a manufacturer?

Embedded emissions per tonne of covered good, per consignment, rather than an annual site figure — and the definitive regime has applied since 1 January 2026 across six sectors: cement, iron and steel, aluminium, fertilisers, electricity and hydrogen. That is a resolution demand, not a reporting-frequency demand, and it cannot be met by reshaping a site-and-year number. It requires energy and process data allocated to production output, which is the stage 2 to stage 3 transition described on this page.

How do we make our ESG numbers assurance-ready without more headcount?

Measure trace time first: sample five disclosed figures and time how long each takes to reach a meter reading or supplier document. That number tells you whether your problem is evidence or organisation. Then do the three cheapest things — freeze the factor set per period, record a grade on every value, and store the lineage as the calculation runs rather than reconstructing it later. Those three convert an assurance engagement from an evidence hunt into a sampling exercise, and none of them requires an additional person.

Who should own ESG data readiness — sustainability, operations or IT?

All three, with distinct accountabilities and a joint owner pair at the working level. The plant energy engineer owns the physical data: meters, retention, plausibility, reconciliation to the invoice. The reporting controller owns the disclosure: boundary, methodology, factor control, restatement. IT owns the join and its reliability. Programmes with only a sustainability owner stall because nobody can change the MES; programmes with only an operations owner stall because nobody is accountable for what gets published.

Can a language model write our sustainability disclosure?

It can draft narrative from a dataset that is already locked and computed, and that is a legitimate, low-risk use — the numbers are fixed, the model is arranging prose around them, and a human signs it. What it must not do is supply or adjust any figure, infer a value that is not evidenced, or generate the explanation of a method that was not actually followed. The line is straightforward: the model may describe the calculation, never perform it, and any field it did influence must stay flagged in the disclosure.

How does the EU AI Act affect AI used in ESG reporting?

It makes the models themselves part of your governance scope. The Act takes a risk-based approach, and its practical effect for a manufacturer is that any AI system contributing to a disclosed figure needs to be inventoried, classified, and shown to have meaningful human oversight and traceable outputs. Management-system standards such as ISO/IEC 42001 and the NIST AI Risk Management Framework give you a structure for that register. In practice the register entry — purpose, version, inputs, reviewer, override rate — is the same artefact your assurer wants anyway.

What about the social and governance parts, not just carbon?

They have the same grade problem with sharper constraints. Workforce disclosures under ESRS S1 need headcount, incident and injury rates, training hours and pay gaps, which live in HR and EHS systems where incident narratives are free text and contractor headcount is often reconstructed from access logs. Classification and extraction help with the narrative data; using models to score, rank or infer things about individual workers does not, and carries obligations that energy data does not. Grade the inputs the same way, and keep the model firmly on the document rather than the person.

About the author

Atomic Loops Engineering

Industrial AI practice

Atomic Loops builds production AI systems for process, electronics, building-products and consumer-goods manufacturers — energy and emissions allocation, supplier-document extraction, anomaly detection on meter series and decision support wired into the MES and ERP, with the lineage and grading that make the resulting numbers admissible rather than merely plausible.

· Energy and emissions allocation built against live MES production-order data
· Supplier-document extraction pipelines with human verification and override logs
· Delivery includes the evidence layer: grades, factor versions, lineage, restatement
· 21 cited sources on this page

Sources

Manufacturing (Non-Automotive) industry practiceProduction AI systems Atomic Loops has built for process, building-products, electronics and consumer-goods manufacturers.Regulations, compliance and governance for manufacturing AIThe governance cell this page hands off to once a model sits inside a disclosed figure — registers, risk classification and human oversight.Transformation phases for factory digitisationThe wider digitisation roadmap the evidence chain runs alongside — the same MES and metering work, read for production rather than disclosure.

Find out what grade your disclosed numbers actually are

We sample real figures from your last reporting cycle, time the trace on each one, grade the inputs behind them, and leave you with a costed 90-day plan for the weakest dimension. You keep the sample results and the plan whether or not we build anything.

Book an evidence-chain review See our manufacturing work

Published 20 August 2026 · Last updated 20 August 2026

Key takeaways

Abbreviations used on this page

Score your ESG evidence chain

What AI readiness for ESG means in a manufacturing plant

The five stages of an ESG evidence chain

Invoice-grade

Metered

Allocated

Assured

Steering

Where manufacturers actually sit on the ladder

What each obligation actually demands from the plant

What the transitions look like in public

The four dimensions that set your stage

The evidence architecture, layer by layer

Sensing and metering

Production context

Methodology and factor register

Calculation and allocation engine

AI assist layer

Disclosure, control and assurance

A 90-day plan: a product carbon footprint for one SKU family

The order matters

Instrumenting readiness: formula, source, cadence

Failure modes that send manufacturers backwards

A factor library updates and history silently changes

Gap-fill quietly becomes the number

The meter that stopped reading

The steering copy and the disclosed copy diverge

The AI use nobody registered

Glossary

Frequently asked questions

What does AI readiness for ESG actually mean in a factory?

Where should AI be used in ESG reporting, and where should it not?

How long does it take to move from metered to allocated data?

Do we need a data platform before we can do any of this?

What is the difference between primary data and a spend-based estimate?

How do we handle emission factors that change between reporting years?

What resolution does CBAM actually require from a manufacturer?

How do we make our ESG numbers assurance-ready without more headcount?

Who should own ESG data readiness — sustainability, operations or IT?

Can a language model write our sustainability disclosure?

How does the EU AI Act affect AI used in ESG reporting?

What about the social and governance parts, not just carbon?

About the author

Sources

Related reading

Find out what grade your disclosed numbers actually are