Manufacturing (Non-Automotive)Future of AI & Visionary Thinking

AI 2030 and manufacturing hyper-efficiency: what a non-automotive plant can actually book

Manufacturing hyper-efficiency is the pursuit of the last few points of output, yield and energy a plant can take from an asset base it already owns. By 2030 it will belong to factories that can name the physical floor under every loss bucket, prove a gain against a control, and hold it through mix and changeover.

Manufacturing (Non-Automotive) · Future of AI & Visionary Thinking

Atomic Loops EngineeringIndustrial AI practiceReviewed by Atomic Loops Solutions ArchitecturePublished 20 August 202661 min read

Key takeaways

Hyper-efficiency is an accounting discipline before it is an AI capability. A plant that cannot decompose its losses line by line cannot prove a gain, and a gain nobody can find in the cost or energy accounts is not a gain — it is a slide.
Every loss bucket has a floor. The US Department of Energy's manufacturing bandwidth studies name four bands — current typical, state of the art, practical minimum and thermodynamic minimum — and nothing runs below the last one. A target set without a floor is benchmark envy.
Efficiency won away from the constraint is not efficiency, it is inventory. Only recovered minutes on the bottleneck convert into throughput; everywhere else they convert into work-in-progress and a busier plant that ships the same volume.
Most 2030 claims are present-readiness claims wearing a date. Closed-loop setpoint control, in-line quality inspection, predictive maintenance and energy scheduling all run in production today; what is scarce is the metering, the loss model and the measurement-and-verification discipline that let a plant book them.
A gain that is not encoded in the recipe, the setpoint or the standard has an expiry date, and it is usually the next changeover. Re-setting the standard after each accepted gain is what makes the second point cost a fraction of the first.

Abbreviations used on this page

MES: Manufacturing execution system
MOM: Manufacturing operations management
SCADA: Supervisory control and data acquisition
PLC: Programmable logic controller
HMI: Human–machine interface (the operator's screen)
OEE: Overall equipment effectiveness (availability × performance × quality)
TEEP: Total effective equipment performance — OEE measured against all calendar time, not just scheduled time
FPY: First-pass yield — good units through the process without rework
SEC: Specific energy consumption — energy per unit of good output, e.g. kWh per tonne
EnPI: Energy performance indicator, the normalised energy metric ISO 50001 asks for
M&V: Measurement and verification — the protocol that decides whether a claimed saving is a saving
CMMS: Computerised maintenance management system

What manufacturing hyper-efficiency actually means by 2030

A definition, the difference between a claimed point and a booked one, and the path an efficiency idea has to travel before it changes a plant's cost base.

Manufacturing hyper-efficiency is the pursuit of the last few points of output, yield and energy that a plant can still take from an asset base it already owns — the residual after the obvious waste has gone. It is a different problem from the one most improvement programmes were built for. Recovering the first ten points of overall equipment effectiveness is largely a matter of doing known things consistently; recovering the last three is a measurement problem, a control problem and, eventually, a physics problem.

That is why the 2030 framing misleads. The interesting question is not which new class of model arrives, but whether a plant can convert an idea into a number that survives an audit and a changeover. Today, in a typical non-automotive plant — food, pharmaceutical, plastics, packaging, building products, industrial equipment — an efficiency idea has to pass through four gates: it must be attributable to a named loss bucket, bound to a control somebody is allowed to change, proved against something that did not change, and encoded where the next shift will find it. Most ideas die at the third gate, and no model fixes that.

The vocabulary matters here because the same word is doing three jobs. A claimed point is an improvement asserted from a before-and-after comparison. A booked point is one finance has accepted into a standard cost, an energy budget or an ISO 50001 (opens in a new tab) energy performance indicator. A held point is one still present two changeovers later. Plants routinely claim several times what they book, and book several times what they hold — and the gap between those three numbers, not the sophistication of the models, is what decides where a factory lands by 2030.

How an efficiency idea becomes a booked, held point

The path from signal to standard, drawn by rung. The rung is decided by where the arrow stops: rungs 1–2 stop at a claim nobody can reproduce, rung 3 stops at an accepted number in a report, and rungs 4–5 continue into the recipe, the standard and the entitlement register. Most plants are in the top lane.

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

The process, in words

At rungs 1–2, PLC counters and end-of-shift notes are exported by hand into a spreadsheet study, compared across a before-and-after window chosen once the result is known, and presented as a claim. Nobody can regenerate the arithmetic, so the number never reaches the cost accounts and the same bucket is rediscovered a year later.
At rung 3, historian tags, MES states and sub-meter readings feed a loss model in which each bucket is bound to a parameter somebody is allowed to change. The recommendation — a setpoint, a schedule or a maintenance trigger — arrives in the field the operator already works in, with the previous value one switch away, and the change is proved against a matched control under a plan written beforehand, so finance can accept the number.
At rungs 4–5 the accepted setting is written into the recipe and control layer that the changeover procedure actually reads, the standard cost or EnPI is re-set so the saving is not counted twice, and the entitlement register is updated. That last step is the compounding mechanism: re-derived headroom changes which bucket the loss model should attack next, so the programme reprioritises itself instead of repeating itself.

Step-by-step insights

Free-text stop reasons — the habit that caps everything downstream: Nothing above a stop-reason field poisoned by free text can be trusted. Operators under pressure type whatever clears the dialogue fastest, so 'other' and 'minor stop' absorb the buckets that matter most, and the Pareto that results is a map of typing convenience rather than of losses. A fixed taxonomy of eight to twelve reasons, agreed with the shift teams and reachable in one touch at the HMI, changes the quality of every downstream number more than any modelling investment of the same cost. It is also the cheapest change on this diagram.
The before-and-after week — why the comparison must be fixed first: A window selected after the result is known will always find an improvement, because production data has enough natural variation to supply one. The fix is procedural rather than statistical: write the boundary, the independent variables and the comparison down before the change, in the style the international performance measurement and verification protocol has used for energy claims for decades. The discipline's real value is that it lets a plant report a null result without embarrassment, and a programme that has never reported one is not measuring anything.
Binding a bucket to a controllable parameter: The step most loss maps never take is naming, for each bucket, the parameter that governs it and the person allowed to change that parameter. Drying energy is governed by outlet-moisture target and inlet temperature; minor stops on a filler are governed by infeed pressure and changeover set-up; compressed-air cost is governed by header pressure and compressor sequencing. Where the governing parameter is fixed by a validated specification, say so and move on — that bucket belongs in the entitlement register as certification-bound, not in the improvement backlog.
The approval step is data collection, not a concession: Keeping an operator in the loop on the first hundred recommendations looks like caution and functions as instrumentation. Every acceptance and every override, with the conditions attached, becomes the evidence that later sets safe bounds for unattended adjustment. Plants that jump straight to closed-loop control have no such record and end up setting limits by argument. It also matters politically: a change proposal that ships with an approval step and a one-switch revert clears a change-control board in weeks, where the same proposal without them sits for two quarters.
Encoding — where the gain actually survives or dies: The changeover procedure is the real system of record for a production setting. If the optimised value is not in the recipe record, the parameter set or the equipment master that a changeover reads, the gain lasts until the next SKU and then quietly reverts, usually without anyone noticing because reporting is a line average. Encoding also forces a useful conversation about validated state: in a regulated plant, moving a parameter from 'engineer's judgement' to 'recipe' is a change-control event, and it is far better to discover that at the design stage than after the trial has succeeded.
The entitlement register — why the last box changes the first one: Updating the register after each booked gain is what turns a series of projects into a programme. It re-prices the remaining buckets: one that looked large may now be within a few per cent of its state of the art, while a mid-sized one nobody touched may hold most of the remaining headroom. That re-pricing feeds straight back into the loss model at the start of the second lane. Without it, target-setting reverts to last year plus a percentage, effort is allocated by narrative, and the plant spends its best engineers on the hardest remaining point instead of the cheapest available one.

None of the mechanism above is speculative, and none of it is new. It is the ordinary machinery of manufacturing operations management described by the MESA model (opens in a new tab) and the ISA-95 (opens in a new tab) integration levels, with a measurement discipline borrowed from energy engineering. What 2030 changes is how much of it can be automated and how fast the loop runs — not whether the loop is required.

The five rungs in detail: Anecdotal to Entitlement-bound

For each rung: what it looks like on the floor, the diagnostic checks a reviewer can run in an afternoon, the anti-pattern that traps plants there, and what leaving costs.

The ladder below tracks one thing only: how much of a claimed efficiency gain a plant can actually convert into a booked, held number. It deliberately says nothing about how advanced the models are, because model sophistication is almost uncorrelated with rung — some of the most capable data-science groups sit at rung 2, producing excellent predictions that no standard has ever moved for.

Efficiency booked and held, against position on the ladder

The curve is flat through rungs 1 and 2 — where most plants are — because claiming is not booking. It inflects at rung 3, when a change is first proved against something that did not change, and steepens at rung 4, when replication stops costing a project. This is why plants that measure progress in improvement events rather than in re-set standards report activity without a moving cost curve.

Efficiency booked and held by stage

Stage 1 · Anecdotal — 21% of operators. Efficiency is a story told after the fact — improvements are claimed from local wins, and nothing in the plant's cost or energy accounts moves in response.
Stage 2 · Loss-mapped — 34% of operators. Losses are decomposed per line against a defined standard, but the loss map is a report: no bucket is bound to a control that can actually move it.
Stage 3 · Attributed — 27% of operators. One AI-supported change is proved against a matched control under a written M&V plan, and finance accepts the resulting number into a standard or a budget.
Stage 4 · Held — 14% of operators. Proved gains are encoded in the recipe and control layer, survive changeover and mix, and the standard is re-set — so the second line costs a fraction of the first.
Stage 5 · Entitlement-bound — 4% of operators. Every line runs against a maintained, physics-anchored entitlement model, so what remains is bounded by capital, certification and regulation rather than by knowledge.

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

Select a rung

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

Stage 1

Anecdotal

21% of operators sit here

Efficiency is a story told after the fact — improvements are claimed from local wins, and nothing in the plant's cost or energy accounts moves in response.

Rung 1 is not a plant that lacks improvement activity — most sites here run continuous-improvement events constantly and genuinely make things better. What is missing is a shared arithmetic. There is no agreed time model, so 'availability' means scheduled time to one supervisor and calendar time to another; there is no loss taxonomy, so the same twenty-minute stop is a changeover on nights and a breakdown on days; and there is no sub-metering, so energy is a single invoice divided by tonnes.

The tell is that improvement claims cannot be reproduced. Ask for the underlying rows behind a quoted three-point gain and you will get a spreadsheet built once, by one engineer, from an export nobody else can regenerate. The claim may well be true. It is simply not checkable, and an uncheckable claim cannot be booked, replicated or defended when the cost per unit fails to move.

This is the cheapest rung to leave and the most expensive to stay on, because every improvement amortises nothing. The tenth event costs what the first did, the same loss bucket is rediscovered every eighteen months, and the plant accumulates a folder of successful projects alongside a flat cost curve. Nothing about that changes because a model is added; a model built on this foundation inherits exactly the same unreproducible arithmetic.

In practice

The three points that never arrived

A household-goods plant ran a supplier trial on its highest-volume filling line and reported a three-point OEE gain. The number came from a hand-picked fortnight compared against a nameplate rate nobody had validated since the line was re-tooled. Cost per case did not move that quarter or the next. Eighteen months later a different vendor proposed the same improvement on the same line, and the site had no record that would have let anyone say whether it had already been done.

What it looks like

Line performance is quoted in percentages nobody can reproduce from the MES
Downtime reasons are free text, entered at the end of a shift from memory
Energy is one site meter and a monthly invoice, with no split by line or utility
Every improvement claim rests on a before-and-after week chosen after the result was known

Diagnostic signals you can check this week

Ask two shift teams to define 'planned downtime' and compare the answers
Ask for the raw rows behind the last quoted efficiency gain — if the answer is a spreadsheet, you are here
Check whether any line has a validated nameplate or ideal cycle rate less than two years old
Ask how many kilowatt-hours the plant's largest motor group used last month; if only the site total exists, energy losses are invisible by construction

Anti-pattern · Buying the OEE dashboard first

The instinctive fix is a visibility platform: connect the PLCs, put screens on the floor, watch the numbers. It fails predictably, because a dashboard renders whatever time model and reason taxonomy it is given, and at this rung there is not one. Six months later the plant has real-time percentages that still nobody trusts, and the argument has moved from 'we do not know' to 'the system is wrong'. Agree the time model and the loss taxonomy on one line with paper and a stopwatch first; the software is the easy part afterwards.

What holds you here

There is no agreed time model or loss taxonomy, so no two efficiency numbers on site are comparable and no claim can be reproduced.

Highest-leverage next move

Pick the plant's constraint line and build one loss decomposition against an agreed time model — validated ideal rate, calendar time, planned and unplanned states — before installing any software.

Cost of leaving

Effort: 2–4 months
Team: One controls or process engineer and one production supervisor, part-time
Risk: Low — the work is definitions and metering; nothing in the control layer changes
To next stage: 2–4 months

If this is you, the next step is

A two-week exercise: one line, one time model, one loss taxonomy you can reproduce.

Map one line's losses

Stage 2

Loss-mapped

34% of operators sit here

Losses are decomposed per line against a defined standard, but the loss map is a report: no bucket is bound to a control that can actually move it.

Rung 2 is where most non-automotive plants sit, and it is a genuine achievement: the arithmetic is finally shared. Availability, performance and quality losses are separated against a time model of the kind standards such as ISO 22400 define, energy is split far enough to compute kilowatt-hours per tonne of good output, and the Pareto of losses is stable enough that people argue about priorities rather than about the numbers. The plant knows where its losses are.

What it does not have is a mechanism. Every bucket on the chart is described rather than owned, and none of them is bound to a control — a setpoint, a schedule, a recipe parameter, a maintenance trigger — that could move it without a project. So the loss review becomes a recurring meeting about the same three bars, and the improvement that eventually happens is whichever one somebody had the appetite to fund, not the one with the most recoverable headroom.

Time spent at rung 2 is not neutral, either. Every quarter of accurate measurement without an attributable change teaches the site that the numbers are a reporting exercise, and operators learn that the loss review is something that happens to them rather than something that changes their shift. Plants that sit here for years are frequently harder to move than plants at rung 1, because the measurement is now associated with an absence of consequence.

In practice

The loss review that ran for two years

A speciality-chemicals site held a Monday loss review for two years. Every week the top bar was drying energy on one of two spray dryers, and every week the discussion concluded that inlet temperature was 'set by the product spec'. Nobody had asked whether the spec set a floor on outlet moisture or on inlet temperature — it set the former. Two years of accurate measurement had produced no change because the bucket was never connected to the parameter that governed it.

What it looks like

OEE is computed from MES states against an agreed time model, not from a spreadsheet
Energy is sub-metered at least to the department, and a specific energy consumption figure per unit exists
Downtime reasons come from a fixed taxonomy operators actually use at the HMI
The weekly loss review names the largest bucket — and names the same one again next week

Diagnostic signals you can check this week

For the top three loss buckets, ask which controllable parameter each one is governed by. Blank answers mean the map is not bound to anything
Check whether the largest bucket sits on the plant's constraint or somewhere else — off-constraint effort is the commonest way to be busy and flat
Look for an M&V plan written before any past improvement, not a result written after it
Ask how the ideal cycle rate was set, and when. A rate inherited from the equipment supplier's brochure inflates every performance loss on the chart

Anti-pattern · Chasing the biggest bar on the Pareto

The largest loss bucket is a poor first target twice over. It is often the least movable — planned maintenance, a validated cure time, a regulated hold — and it is frequently off the constraint, where recovered minutes become work-in-progress rather than shipped units. Rank buckets by recoverable headroom against a stated floor and by whether they sit on the bottleneck, not by bar height. The right first bucket is usually mid-sized, on the constraint, and governed by a parameter somebody is already allowed to change.

What holds you here

Nothing binds a loss bucket to a control: the map explains the past and changes no setpoint, schedule, recipe or maintenance trigger.

Highest-leverage next move

Take one bucket on the constraint line, write the measurement-and-verification plan before touching anything, and prove a single change against a matched control.

Cost of leaving

Effort: 4–8 months
Team: Process engineer, data engineer, a named line owner and a finance counterpart
Risk: Medium — the first M&V plan will contradict at least one number the site already reports
To next stage: 4–8 months

If this is you, the next step is

We take your existing loss map and turn one bucket into a controllable, provable change.

Bind one bucket to a control

Stage 3

Attributed

27% of operators sit here

One AI-supported change is proved against a matched control under a written M&V plan, and finance accepts the resulting number into a standard or a budget.

Rung 3 is the first rung where the word 'efficiency' means the same thing to the plant and to the ledger. A change is proposed, the way it will be judged is written down before it is made, a comparable piece of the plant is deliberately left alone, and the difference between them is reported with an honest uncertainty band. The discipline is old and well documented — the international performance measurement and verification protocol has governed energy-savings claims this way for decades — and importing it wholesale is far faster than reinventing it.

The hard part is rarely statistical. It is that a written M&V plan removes the escape routes. Once the boundary, the independent variables and the comparison are fixed in advance, a change either moved the number or it did not, and the first honest 'it did not' is the moment a programme becomes credible. Sites that skip straight to reporting savings accumulate a stack of claims that sum to more than the plant's total cost base, and finance quietly stops reading them.

What still limits rung 3 is durability. The proof is a project artefact — a report, a spreadsheet, a validated setpoint written on a whiteboard by the engineer who ran the trial. It is not in the recipe, the control layer or the standard, so it survives exactly as long as the people and the product mix that produced it. The next changeover, the next campaign, the next shift pattern is where it quietly goes back.

In practice

The number that survived year-end

A pharmaceutical packaging site proved a reduction in unplanned stops on one blister line using a model that flagged feeder faults from vibration and torque signatures. Because the M&V plan named a sister line as the control and specified an eight-week alternating design, the finance business partner accepted the delta into the site's standard cost at year-end. It was the first efficiency number that plant had ever booked rather than reported — and the first one that survived the auditors' questions about what would have happened anyway.

What it looks like

An M&V plan exists before the trial — boundary, independent variables, test and acceptance criteria
A matched control carries the comparison: a parallel line, a period-matched baseline model, or an alternating-week design
The result is expressed in the plant's own units — kWh per tonne, cost per case, good units per shift — with an uncertainty band
Finance has accepted at least one such number into a standard cost, an energy budget or an EnPI

Diagnostic signals you can check this week

Ask to see the M&V plan for the most recent improvement, dated before the change
Ask what the control was. 'The same line last quarter' is a baseline, not a control, unless the model normalises for volume, mix and ambient conditions
Ask whether any proved gain has been rejected. A programme with a 100% success rate is not measuring
Check whether the accepted number changed a standard, a budget line or an EnPI — or only a slide

Anti-pattern · Booking the same saving twice

The proved gain gets counted once in the project's business case and again in next year's budget, because the standard was never re-set. The plant now expects a cost base it has already spent, and when the numbers do not reconcile the improvement programme is blamed for the gap. Re-setting the standard at the moment finance accepts the number is unglamorous bookkeeping that protects the whole programme's credibility — and it is the step most often skipped, because the project is finished by then and nobody owns it.

What holds you here

The proof lives in a report rather than in the recipe, the setpoint or the standard, so it decays at the next changeover, campaign or shift-pattern change.

Highest-leverage next move

Encode the proved setting in the recipe or control layer under change control, re-set the standard, and replicate to a second line to find out what the copy actually costs.

Cost of leaving

Effort: 6–12 months
Team: Process engineer, ML engineer, automation engineer, named line owner, finance partner
Risk: Medium — the first write into a recipe or setpoint needs change control and a drilled fallback
To next stage: 6–12 months

If this is you, the next step is

Getting a validated setting out of a report and into the recipe, under change control.

Encode a proved gain

Stage 4

Held

14% of operators sit here

Proved gains are encoded in the recipe and control layer, survive changeover and mix, and the standard is re-set — so the second line costs a fraction of the first.

At rung 4 the marginal cost of the next point collapses, and it collapses for an unglamorous reason: the plant has stopped re-deriving things. The loss taxonomy, the time model, the M&V template, the change-control route into the recipe and the shape of a matched control are all reusable, so a new bucket on a new line is mostly specification — which parameter, which control, which acceptance test. Engineering time moves from plumbing to argument about targets, which is the healthier argument.

The distinguishing discipline is persistence rather than sophistication. A gain is only held if the setting that produced it is in a versioned artefact the changeover procedure reads, if the standard moved with it, and if somebody is watching for the bucket to come back. Product mix is the usual solvent: a setting proved on the long-running SKU quietly under-performs on the short campaign, and without per-SKU monitoring the average hides it for a quarter.

The constraint that emerges here is knowledge of the remaining headroom. A rung-4 plant is good at proving and holding gains, and it still sets next year's target by peer benchmark or by last year plus a percentage. That method cannot distinguish a bucket that is close to its physical floor from one that has never been touched, so effort is allocated by narrative rather than by entitlement, and the programme spends its best engineers on the hardest remaining percent.

In practice

The second line that cost a fifth

A plastics processor proved a cycle-time gain on one injection-moulding cell by predicting the point at which the part was cool enough to eject from mould-temperature and melt-history signals rather than from a fixed timer. The setting went into the mould's recipe record under change control, the standard cycle was re-set, and the plant kept per-mould monitoring. Copying it to the second cell took three weeks: two days of engineering, the rest spent revalidating the part on the new tool.

What it looks like

Optimised setpoints live in the recipe or control layer under change control, not in an operator's memory
The standard rate, standard cost or EnPI is re-set the moment a gain is accepted
A second line reproduced the gain in weeks, mostly by configuration rather than by a new project
Erosion of a recovered bucket raises an alarm during the month, not a surprise at quarter end

Diagnostic signals you can check this week

Ask where the optimised setpoint physically lives, and whether the changeover procedure reads it
Compare the elapsed time and cost of the last three replications; if they are not falling, nothing is being reused
Ask whether the standard was re-set when the last gain was accepted, and who did it
Ask how a target for next year was set. 'Benchmark' or 'last year plus three' means entitlement is unknown

Anti-pattern · Freezing the setting and calling it finished

A validated setpoint gets locked into the recipe and never revisited, which looks like discipline and behaves like decay. Heat exchangers foul, filters load, motors age, tooling wears and raw-material specification drifts, so the setting that was optimal at commissioning becomes conservative within a year and unsafe-for-quality within three. Hold the gain by versioning the setting and scheduling its revalidation against the asset's condition — held is not the same as frozen.

What holds you here

Nobody can say how much is left: targets come from peer benchmarks or last year plus a percentage, so a nearly exhausted bucket looks identical to an untouched one.

Highest-leverage next move

Build a per-line entitlement model — thermodynamic floor, state of the art, practical minimum — and re-express every remaining target as headroom against it.

Cost of leaving

Effort: 12–18 months
Team: Platform and automation engineers, process engineering, quality, plus a standing change-control route
Risk: Higher — writes into recipes and setpoints touch quality and, in regulated plants, validated state
To next stage: 12–18 months

If this is you, the next step is

Floor, state of the art and practical minimum per line, so targets stop being guesses.

Build the entitlement model

Stage 5

Entitlement-bound

4% of operators sit here

Every line runs against a maintained, physics-anchored entitlement model, so what remains is bounded by capital, certification and regulation rather than by knowledge.

Rung 5 is narrower and less romantic than the phrase suggests. It is not a self-running factory. It is a plant that can say, for each line and each loss bucket, how much is theoretically available, how much is available with technology that exists, how much of that is reachable without a requalification, and what the remaining gap would cost in capital or in regulatory work. Efficiency stops being an aspiration and becomes a register with owners and dates.

The work at this rung is mostly maintenance of that register. Floors move when the product changes; the state of the art moves when somebody else demonstrates it; practical minimum moves when an R&D technology becomes purchasable. A register that is not re-derived is worse than none, because it converts a stale assumption into an authoritative-looking number and quietly retires opportunities that reopened two years ago.

Autonomy appears here, and it appears narrowly. Continuous setpoint adjustment inside proved bounds on a well-instrumented utility is defensible; unattended change to a validated pharmaceutical process is not, and will not be by 2030, because the binding constraint is regulated change control rather than model capability. Plants that understand this stop arguing about how clever the model is and start arguing about which decisions are eligible — which is the argument that actually determines the 2030 number.

In practice

The register that closed an argument

A food manufacturer's utilities team was asked for another five per cent on steam. The entitlement register showed the boiler house already within a few per cent of its state of the art, with the remaining gap sitting in a heat-recovery retrofit costed at eight figures — while the pasteuriser hold time, never examined, was carrying a much larger recoverable loss governed by a validated food-safety parameter. The register turned a target negotiation into a capital-and-certification decision, which is what it always was.

What it looks like

Each major line has a documented floor, state of the art and practical minimum, with the assumptions written down
Remaining headroom is stated per bucket and reviewed like a risk register, not like a wish list
Residual opportunities are sequenced by what they cost to certify, not by what they cost to build
Autonomous adjustment runs inside stated bounds on the buckets where the evidence supports it, and escalates outside them

Diagnostic signals you can check this week

Ask when the entitlement model was last re-derived and against what published reference
Ask which residual opportunities are blocked by capital and which by certification — a plant at this rung can separate them
Check whether autonomous adjustment is bounded by written limits, and whether the escalation rate is monitored
Ask whether any opportunity was reopened in the last year because the state of the art moved

Anti-pattern · Letting the register become a spreadsheet nobody re-derives

The entitlement model is built once, beautifully, by an engineer who then moves on. Two years later it is still quoted in target-setting meetings with assumptions — product mix, ambient conditions, available technology — that no longer hold. Its authority is now doing damage, because it forecloses opportunities that have reopened. Give the register an owner, a review cadence and a rule that every quoted floor names the published source and the date it was taken from.

What holds you here

The remaining headroom needs capital, a requalification or a regulatory change — engineering is no longer the constraint on the number.

Highest-leverage next move

Treat the entitlement register as a governed artefact with an owner and a review cadence, and sequence the residual by what it costs to certify rather than by what it costs to build.

Cost of leaving

Effort: Continuous
Team: Process engineering, energy management, quality and finance, with a standing review forum
Risk: Concentrated — the residual is capital-intensive and, in regulated plants, requalification-bound

If this is you, the next step is

We re-derive two lines' floors against published references and mark what has moved.

Stress-test an entitlement register

Where non-automotive plants actually sit on the ladder

The distribution across the five rungs, and why the rung 2 → 3 step is the one most programmes never take.

Most non-automotive plants are at rung 2 — they can measure their losses and cannot yet prove that anything they did moved them. The distribution below is weighted heavily toward that rung: a clear majority of sites have a working loss map, a minority have proved a single change against a control, and very few have a maintained entitlement model that tells them how much is left.

Distribution of non-automotive plants across the five rungs

Illustrative distribution — a model-derived synthesis, not a survey. It is shaped to be consistent with the staging in acatech's Industrie 4.0 Maturity Index, the small designated population of the World Economic Forum's Global Lighthouse Network, and the persistent adoption-versus-impact gap reported in MHI's annual industry survey. Treat the shape as the argument and the exact percentages as illustrative.

Share of plants

21% — 1 · Anecdotal
34% — 2 · Loss-mapped (the plateau)
27% — 3 · Attributed
14% — 4 · Held
4% — 5 · Entitlement-bound

Source: Illustrative synthesis, anchored to acatech, WEF Global Lighthouse Network and MHI research

The shape is not a manufacturing peculiarity. Cross-industry research has consistently described a wide gap between organisations experimenting with AI and organisations reporting material bottom-line impact — see McKinsey's State of AI series (opens in a new tab) — and MHI's annual industry report (opens in a new tab) tracks the same adoption-versus-impact divergence across material handling and supply chain. What is specific to a factory is where the gap sits: not in the model, and not usually in the data, but in the absence of a comparison that would let anybody say the change worked.

The step from rung 2 to rung 3 is where the distribution thins, and it is the cheapest step on the ladder in engineering terms and the most expensive in organisational ones. Nothing technical is required beyond a control group and a plan written in advance. What is required is the willingness to run a change that might turn out not to have worked, and to say so — which is a governance decision, not a capability.

The 2030 efficiency ledger: every loss bucket and its floor

The centrepiece of this page. Eight loss buckets, what sets the floor under each, what AI moves today, what is credibly available by 2030 — and the overclaim attached to each one.

The remaining efficiency in a non-automotive plant sits in eight buckets, and each one has a different kind of floor beneath it. Some floors are thermodynamic and absolute; some are set by process capability and measurement uncertainty; some are set by a validated specification that can only move through a regulator; and one — the constraint — is set by where the bucket sits in the flow, because minutes recovered away from the bottleneck do not convert into throughput at all. The ledger below is how we scope efficiency programmes: bucket, floor, what is genuinely available now, what is credible by 2030, and the claim that is routinely made instead.

Loss bucket	What sets the floor	What AI moves today (2026)	Credible by 2030	The overclaim
Unplanned downtime	Failure physics and the spares and labour you can actually stage; some failures give no usable warning	Anomaly detection on vibration, motor current and torque signatures, feeding the CMMS as a scheduled job	Most warnable failure modes caught with enough lead time to become planned work	"Zero unplanned downtime" — unwarnable and random failures do not disappear
Changeover and set-up	The mechanical work that must physically happen, plus any validated clean or flush between products	Sequencing and campaign optimisation that reduces the number of changeovers, and guided set-up that reduces variance between operators	Changeover duration close to the best operator's time, every time, on every shift	"Instant changeover" — the clean-down and the qualification run are process, not scheduling
Minor stops and speed loss	Infeed variability, material specification spread and the machine's stable operating window	Predicting jam-prone conditions from upstream signals and adjusting infeed, plus per-SKU rate targets instead of one line rate	The recovery of most of the gap between the line's best hour and its average hour	"Run above nameplate" — running above the validated stable window buys stops later
Scrap, rework and giveaway	Process capability and measurement uncertainty: you cannot target a fill closer to the limit than you can measure it	In-line vision and sensor inspection that catches defects at the station that caused them, and fill-target control against real-time weight variance	Giveaway pulled close to the legal or specification limit plus measured uncertainty	"Zero defects" — capability and gauge error set a floor no model reaches below
Thermal and drying energy	Thermodynamics: latent heat, reaction enthalpy, and the minimum work of separation for the product you actually make	Soft sensors for moisture and quality that let setpoints track the true target instead of a conservative margin	Operation near the state of the art for the installed asset; the practical minimum needs new equipment	"AI cuts energy 30%" — quoted without saying from which band, or on what asset
Motive and compressed-air energy	Motor and compressor efficiency classes, and the isothermal work of compression for the air you genuinely need	Pressure-setpoint scheduling, compressor sequencing, and leak and artificial-demand detection from off-shift flow signatures	Header pressure held at the true requirement rather than at a historical margin, continuously	Crediting a model for savings that came from a leak survey nobody ran until the project started
Constraint and scheduling loss	The bottleneck's capacity, plus the variability the schedule must absorb to keep it fed	Finite-capacity scheduling that protects the constraint, and sequencing that reduces changeover load where it matters	A schedule that keeps the constraint fed under realistic variability, re-planned within the shift	Reporting plant-wide OEE gains as throughput when the constraint never moved
Quality release and documentation	Regulated review, sampling and release requirements — GMP, food-safety verification, product certification	Automated review-by-exception on batch records, and anomaly triage that shortens investigation time	Review-by-exception on routine batches, with human review concentrated on flagged ones	"Real-time release for everything" — release scope is a regulatory decision, not a modelling one

The 2030 efficiency ledger for a non-automotive plant. 'Credible by 2030' assumes existing technology deployed well, not a new class of model. Floors are stated as the kind of limit that binds, not as a universal number — the value is plant-specific and belongs in your own entitlement register.

Two rows of that table deserve to be read together, because they are the most common source of double counting. Motive energy and thermal energy are both usually attacked with setpoint models, and both are usually preceded by a physical clean-up — a leak survey, a steam-trap audit, an insulation repair — that produces a large part of the measured saving. If the M&V boundary does not separate those, the model is credited with work a maintenance technician did, and the next site's business case is built on a number that will not reproduce.

Rank by recoverable headroom, not by bar height
A bucket's size on the Pareto tells you what it costs today, not what you can get back. Multiply each bucket by an honest recovery fraction against its floor before ranking, and the order usually changes — the largest bar is often the one closest to its limit.
Ask which floor binds before asking which model helps
A thermodynamic floor is absolute, a capability floor moves with better measurement, and a certification floor moves only with a regulator. Those three demand completely different work: engineering, metrology and regulatory affairs respectively. Deciding which one binds takes an afternoon and saves quarters.
Separate the constraint from everything else
Recovered minutes on the bottleneck become units shipped. Recovered minutes anywhere else become work-in-progress, a busier plant and, occasionally, a worse one. Any efficiency claim expressed in plant-average OEE rather than in constraint throughput should be treated as unproven until the constraint's number is shown.
Price the residual in certification, not just capital
In regulated plants the remaining headroom is often reachable technically and blocked procedurally. A change to a validated parameter costs a qualification exercise; a change to a release strategy costs a regulatory submission. Putting those costs in the ledger converts an argument about ambition into a sequencing decision.

The energy rows can be anchored more precisely than the others, because the US Department of Energy's bandwidth studies (opens in a new tab) publish exactly this decomposition per sector — current typical, state of the art, practical minimum and thermodynamic minimum — and the Industrial Efficiency and Decarbonization Office (opens in a new tab) maintains the underlying analysis. The IEA's industry programme (opens in a new tab) provides the equivalent view at sector level internationally. Between them, a plant can place its own kilowatt-hours per tonne on a published scale rather than against a sister site that may itself be far from good.

What 2030 actually changes — and what it does not

Six claims made about the AI factory of 2030, separated into what runs in production today, what published work actually claims, and what remains speculation.

Very little of what is promised for 2030 is a new capability; most of it is today's capability with the deployment problem assumed away. Closed-loop setpoint control, in-line vision inspection, predictive maintenance, finite-capacity scheduling and soft sensors all run in production plants now, and have for years. What is genuinely scarce is the metering, the loss model, the change-control route and the measurement discipline that let a plant book what those techniques produce — which is why the honest 2030 forecast is mostly a forecast about present-readiness.

The 2030 claim	Real today (2026)	What would have to be true by 2030	Honest verdict
Lights-out factories become normal	Lights-out cells and shifts exist in high-volume, low-mix, well-fenced operations — mostly machining, moulding and electronics test	Material presentation, changeover, deviation handling and maintenance would all have to be automated for a high-mix plant, and the safety case rewritten for each	Real but narrow. Growth continues in the operations already suited to it; general high-mix lights-out is speculation
Plants self-optimise end to end	Closed-loop optimisation runs on bounded, well-instrumented sub-systems: utilities, individual units, single lines	A validated plant-wide model with enough sensing to constrain it, plus an accepted safety and quality case for unattended cross-unit change	Plausible per sub-system, speculative plant-wide. The sensing and the safety case, not the optimiser, are the binding constraints
Generative models write process recipes	Language models draft work instructions, summarise deviations and help engineers search process history	A recipe change is a validated change; a generated recipe would still need the same qualification as a human-written one	Useful for drafting and search today. It changes who writes the document, not what it costs to approve it
Digital twins replace physical trials	Calibrated first-principles and hybrid models genuinely replace some trials — heat and mass balance, flow, scheduling what-ifs	Model validity would have to be demonstrated across the operating envelope the trial was covering, which is itself a measurement programme	Real where the physics is well characterised and the model is calibrated against the specific asset; overstated as a general replacement
Agents run the production schedule	Finite-capacity scheduling with automated re-planning inside defined rules is ordinary practice	Autonomy over commercial trade-offs — which customer waits — needs an owner, a policy and an audit trail more than it needs a better planner	Mostly a governance question. The scheduling technology is not the limitation and has not been for years
Foundation models for industrial signals	Pre-trained time-series and vision models measurably shorten the cold-start on a new asset or a new defect class	Transfer would have to hold across genuinely different assets and product mixes, and be demonstrable to a quality function	Promising and actively researched; treat published transfer results as evidence about the studied assets, not about yours

Separating the three classes of statement. 'Real today' means running in production plants now; 'what would have to be true' names the specific blocker, which in five of six rows is not a modelling problem.

Read down the third column and the pattern is hard to miss: five of the six blockers are sensing, validation, safety case or governance. Only one — transfer learning across assets — is a genuine modelling frontier. That is the strongest argument available for treating future-readiness as present-readiness, and it is the reason a plant that spends 2026 fixing its stop-reason taxonomy and its sub-metering will be better placed in 2030 than one that spends the same money on a platform.

The thermodynamic minimum is the least amount of energy required under ideal conditions, which typically cannot be attained in commercial operations.
US Department of Energy, manufacturing energy bandwidth studies

That sentence is the discipline of this whole page in one line. There is a floor, it is calculable, it is not reachable, and the useful question is how far above it you are running and what the next band costs. A 2030 target expressed as a percentage improvement over today says nothing about any of that. A target expressed as headroom against a named band — state of the art for this asset class, practical minimum with technology under development — can be argued about, costed and sequenced.

Energy bands DOE places every manufacturing process between

US Department of Energy

5 of 6

Claims in the table above whose blocker is sensing, validation or governance rather than modelling

Counted from the table above

1 of 6

Genuine modelling frontier among them — transfer across different assets and product mixes

Counted from the table above

What booked efficiency looks like in public

Three publicly reported programmes at non-automotive manufacturers, read against the ladder. None is an Atomic Loops engagement — each links to the operator's own published material.

The clearest public evidence for the ledger thesis is what large non-automotive manufacturers chose to report. In each case below the reported outcome is expressed in the plant's own operating units — breakdowns, cycle time, doses per batch, productivity — rather than in model accuracy, and in each case the improvement was tied to a specific loss bucket rather than to a general capability.

Three programmes read against the ladder

Outcomes as reported by the operators themselves; verify figures against the linked source before reusing them, as we have not independently audited them. Images are illustrative industrial scenes from a generated library, not operator photography, and imply no endorsement.

Illustration of a snack-food production line with inspection cameras and analytics overlays

PepsiCo (Frito-Lay)Global food and beverage manufacturer · high-volume snack plants24

Challenge: Two different loss buckets at once: unplanned mechanical stoppages across a large plant estate, and product-quality variation that could only be sampled off-line, after the conditions that caused it had passed.
Approach: PepsiCo has publicly described monitoring plant machinery to predict mechanical failures before they occur, and separately an AI system trained to assess individual product pieces on the line — curvature, density and puffiness on a Cheetos line — and adjust process conditions in response rather than reporting the variation afterwards.
Reported outcome: PepsiCo reports that after one year, the plants concerned saw zero unexpected breakdowns or interruptions, with mechanics able to concentrate on planned rather than emergency maintenance. The in-line quality system was described as being tested in Spain with wider rollout planned.; What it shows about the curveBoth interventions bound a named loss bucket to a control that could move it — maintenance scheduling for downtime, process conditions for quality. Neither is reported as a model metric, which is why the numbers are usable.

PepsiCo — Artificial intelligence at PepsiCo (opens in a new tab)

Illustration of a pharmaceutical cleanroom filling line with process analytics displayed above it

PfizerGlobal pharmaceutical manufacturer · regulated GMP production34

Challenge: Scaling manufacturing output under regulated change control, where every process adjustment is a validated change and the usual efficiency levers are constrained by qualification rather than by engineering.
Approach: Pfizer has published that it applied data and AI to the manufacture of PAXLOVID, and separately that it deployed a Digital Operations Center giving an end-to-end view of manufacturing, used to predict issues and adjust operations during production rather than after it.
Reported outcome: Pfizer reports reducing the cycle time of a critical step in the supply chain by 67%, enabling the production of 20,000 extra doses per batch — an outcome expressed in units the business books rather than in model performance.; What it shows about the curveThe most tightly regulated plants can still book large gains, provided the target is a loss bucket whose governing parameter is inside the qualified envelope. The ceiling here is change control, not capability — exactly the ledger's certification-bound row.

Pfizer — data and AI are helping to get medicines to patients faster (opens in a new tab)

Illustration of an electronics assembly line with a robot-mounted inspection camera scanning circuit boards

Nokia (Oulu)Telecoms equipment manufacturer · 5G base-station assembly35

Challenge: Holding productivity and quality while ramping a high-mix electronics assembly operation, where per-line improvements historically did not replicate across the factory.
Approach: Nokia has published that its Oulu factory runs on private wireless connectivity across factory assets, IoT analytics on edge cloud and a real-time digital twin of operations data — infrastructure that makes the same measurement and control available on every line rather than on a pilot line.
Reported outcome: Nokia reports more than 30% productivity improvement, 50% savings in time of product delivery to market and annual cost savings of millions of euros at the site, which the World Economic Forum designated an advanced Fourth Industrial Revolution lighthouse.; What it shows about the curveThe rung-4 signature is visible here: the investment was in the shared measurement and connectivity layer, so improvements replicated instead of being re-engineered per line. That is what turns a series of gains into a compounding one.

Nokia — Oulu 5G factory recognised as an advanced lighthouse (opens in a new tab)

All three sites sit inside the World Economic Forum's Global Lighthouse Network (opens in a new tab) orbit or its equivalent public reporting, which is worth reading for what it does and does not prove. It is evidence that these approaches produce reported operating gains at real plants; it is not evidence about the median plant, because designation selects for sites that measured and published. The useful thing to extract from a lighthouse is the sequence — what was instrumented, what was bound to a control, what was replicated — rather than the headline percentage.

The efficiency stack, layer by layer

What actually has to exist before a point can be booked and held — and which layer a plant can legitimately defer.

Booking an efficiency point requires five layers, and the order in which they are built decides whether a programme compounds or resets. The stack below is deliberately vendor-neutral: each layer is defined by what it has to guarantee, not by which product supplies it, because the same guarantees are met by a historian and a PLC in one plant and by a cloud platform in the next.

Layers required by rung

Each layer is annotated with the rung that first requires it. A plant trying to reach rung 3 without the attribution layer is running a rung-2 loss review with a model attached.

Sensing and metering
Stage 1+
- Line counters and machine statesFrom the PLC, with a taxonomy operators use
- Utility sub-meteringElectricity, gas, steam, compressed air — per line where it matters
- Quality measurementIn-line gauges, checkweighers, vision, with known gauge error
→
Contextualised loss model
Stage 2+
- Agreed time modelCalendar, scheduled, planned and unplanned states, defined once
- Loss taxonomy bound to parametersEach bucket names the control that governs it
- Normalised energy indicatorSEC and EnPI per unit, corrected for mix and ambient
→
Model and recommendation
Stage 3+
- Predictive and soft-sensor modelsFailure warning, moisture, quality, demand
- Optimisation under constraintsSetpoints and schedules inside the validated envelope
- Uncertainty exposedA recommendation without a confidence is not actionable
→
Attribution and standards ledger
Stage 3+
- M&V plan templateBoundary, independent variables, test, written first
- Matched control designParallel line, alternating weeks or normalised baseline
- Standard re-set routeWho moves the standard cost, budget or EnPI, and when
→
Control binding and governance
Stage 4+
- Recipe and parameter write-backUnder change control, read by the changeover procedure
- Fallback and boundsPrevious value one switch away; limits enforced below the model
- Entitlement registerFloor, state of the art, practical minimum, with owners (rung 5)

Pipeline described

Sensing and metering (stage 1+) — Line counters and machine states: From the PLC, with a taxonomy operators use; Utility sub-metering: Electricity, gas, steam, compressed air — per line where it matters; Quality measurement: In-line gauges, checkweighers, vision, with known gauge error
Contextualised loss model (stage 2+) — Agreed time model: Calendar, scheduled, planned and unplanned states, defined once; Loss taxonomy bound to parameters: Each bucket names the control that governs it; Normalised energy indicator: SEC and EnPI per unit, corrected for mix and ambient
Model and recommendation (stage 3+) — Predictive and soft-sensor models: Failure warning, moisture, quality, demand; Optimisation under constraints: Setpoints and schedules inside the validated envelope; Uncertainty exposed: A recommendation without a confidence is not actionable
Attribution and standards ledger (stage 3+) — M&V plan template: Boundary, independent variables, test, written first; Matched control design: Parallel line, alternating weeks or normalised baseline; Standard re-set route: Who moves the standard cost, budget or EnPI, and when
Control binding and governance (stage 4+) — Recipe and parameter write-back: Under change control, read by the changeover procedure; Fallback and bounds: Previous value one switch away; limits enforced below the model; Entitlement register: Floor, state of the art, practical minimum, with owners (rung 5)

Step-by-step insights

Sensing and metering — the layer whose absence is invisible: Missing sensing does not announce itself; it shows up as a loss bucket that never appears on the Pareto because nothing measures it. Compressed air is the classic case: without a flow meter at the header, the entire artificial-demand and leak bucket is folded into the site electricity total and no analysis will ever find it. Before any modelling work, list the buckets in the ledger and ask which instrument would have to exist for each to be measurable. The gaps in that list are the first capital request, and it is usually a small one.
The time model — one definition, agreed once, defended forever: Almost every irreproducible efficiency number traces back to an unstated time model. Standards such as ISO 22400 exist precisely to fix which time counts as scheduled, planned and unplanned, and adopting one — any one — consistently matters far more than which one you pick. Publish the definition, put it on the wall next to the board, and require every quoted percentage to state the model it was computed under. It costs an afternoon and removes the most common way for two accurate numbers to contradict each other.
Models — the layer plants over-invest in first: This is the layer everyone wants to start at and the one with the least leverage in isolation. A moderately accurate failure warning that lands as a scheduled job in the CMMS prevents more downtime than an excellent one that lands in an inbox, and a soft sensor that tracks outlet moisture is only worth building once somebody is allowed to change the inlet temperature it would inform. Sequence models after the loss model that names their target and the change-control route that lets their output do something.
Attribution — the layer that decides whether any of it counts: The attribution layer is a template and a habit rather than a technology: an M&V plan written before the change, a control that stays on the old setting, and a route by which an accepted number moves a standard. Its absence is why so many plants have a folder of successful projects and a flat cost curve. Building it is nearly free and reliably unpopular, because it is the layer that occasionally reports that a change did nothing — which is precisely the property that makes the other reports believable.
Control binding — where safety, quality and efficiency meet: Writing an optimised value into a recipe or setpoint crosses into territory with its own rules. Safety interlocks and instrumented functions stay in the safety layer, below and independent of any model, and hard limits belong in the PLC rather than in the recommendation. In a regulated plant, moving a parameter into the recipe is a change-control event with a qualification cost, which is far better discovered while designing the trial than after it succeeds. The fallback — the previous value, one switch away, drilled once on a quiet shift — is what usually unlocks the approval.

The layer most often skipped is attribution, and skipping it is what produces the characteristic rung-2 plant: excellent instrumentation, capable models, and an improvement history nobody outside the team believes. The layer most often built too early is the model layer. If you have budget for exactly one thing this year and you are at rung 2, buy the metering that makes an invisible bucket visible and the template that makes a claim checkable — in that order.

Interoperability deserves a note, because it decides what the stack costs to build twice. Machine data reaching the loss model through OPC UA (opens in a new tab) rather than through a bespoke driver per asset is the difference between a second line taking three weeks and taking a project, and the ISA-95 (opens in a new tab) level model is still the cheapest shared language for arguing about where a function belongs. Neither is glamorous and both are load-bearing for replication.

A 90-day plan: booking a compressed-air saving on one plant

The rung 2 → 3 transition made concrete on the loss bucket almost every non-automotive plant carries and almost none has metered — compressed air. Contains no model development in the first six weeks.

Moving one rung takes about 90 days when it is scoped to a single loss bucket, and several years when it is scoped to a plant. To make that concrete, the plan below runs the transition on compressed air: the utility that typically consumes a substantial share of a plant's electricity, is almost never metered at the header, and whose losses divide cleanly into four buckets that can be separated with data most sites already have. It is chosen deliberately because it touches no product quality parameter, so the change-control path is short even in a regulated plant.

Rung 2 → rung 3 on compressed air, in one quarter

One compressor house, one header, one named owner. If a phase needs longer than its window, narrow the scope — one shift pattern, one production area — rather than extending the plan.

Days 1–15
Write the M&V plan, then set the boundary
Before any analysis: define the measurement boundary (compressor house electricity plus header flow), the independent variables that will normalise it (production tonnage by product family, ambient temperature and humidity, shift pattern), the comparison design, and the acceptance criteria. Install or commission a header flow meter and compressor-level power metering if they do not exist. Name the site utilities engineer as owner and agree the finance counterpart who will accept the number.
A dated M&V plan and a metered boundary
Days 16–45
Decompose the bucket before modelling anything
Split the compressor house's kilowatt-hours four ways using data you now have: air genuinely used at end-use, artificial demand (header pressure held above the highest true requirement), leakage (flow during a planned no-production window), and control loss (compressors running unloaded because of sequencing). Walk the three largest end-uses and confirm their actual pressure requirement rather than the one on the drawing. Fix nothing yet — this phase is measurement, and its output is the split.
A four-way split of the bucket, with the leak rate measured
Days 46–70
Put the recommendation where the decision is made
Deploy a pressure-setpoint schedule and compressor-sequencing recommendation into the setpoint field the utilities operator already uses in the SCADA or energy management system — not a separate screen. The operator approves each change; hard pressure limits and safety interlocks remain in the PLC below the model; the fixed historical setpoint stays one switch away as the fallback and is exercised once, deliberately, on a quiet shift. Log every acceptance and override with conditions.
Recommendations live, with approval, limits and a drilled fallback
Days 71–90
Prove it, book it, re-set the standard
Run the comparison the plan specified — an alternating-week design or a normalised baseline regression using the agreed independent variables — and report kilowatt-hours per tonne with an uncertainty band. Separate the saving attributable to the setpoint schedule from any saving produced by the leak repairs, so the model is not credited with a technician's work. Take the number to the finance counterpart, book it into the site energy budget, and re-set the EnPI so it cannot be claimed again next year.
A booked kWh-per-tonne delta and a re-set standard

The order that keeps it honest

The M&V plan comes before the model
Writing the boundary, the independent variables and the acceptance test first is what converts the result from an assertion into evidence. The protocol for doing this is published — the IPMVP (opens in a new tab) has governed energy-savings claims this way for decades — and adopting the template costs nothing.
Kill artificial demand before optimising the setpoint
Repairing leaks and correcting a header pressure held above the true requirement will produce most of the early saving. Do it, measure it separately, and attribute it separately. A business case for the next site built on a model credited with a leak survey will not reproduce, and everyone will conclude the model was oversold.
Limits stay below the model
Minimum pressure for safety-critical end-uses, compressor protection interlocks and any process-critical floor belong in the PLC and the safety layer, enforced independently of whatever the model recommends. A recommendation that can be clipped by a hard limit is a recommendation a change-control board will approve.
Re-set the standard, or the saving gets spent twice
The moment finance accepts the number, the energy budget and the EnPI move with it. Skipping this is how a plant ends up expecting a cost base it has already banked, and how the same saving appears in two consecutive years' improvement totals.

The reason this plan generalises is that its structure is bucket-agnostic. Substitute drying energy, minor stops or fill giveaway and the four phases keep their shape: define how it will be judged, decompose before modelling, put the recommendation where the decision already happens, prove it against something that did not change. Only the instrumentation in phase one and the change-control path in phase three vary — and in a regulated plant, phase three is where the extra time goes.

Verifying a point: the measurements that decide the argument

Where each verification metric comes from — the formula, the source system, the cadence, and the rung at which it first means something.

A verification metric you cannot name a source system for is an opinion with a decimal place. Every measure below reduces to counts, timestamps and meter readings that the PLC, MES, historian or energy management system already records; the work is joining and normalising them, not creating them. The table is the build sheet, and the right-hand column is the honesty check — the rung at which each metric first measures something real rather than something assumed.

Measure	Formula / read	Source	Cadence	Honest from
Constraint throughput	Good units through the bottleneck ÷ scheduled time on the bottleneck	MES states + line counters	Per shift	Rung 2
Loss split by bucket	Time or units lost per taxonomy code ÷ total loss	MES stop records against the agreed time model	Weekly	Rung 2
Specific energy consumption	kWh (or MJ) ÷ tonnes of good output, normalised for mix and ambient	Sub-meters + MES production records	Daily, reviewed monthly	Rung 2
Baseline model fit	R² and CV(RMSE) of the pre-change energy or rate model against its independent variables	M&V workbook	Once per plan, re-checked quarterly	Rung 3
Attributed delta	Post-change actual − baseline model prediction, with uncertainty band	M&V workbook + meters	Per trial, then monthly	Rung 3
Recommendation acceptance	Accepted recommendations ÷ recommendations shown	Approval log in the SCADA or MES	Weekly	Rung 3
Gain retention after changeover	Performance on the improved parameter, first full run after each changeover ÷ proved level	MES per-SKU records	Per changeover	Rung 4
Replication cost	Engineering days to reproduce the gain on line n ÷ days on line 1	Delivery tracker	Per replication	Rung 4
Headroom remaining	(Current performance − stated floor) ÷ (original performance − stated floor)	Entitlement register	Quarterly	Rung 5

Instrumentation build sheet for verifying an efficiency gain in a non-automotive plant. 'Honest from' is the rung at which the metric stops being an assumption.

Two of those rows carry most of the weight. Baseline model fit is what stops a normalised comparison from quietly becoming a guess: a baseline that explains little of the historical variation cannot detect a small saving, and a plan that discovers this after the trial has run has wasted a quarter. Gain retention after changeover is the metric that separates rung 3 from rung 4, and almost nobody measures it — which is precisely why so many booked savings evaporate between the project close-out and the following year's budget.

Rung 3 readiness checklist

If you cannot tick all seven, the next efficiency number your plant produces will be a claim rather than a booking, regardless of how good the model is. Tick as you go — this list works without JavaScript.

An M&V plan exists, dated before the changeBoundary, independent variables, comparison design and acceptance criteria.
Something comparable is deliberately left unchangedA parallel line, an alternating-week design, or a validated normalised baseline.
The loss bucket names the parameter that governs itAnd names the person allowed to change that parameter.
The measure is expressed in plant units, with an uncertainty band
Hard limits and interlocks sit below the modelIn the PLC and safety layer, enforced independently of the recommendation.
A fallback to the previous setting has been exercised, not just documented
A finance counterpart has agreed in advance to accept the numberIncluding agreeing what result would mean the change did not work.

0 of 7 ticked

Nothing ticked — start with the plan, not the platform

A blank list is common at rung 2 and it is not a tooling problem. Pick the single bucket on your constraint line that you most want to move, write the M&V plan for it this week, and only then decide what needs measuring. The compressed-air plan above is the cheapest version of this because it touches no quality parameter.

Scope the 90 days with us

The bounds that set the 2030 ceiling

Six limits that no model moves: thermodynamics, process capability, the constraint, the safety layer, regulated change control and asset physics.

The ceiling on manufacturing efficiency in 2030 is set by six bounds, and only one of them is technological. Knowing which bound is currently binding on a given bucket is the most useful thing an efficiency programme can establish, because each bound demands entirely different work — a thermodynamic bound needs a different process, a capability bound needs better measurement, a certification bound needs a regulatory submission, and the constraint bound needs the bottleneck to move. Conflating them is how programmes spend engineering effort on problems that are not engineering problems.

Bound	What it is	What it forbids	What it still allows
Thermodynamic	The minimum energy the physical transformation requires — latent heat, reaction enthalpy, minimum work of separation	Any claim of energy per unit below the theoretical minimum, however the model is described	Closing the gap between current typical practice and the state of the art for the installed asset, which is usually large
Process capability	The spread of the process plus the uncertainty of the measurement used to control it	Targeting a fill weight, dimension or concentration closer to a limit than the gauge can resolve	Reducing the spread itself, and reducing gauge error — both of which move the achievable target
The constraint	The bottleneck's capacity, and the buffers the schedule needs to keep it fed	Converting recovered minutes elsewhere into shipped units; those minutes become work-in-progress	Elevating or protecting the constraint, and reducing the variability the schedule has to absorb
The safety layer	Interlocks and instrumented protective functions, independent of the control layer by design	Any model output that would override, bypass or degrade a protective function	Optimisation inside the envelope those functions define, with limits enforced beneath the model
Regulated change control	Validated state in GMP manufacture, verified control points in food safety, product certification elsewhere	Unattended change to a validated parameter, or a release decision made outside the approved strategy	Change inside the qualified envelope, review by exception, and a documented route to widen the envelope
Asset physics	Motor and compressor efficiency classes, heat-transfer surfaces, tooling wear, fouling rates	Recovering a loss that is intrinsic to installed equipment without replacing or modifying it	Operating the asset at its best available point continuously, which most plants do not do

The six bounds, what each forbids and what it still allows. Every bound is a real limit; none is a reason not to act, because in every case the space beneath the bound is larger than most plants are using.

Two of these bounds are frequently misread as technological, and both cost plants years. The first is the safety layer: teams occasionally propose that a model manage a protective function because it could do so more precisely, which misunderstands the purpose of independence — a protective function's value comes from not sharing failure modes with the system it protects. The second is regulated change control. The question in a pharmaceutical or food plant is never whether a model could set a parameter better, but whether that parameter sits inside the qualified envelope and what widening it would cost.

Regulation is a scope question, not a prohibition
The EU AI Act framework (opens in a new tab) is risk-based: obligations attach to the use, and an AI system that is a safety component of machinery sits in a materially different category from one that recommends a compressor sequence. The practical effect on a plant is a classification exercise per use case and a documentation obligation, not a bar on industrial AI. Doing the classification early is far cheaper than retrofitting it.
Energy management gives the ledger a legal home
Sites operating an ISO 50001 (opens in a new tab) energy management system already have the artefacts this page asks for under different names: an energy baseline, energy performance indicators, and a review cycle. Booking an AI-driven saving as a change to an EnPI is usually less work than inventing a parallel process, and it puts the number somewhere auditors already look.
Capability bounds move, thermodynamic bounds do not
It is worth stating plainly because the two are routinely confused in vendor material. Better measurement genuinely raises what a process can target; no measurement raises what thermodynamics permits. When a claim implies the latter, the arithmetic is wrong somewhere — usually in an unstated change of boundary or product basis.
The constraint bound is the one most often ignored
It is also the cheapest to check. Identify the bottleneck, express the proposed gain as its effect on bottleneck throughput, and see whether anything survives. Improvements that do not survive that test are still worth doing sometimes — for cost, for safety, for energy — but they should not be sold as capacity.

For plants without in-house resource to establish these bounds, the public infrastructure is better than most engineers realise. NIST's manufacturing programmes (opens in a new tab) publish measurement science that underpins several of them, the Manufacturing Extension Partnership (opens in a new tab) works directly with smaller manufacturers, and the DOE bandwidth studies give a sector-level starting point for the energy bounds. Beginning from published references rather than from a sister plant's performance is what makes an entitlement register defensible.

Failure modes that turn a booked gain back into a claim

Efficiency is not monotonic. Four regressions account for almost all of the gains that quietly disappear between the project close-out and the next budget.

Efficiency is not monotonic, and the most expensive losses are the ones that were already won. A gain regresses when the conditions that produced it stop holding and nothing is watching — which is common, because reporting is usually a line average and an average absorbs a single SKU's regression for a quarter. Four patterns account for nearly all of it, and each has a cheap preventive measure that costs less than rediscovering the bucket.

Likelihood: highImpact: high

The setting goes back at changeover

The proved value lived in a work instruction, an engineer's notebook or an operator's habit rather than in the recipe record the changeover procedure actually reads. The next SKU restores the old parameter set, the average hides it, and the loss reappears on the Pareto a year later as a new discovery.

PreventionEncode the value in the recipe or parameter set under change control, and track performance per SKU rather than per line.

Likelihood: highImpact: medium

The standard was never re-set, so the saving is spent twice

Finance accepted the number into the project's business case but the standard cost, budget line or EnPI stayed where it was. The following year the same improvement is counted again, the plant's expected cost base no longer reconciles, and the improvement programme is blamed for a gap it did not create.

PreventionMake the standard re-set part of the acceptance step, with a named owner, rather than a task for a project that has closed.

Likelihood: mediumImpact: high

The model is credited with a technician's work

A leak survey, a steam-trap repair or a filter change happened during the trial window and was not separated in the M&V boundary. The reported saving is real and the attribution is wrong, so the next site's business case assumes a model effect that will not reproduce and the approach acquires a reputation for over-promising.

PreventionName every concurrent physical intervention in the M&V plan and measure it as a separate term, even when it is inconvenient.

Likelihood: mediumImpact: medium

The frozen setpoint ages with the asset

An optimal value is locked into the recipe and never revalidated. Heat-transfer surfaces foul, filters load, tooling wears and raw-material specification drifts, so a setting that was optimal at commissioning becomes conservative within a year and, on quality-critical parameters, eventually unsafe for the process it was tuned on.

PreventionVersion the setting and schedule its revalidation against asset condition, not against the calendar alone.

Every one of these is a bookkeeping failure rather than a technical one, which is the recurring theme of this page and the reason the 2030 question resolves the way it does. A plant that fixes its stop-reason taxonomy, meters its largest utility, writes an M&V plan before each change and re-sets the standard afterwards will be able to absorb whatever modelling capability arrives. A plant that does none of those will convert better models into better claims.

Glossary

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

Entitlement: The performance a line or process could reach given its physical limits and installed equipment, stated per loss bucket. Remaining headroom is measured against entitlement rather than against last year, which is what allows a nearly exhausted bucket to be told apart from an untouched one.
Loss bucket: One category in a plant's loss decomposition — unplanned downtime, changeover, minor stops, scrap and giveaway, thermal energy, motive energy, constraint loss, quality release. A bucket is only actionable once it names the controllable parameter that governs it.
Six big losses: The classical OEE loss decomposition: breakdowns and set-up losses against availability, minor stops and reduced speed against performance, start-up rejects and production rejects against quality. It remains the most useful starting taxonomy for a discrete or packaging line.
Specific energy consumption (SEC): Energy per unit of good output — kilowatt-hours per tonne, megajoules per case, and similar. Only meaningful when normalised for product mix and ambient conditions, otherwise a mix change reads as an efficiency gain.
Artificial demand: Compressed air or steam consumed only because the header is held above the highest genuine requirement of any end-use. It is one of the largest recoverable buckets in a typical plant and is invisible without header metering.
Measurement and verification (M&V): The discipline of proving that a claimed saving happened: a boundary, a set of independent variables, a comparison design and acceptance criteria, all fixed before the change. The international protocol for it has governed energy-savings claims for decades.
Matched control: The comparable piece of plant deliberately left on the old setting so the difference is attributable rather than asserted — a parallel line, an alternating-week design, or a normalised baseline model. Without one, natural variation supplies whatever result was hoped for.
Standard re-set: Moving the standard cost, standard rate, budget line or energy performance indicator the moment a gain is accepted, so the saving cannot be counted a second time and the plant's expectation matches its new cost base.
Constraint: The bottleneck that sets the plant's throughput. Recovered minutes on the constraint become units shipped; recovered minutes elsewhere become work-in-progress. Any efficiency claim expressed as plant-average OEE should be re-expressed against the constraint before it is believed.
Thermodynamic minimum: The least energy a physical transformation requires under ideal conditions — the absolute floor beneath an energy loss bucket. Published per sector in the US Department of Energy's bandwidth studies alongside current typical practice, state of the art and practical minimum.
Giveaway: Product delivered above the declared or specified quantity — overfill, overweight, over-thickness. Its floor is set by process spread plus measurement uncertainty, so reducing it is a metrology problem at least as much as a control problem.
TEEP: Total effective equipment performance: OEE measured against all calendar time rather than scheduled time. It exposes the capacity hidden in unscheduled hours, which is often larger than the entire OEE loss and is a scheduling and demand question rather than an engineering one.

Frequently asked questions

The questions plant, engineering and finance leaders ask most often when a 2030 efficiency target lands on their desk.

What does manufacturing hyper-efficiency actually mean?

It means pursuing the residual efficiency in a plant after the obvious waste has gone — the last few points of output, yield and energy available from assets you already own. The distinction that matters is between claimed, booked and held points. A claimed point is asserted from a before-and-after comparison; a booked point has been accepted by finance into a standard or budget; a held point is still present two changeovers later. Most plants claim several times what they book, and book several times what they hold.

Will AI make factories dramatically more efficient by 2030?

In specific loss buckets, yes; as a general step change, the evidence does not support it. Predictive maintenance, in-line quality inspection, soft sensors, setpoint optimisation and finite-capacity scheduling all run in production plants today and reliably move named buckets. What limits their contribution is not model capability but sensing, validation, change control and attribution. The honest 2030 forecast is therefore mostly a forecast about present-readiness: plants that fix their metering, loss model and measurement discipline now will convert whatever arrives, and plants that do not will convert better models into better claims.

How do we know how much efficiency is actually left on a line?

Build an entitlement model per line: state the floor for each loss bucket, the state of the art for that asset class, and the practical minimum with technology under development, each citing a published reference and a date. The US Department of Energy's bandwidth studies publish exactly this decomposition for energy at sector level and are a defensible starting point. Express every remaining target as headroom against those bands rather than as a percentage over last year, and the sequence of what to attack usually changes immediately.

Why does an efficiency gain disappear after a changeover?

Because the proved setting lived somewhere the changeover procedure does not read. If the optimised value is in a project report, a work instruction or an experienced operator's habit rather than in the recipe record or parameter set, the next product restores the old values. Line-average reporting then hides the regression for a quarter or more. The fix is to encode the value under change control in the artefact the changeover actually loads, and to monitor performance per SKU so erosion raises an alarm within the month.

Do we need a digital twin before we can do any of this?

No, and starting there is a common way to spend a year without booking anything. Calibrated models genuinely replace some physical trials where the physics is well characterised and the model has been validated against that specific asset — heat and mass balance, flow, scheduling what-ifs. But nothing in the 90-day plan on this page requires one. Instrument one bucket, bind it to a controllable parameter, prove one change against a control. If a twin is warranted, its real requirements will be visible after that exercise rather than guessed before it.

How do you prove an AI-driven energy saving to finance?

Write the measurement-and-verification plan before the change: define the boundary, the independent variables that normalise it — production volume, product mix, ambient conditions — the comparison design and the acceptance criteria. Then run either an alternating-week design or a normalised baseline regression, and report the delta with an uncertainty band in the plant's own units. Separate any concurrent physical work such as leak repairs as its own term. Finally, get the number accepted into the energy budget or the ISO 50001 energy performance indicator, and re-set that indicator so it cannot be claimed again.

Is OEE the right headline metric for hyper-efficiency?

It is the right decomposition and the wrong headline. OEE is excellent for splitting losses into availability, performance and quality on a single machine, and standards such as ISO 22400 fix the definitions so two plants can compare. But plant-average OEE is misleading as a summary, because gains away from the constraint do not become throughput. Report the constraint's throughput as the headline and use OEE beneath it to explain where the losses sit. Consider TEEP alongside it if unscheduled calendar time is significant.

What is the cheapest first efficiency win in a typical plant?

Usually compressed air, for three reasons. It consumes a meaningful share of plant electricity, it is almost never metered at the header so its losses are invisible rather than absent, and its four loss components — genuine end use, artificial demand, leakage and control loss — can be separated with a flow meter and an off-shift measurement. It also touches no product-quality parameter, so the change-control path is short even in a regulated plant. That combination makes it the ideal bucket for a plant's first properly attributed saving.

How does regulated manufacturing change the timeline?

It changes where the time goes rather than what is possible. In GMP pharmaceutical manufacture or a food plant with verified control points, the analysis and modelling work is unchanged; what lengthens is moving a parameter into the recipe, because that is a change to validated state with a qualification cost. The practical response is to sequence early work onto buckets whose governing parameters sit inside the qualified envelope — utilities, maintenance scheduling, packaging-line stops — and to treat parameters that would require widening the envelope as certification-bound entries in the entitlement register.

Does the EU AI Act stop us putting AI in the control loop?

No. The framework is risk-based, so obligations attach to the use rather than to the technology. A system acting as a safety component of machinery sits in a materially different category from one recommending a compressor sequence or flagging a likely bearing failure, and most plant-efficiency uses fall well below the highest tier. The practical effect is a classification exercise per use case plus documentation and human-oversight obligations. Doing that classification while designing the first deployment costs days; retrofitting it across a deployed estate costs quarters.

Should the model write directly to the PLC?

Almost never, and not for the reason people assume. Recommendations should reach the setpoint the operator or the control system already uses, with hard limits and any protective functions enforced beneath the model and independently of it — a protective function's value comes precisely from not sharing failure modes with what it protects. Keep a human approval step through the first few hundred recommendations: it collects the acceptance and override record that later justifies wider bounds, and a proposal that ships with approval plus a one-switch fallback clears change control far faster than one without.

What does a 90-day efficiency programme cost in team terms?

Roughly one process engineer and one data or ML engineer for the quarter, meaningful time from a named plant owner such as the utilities or line manager, and a few hours from a finance counterpart at the start and the end. Instrumentation is usually a small capital item — a flow meter, a few power meters. The dominant cost is rarely engineering; it is the change-approval route into whatever system holds the parameter, which is why picking a bucket whose parameter you already control is the single biggest lever on the timeline.

About the author

Atomic Loops Engineering

Industrial AI practice

Atomic Loops builds production AI systems for manufacturing, logistics and energy operators — forecasting, scheduling, vision inspection, setpoint optimisation and decision support running against live plant data, integrated into the MES, historian and control layer rather than delivered as dashboards.

· Production deployments in food, pharmaceutical, plastics and industrial-equipment plants
· Loss decomposition and entitlement modelling run jointly with plant engineering
· Integration-first delivery: MES and SCADA write-back, change control, tested fallback
· 19 cited sources on this page

Sources

Lagging vs leading: the manufacturing AI gapWhy the distance between AI-leading and AI-lagging plants widens, and what crossing between the two operating loops involves.Factory digitisation phases and transformation roadmapThe sequencing decisions — connectivity, context, prediction, adaptive control — that decide whether digitisation delivers in quarters or years.Manufacturing industry practiceProduction AI systems Atomic Loops has built for food, pharmaceutical, plastics and industrial-equipment manufacturers.

Find out what your plant can actually book — then go and book it

We run the assessment with your process engineering, operations and finance leads, rank your loss buckets by recoverable headroom against a stated floor, and leave you with a costed 90-day plan for the bucket with the shortest path to an accepted number. You keep the plan whether or not we build it.

Book an efficiency review See our manufacturing work

Published 20 August 2026 · Last updated 20 August 2026

Key takeaways

Abbreviations used on this page

Score how much efficiency your plant can actually book

What manufacturing hyper-efficiency actually means by 2030

The five rungs in detail: Anecdotal to Entitlement-bound

Anecdotal

Loss-mapped

Attributed

Held

Entitlement-bound

Where non-automotive plants actually sit on the ladder

The 2030 efficiency ledger: every loss bucket and its floor

What 2030 actually changes — and what it does not

What booked efficiency looks like in public

The efficiency stack, layer by layer

Sensing and metering

Contextualised loss model

Model and recommendation

Attribution and standards ledger

Control binding and governance

A 90-day plan: booking a compressed-air saving on one plant

The order that keeps it honest

Verifying a point: the measurements that decide the argument

The bounds that set the 2030 ceiling

Failure modes that turn a booked gain back into a claim

The setting goes back at changeover

The standard was never re-set, so the saving is spent twice

The model is credited with a technician's work

The frozen setpoint ages with the asset

Glossary

Frequently asked questions

What does manufacturing hyper-efficiency actually mean?

Will AI make factories dramatically more efficient by 2030?

How do we know how much efficiency is actually left on a line?

Why does an efficiency gain disappear after a changeover?

Do we need a digital twin before we can do any of this?

How do you prove an AI-driven energy saving to finance?

Is OEE the right headline metric for hyper-efficiency?

What is the cheapest first efficiency win in a typical plant?

How does regulated manufacturing change the timeline?

Does the EU AI Act stop us putting AI in the control loop?

Should the model write directly to the PLC?

What does a 90-day efficiency programme cost in team terms?

About the author

Sources

Related reading

Find out what your plant can actually book — then go and book it