AI compliance for carbon reduction targets in automotive manufacturing: making a modelled emissions number audit-grade

Stage 1 is not the absence of a carbon number — most automotive manufacturers have had one for years. It is the absence of any resolution beneath it. A spend-based Scope 3 line is computed by taking what you paid a supplier and multiplying it by a sector-average factor, which means the figure moves when procurement negotiates a discount and does not move when the supplier switches to renewable electricity. It is directionally useful for a first inventory and actively misleading as a target-tracking instrument.

The tell is what happens when someone asks a follow-up question. Ask which of your aluminium castings suppliers has the highest footprint per kilogram and the honest answer at stage 1 is that the question cannot be answered from the account, because the account never had a kilogram in it. Ask what your reduction programme achieved last year and the answer is a number that also reflects volume, mix, currency and commodity prices. Every finance director eventually asks a version of this question, and stage 1 has no answer that survives it.

AI at this stage is nearly always doing one job: classifying purchase lines into spend categories so the average factors can be applied. That is genuine, useful work and it is also the highest-risk work on the page, because it feels like accounting automation while it is in fact making thousands of unreviewed judgements that determine which factor gets applied to which euro. When the classifier is wrong it is wrong systematically, and the error lands in a legally required disclosure with no marker on it.

Stage 2 is the most dangerous stage on this ladder, precisely because it looks like a large step forward — and technically it is. The account now speaks in kilograms and process routes. A model trained on the PCFs suppliers have declared can produce a credible estimate for the several thousand parts where none exists, and the resulting total is far more useful for engineering decisions than anything stage 1 produced.

What is missing is the marker. In almost every stage-2 estate the modelled value and the supplier-declared value are written into the same field, and the only way to tell them apart is to ask the analyst who built the extract. That is fine until the figure leaves the building. Under limited assurance, the first thing a provider does is sample figures and trace them to source; a sampled figure that resolves to a model with no version, no back-test and no uncertainty band is not a finding about the model, it is a finding about the control environment.

The second thing missing is version discipline. Emission-factor sets are updated, global warming potential sets change between assessment reports, allocation rules get refined, and the model itself is retrained. At stage 2 none of these events is recorded against the figures they changed, so when this year's number differs from last year's, nobody can decompose the difference into real-world change and method change. That decomposition is exactly what a restatement conversation turns on.

Stage 3 is the first stage where the carbon number survives contact with someone who did not build it. The change is not analytical, it is structural: the ledger stops holding numbers and starts holding statements. A row no longer says 4.81 kg CO2e; it says 4.81 kg CO2e, modelled, estimator v7, factor set 2026.1, cradle-to-gate, plus or minus 22% at the 80th percentile, owned by the commodity lead, superseding the value published on the previous cycle.

The discipline that gets you here is closer to quality engineering than to data science, which is convenient, because automotive manufacturers already have it. A part characteristic that matters is not simply measured — the gauge is qualified through measurement systems analysis, the study is recorded, the calibration has a date, and the result is quoted with a tolerance. Apply that vocabulary to a gap-fill estimator and every question an assurance provider asks has an existing answer shape. The organisations that struggle at this stage are usually the ones treating the carbon model as reporting software rather than as instrumentation.

The constraint that emerges is supplier data. Once provenance is visible, the primary-data share stops being a rhetorical claim and becomes a number on a slide, and it is almost always lower than anyone expected. This is the point at which the programme's centre of gravity moves out of the sustainability team and into purchasing, because the only way to raise it is to make a part-level PCF a condition of the award and to give suppliers a machine-readable path to send one.

Stage 4 is defined from the outside in. The question is no longer whether your team understands how the number was produced but whether a reasonably diligent stranger can rebuild it from artefacts you already keep. That is the practical content of a limited-assurance engagement and it is the same test a CBAM verifier, an OEM's supplier-audit team and a customer's procurement quality function each apply in their own vocabulary.

The work at this stage is control design, and it is mostly borrowed. Change control for the estimator is the change control you already run on MES software. The bias and spread study is a gauge study with a different unit. The reproduction rehearsal is an audit rehearsal. The evidence pack is a control plan. Manufacturers that already hold IATF 16949 certification and run ISO 14001 environmental management have every one of these muscles; what is usually missing is the decision to apply them to a system the sustainability team owns and manufacturing quality has never looked at.

The characteristic stage-4 discovery is that the models were never the weak point. Reproduction rehearsals overwhelmingly fail on inputs: a factor set someone downloaded and stored locally, a supplier declaration that exists only as a PDF attachment, an allocation decision made in a meeting and never written down. Fixing those is dull, cheap and the difference between an assurance engagement that costs five weeks of the team's year and one that costs five days.

Stage 5 is narrower than the marketing language around it suggests. It is not an autonomous carbon-management system. It is the specific arrangement in which one ledger serves both the outward-facing disclosure and the inward-facing decision, so that the number a purchasing manager sees when scoring a quotation is the number that will appear, months later, in an assured statement. That single-source property is the whole point: the moment the decision runs on a different figure from the disclosure, both become negotiable.

The hard part here is governance rather than engineering. If carbon intensity carries weight in an award decision, the weighting is a commercial policy with legal consequences and it has to be versioned, reviewed and defensible to a supplier who lost on it. If an engineering change is approved partly on a footprint delta, the estimator behind that delta has just become part of the product development process and inherits its documentation expectations. The organisations that do this well tend to route the policy through the same forum that handles customer-specific requirements, because it is the forum that already knows how to version a commitment.

Sustaining stage 5 is the part most often underestimated. Factor sets move, suppliers change plants and power contracts, regulations amend their scope and reporting dates, and a threshold set against 2026 conditions quietly stops being valid. The leading indicator worth watching is the rate at which decisions fall outside the policy's stated bounds and route to a human: when it rises, the world has moved and the policy needs a review before an incident forces one.