Lessons from CEOs on AI adoption: what automotive leadership teams can actually use

Stage 1 is not scepticism — it is distance. Most delegated leadership teams believe AI matters; they have simply classified it as a technology procurement rather than an operating change, and routed it to the function that buys technology. The classification is the error. Every lesson in the public record — Toyota's, BMW's, Volkswagen's — begins with a chief executive deciding that AI is a question about how the company makes cars, not about what the IT estate contains.

The tell is the agenda. At stage 1, AI reaches the executive team as an occasional innovation update: a slide of pilots, a vendor demo, a benchmark of what competitors announced. Nothing on that slide has an owner in production, a KPI the plant recognises — OEE, FTT, scrap — or a date tied to a programme SOP. The update is received with interest and produces no decision, because it was not structured to require one.

The cost of staying here is not the missed pilots; it is that the organisation learns to treat AI as weather. Plants experiment locally, a tier-1 supplier or a software vendor sets the pace, and by the time a competitor's system is inspecting every body-in-white in cycle time, the delegated team has no institutional muscle to respond with — no review, no gate, no owner, no vocabulary.

Stage 2 is the mode of the industry, and it is where the announcement–delivery gap opens. The commitment is genuine: the CEO has said, publicly and often personally, that AI is strategic. Budget follows, a programme acquires a name, and a leader is appointed. What has not yet happened is the unglamorous conversion of message into operating machinery — owners with production KPIs, a review cadence with teeth, gate criteria a plant manager recognises. The words are ahead of the plants, and every quarter the gap persists, it compounds.

The compounding is the dangerous part. Each public claim that the organisation cannot yet verify creates expectation debt — with markets, with customers, with the workforce, and inside the leadership team itself. Servicing that debt distorts behaviour: pilots are counted as adoption because the count is needed for the next keynote; internal reporting drifts optimistic because nobody wants to be the slide that contradicts the CEO. The automotive record of the early 2020s is rich in exactly this pattern, and the group-level software programmes that publicly reset their timelines — extensively covered in the business press — paid the debt down at the cost of leadership credibility.

Leaving stage 2 does not require retracting ambition. It requires an announcement ledger — every public AI commitment, mapped to what verifiably runs — and the discipline to let the ledger, not the ambition, write the next statement. The CEOs who navigated this well made commitments that were enormous but checkable: a named platform, a named partner, a stated number of plants. A checkable pledge can be tracked, missed, corrected and ultimately delivered. An uncheckable one can only inflate.

Stage 3 is where the lessons stop being borrowed and start being earned. The defining artefact is the executive review: a fixed cadence — monthly is typical — chaired at genuine executive level, working from a one-page-per-initiative format that states the owner, the production KPI, the evidence class and the decision required. The meeting's value is not oversight for its own sake; it is that a known, unavoidable review changes behaviour upstream. Initiatives arrive with baselines because they know they will be asked. Zombie pilots get killed because the page makes their state visible. The organisation learns that the CEO reads the ledger.

The second artefact is the anchor deployment — one plant, one process, instrumented end to end. The automotive-authentic choice is a quality application: paint-shop surface inspection or body-shop weld verification, written into the MES and the andon loop, measured in FTT and rework hours against a holdout line or shift. The anchor matters because it converts the executive conversation from analogy to evidence. When the CEO can say 'at our anchor plant, first-time-through moved, and here is the holdout that proves it', every subsequent decision — scale, fund, stop — happens against a shared, checkable fact.

What stage 3 has not yet solved is durability. The review runs because the CEO runs it; the gate criteria live in the programme office; the works-council agreement covers one plant. Remove the individuals and the machinery stalls. That is not failure — it is the natural state of a system one budget cycle old — but it is why the record's stage-4 lessons are all about institutionalisation: moving the machinery from people into the operating system, where it survives reorganisations and successions.

Stage 4 is the industrialisation of the leadership behaviour itself. What the CEO personally did at stage 3 — demand evidence, chair the review, arbitrate trade-offs — is rebuilt as machinery that does not need the CEO in the room. Gate criteria for taking an AI system past pilot are written into the QMS the company already certifies against IATF 16949, which means an auditor now checks them as a by-product of an audit that was happening anyway. Decision rights are published: the plant proceeds within criteria, the executive review arbitrates genuine trade-offs, and a named role — not a committee — can stop a degraded system the way anyone can stop a line.

The distinctly automotive layer at this stage is the social one. In co-determined environments, the operators that scaled cleanly negotiated framework agreements with employee representatives covering the plant network — what AI systems may monitor, how operator data is handled under TISAX-consistent rules, how roles change and who is retrained — rather than renegotiating plant by plant, deployment by deployment. Leadership teams that treated the works council as a launch-week stakeholder paid for it in delayed SOPs; the ones that brought representatives into the gate design got something better than consent: a co-owner of the machinery.

The final stage-4 discipline is leadership development with the plants, not above them. The record's institutionalised operators send executives to running deployments — gemba, in the vocabulary the industry already owns — and rotate plant managers through AI ownership so the capability stops being a staff function's property. By the end of stage 4, the question 'who owns AI here?' has the same slightly puzzled answer as 'who owns quality?': a named system with named roles, and everyone.

Stage 5 is rarer than the keynote circuit suggests, and quieter. The signature is succession-proofing: a chief executive has changed — retirement, rotation, the ordinary churn of the industry — and the AI operating machinery carried on without a strategy reset, because it was never the person's property. In an industry where programme lifecycles outlast most executive tenures, this is the difference between a capability and an era. The honest test cannot be run in advance; the proxy is whether every stage-3 and stage-4 artefact — the review, the gates, the agreements — is owned by a role rather than a name.

The second signature is that the direction of lesson-flow reverses. Compounding operators export their machinery: they anchor shared data infrastructure like Catena-X rather than merely joining it, contribute the audit patterns that harden into industry practice under VDA and IATF processes, and engage regulators — the EU AI Act's provisions for industrial AI, UNECE's vehicle software regulations — as participants in the rule-making rather than recipients of it. This is not altruism; a company whose internal machinery matches the emerging external rules pays near-zero compliance retrofit cost, and helps write rules it can already pass.

What remains for the CEO personally is the frontier that machinery cannot own: capital allocation between plants and software, build-versus-partner calls with hyperscalers and tier-1s, and the discipline of not extending automation past its evidence just because the organisation now can. The record suggests stage 5 is also the easiest stage to regress from — a merger, a margin crisis or a charismatic successor with a new brand can unwind institutionalised machinery in two budget cycles. Compounding is a practice, not a destination.