The future of AI adoption in automotive: three fronts, two clocks, and what an OEM can absorb

Experimenting is rarely a shortage of activity. In a large vehicle manufacturer it is usually the opposite: a paint shop trialling a vision model, an ADAS team running a perception experiment, a purchasing group with a generative tool for supplier documents, an aftersales team clustering warranty claims. Every one is defensible. What is missing is any single place where they are visible together, so nothing about them can be sequenced, compared or stopped.

The tell is the inventory question. Ask for a list of every AI system running anywhere in the company — in vehicles in the field, in plants, in the back office. At this stage the honest answer is weeks of asking around and a list that is still wrong, because a substantial share of the work sits inside licences individual teams bought themselves. That invisible tail is where data leaves the company under terms nobody reviewed.

The stage is cheap to leave and expensive to occupy, and the expense is not the failed trials — it is that each trial amortises nothing. The tenth vision pilot in the tenth plant costs what the first did, because the second plant learned nothing transferable: no shared labelling convention, no shared evaluation harness, no shared answer to who is paged when a model degrades on a Sunday shift.

Pocketed is where most vehicle manufacturers are, and it is an achievement rather than a failure. Something works: a vision system catches a defect class the line was missing, a driver-assistance feature ships and reviews well, a warranty-triage model saves a team real hours a week. These are not demos — the users would object loudly if they were switched off, which is the honest test of production.

The problem is that each pocket is a vertical, with its own data path, vendor, evaluation method, definition of good and answer to what happens when it is wrong. None of it is available to the next team, so the marginal cost of capability number three equals number one. Manufacturers here often report rising AI spend and flat AI capability, and both numbers are correct.

This is also where the three fronts diverge without anyone deciding they should. The product front is usually furthest ahead, because a vehicle programme has a hard date and an integration culture; the plant front is next; the enterprise front is a scatter of licences. That imbalance is the residue of which sponsor was loudest, not a strategy — and time here is not neutral, because pockets harden into constituencies that later defend their own stack against any shared foundation.

Programmed is the first stage where the word portfolio means something. There is a register, the register has owners, and the owners meet. Crucially the meeting is about sequence rather than approval: everyone accepts that more good ideas exist than can be delivered, so the discussion is which quarter each one lands in and what it displaces. That conversation is what first surfaces absorption capacity as a real number.

The character of the work changes too. At stages 1 and 2 the hard problems are technical. At stage 3 they are calendar problems: the plant use case is ready but the change window is in October and the model-year freeze lands in September; the product feature is ready but the type-approval evidence is not; the enterprise agent is ready but the works council has not reviewed how its output affects job content. None of these is solved by better modelling, and teams that keep trying stall here for years.

The constraint that emerges is that everything below the portfolio layer is still bespoke. The register is shared; the plumbing is not. Each initiative builds its own data path, evaluation and rollback, so throughput is capped by the number of teams capable of all three — a number usually far smaller than leadership believes. This is where the honest absorption figure first gets measured, and where it is usually a shock.

At stage 4 the marginal cost of the next AI capability finally falls, and the reason is unglamorous: the shared pieces are the boring ones. A common way to hold and version training data. A common evaluation harness that answers 'is this better than what we ship today' for a paint-shop model and an in-cabin assistant alike. A common release path with a rehearsed rollback. Once those exist, a new initiative is mostly specification and delivery is measured in weeks.

The distinctive automotive property of this stage is that the platform must serve two very different risk regimes without pretending they are the same. A plant model that misclassifies a weld is a scrap and rework problem; a vehicle function that misjudges a lane boundary is a homologation and liability problem. A platform that treats them identically will either strangle the plant front in evidence it does not need or under-serve the product front in evidence it absolutely does. Share the mechanics; keep the evidence bar per-front.

What binds here is no longer technology or sequencing. It is people and bets. The platform makes it cheap to do the next obvious thing, so the portfolio fills with obvious things while the genuinely strategic questions — how far to go on autonomy, whether to own the in-vehicle model stack, whether the plant should run its own inference — are deferred because they are uncomfortable and never urgent. Manufacturers plateau here in comfort rather than in pain.

Compounding is narrower than the word suggests. It does not mean an AI-run company; it means the loop closes. A warranty pattern seen in the field becomes a test case in engineering and an inspection class on the line, automatically, because all three sit on the same data and evaluation foundation. The tenth capability is worth more than the first not because the model is better but because the estate it plugs into is richer.

The engineering is largely solved by the time a manufacturer arrives here; the hard part is organisational memory. Compounding requires that capabilities die without taking their assets with them — when a vision system is replaced, its labelled images, evaluation set and failure catalogue stay. Most organisations lose exactly this, because the assets belong to a vendor contract or a programme that ends. Asset survival is a dull supplier clause with an outsized effect on the ten-year trajectory.

Sustaining this stage is a governance discipline, and it is the stage most likely to regress. A reorganisation splits the fronts apart; a cost programme cancels the platform group because it ships no features; a new model programme negotiates its own stack for good local reasons. Compounding is not a plateau you reach but a position you defend, and the defence is mostly about who owns the shared assets when budgets tighten.