Every board asks the same question eventually: which model should we buy? I run sixty-three of them in parallel on a platform I designed and built myself, and I swap them the way a machinist swaps carbide inserts – weekly, without ceremony, the moment a better one ships. The insert does not know the part. The fixture does. This week the industry caught up with what anyone who has run a plant already knew: Caterpillar repackaged decades of mining automation as an "enterprise AI blueprint", and manufacturing writers argued that product context beats model choice. Different vocabulary, identical conclusion. The question is wrong. The model is the least valuable line item in the stack, and what decides whether AI survives on a shop floor never appears on a purchase order.

What Caterpillar actually proved – and it is not the model

Strip the enterprise-AI framing away and look at what the mining company genuinely holds: decades of autonomous haulage telemetry. Every abort. Every grade deviation and operator override. Every component that failed early, and the exact conditions it failed under. That corpus – not a smarter model – is what is being productised. The blueprint sells because the context was captured long before anyone thought to call it AI. Nobody retro-fitted that. It accumulated, because somebody decades ago decided the abort data was worth keeping.

The same news cycle carried the counter-argument from manufacturing writers: on a plant floor, product context – process parameters, tolerance stacks, material lots, changeover recipes – determines whether AI works, not which foundation model sits underneath. It also carried a survey reporting that four in five enterprise AI tools now run without IT oversight. That is what happens when organisations buy the shiny layer and skip the plumbing. Three stories, one message: intelligence is purchasable by the hour. Knowing what to point it at is not.

I swap models weekly – the scaffold is the product

MultiPS, the orchestration platform I built and run in private beta, executes on 63+ models in parallel and synthesises consensus among them. I do not train models; that is not my trade. I deploy and integrate. When a stronger model ships, I re-benchmark and reroute. The interface does not blink. The models are inserts. The scaffold – routing, evaluation, contradiction detection, synthesis – is the product.

Disagreement between models is signal, not noise. When sixty models converge and three dissent, I read the dissents first. That is where ambiguous prompts, hallucinations and genuine edge cases hide. Any quality professional will recognise the pattern: a QRQC session at machine speed, where the one engineer who disagrees with the root cause is the most useful person in the room. Most of us have sat in the meeting where that person was told to be quiet. The platform is built so nobody quiets them.

Your morgue file is the asset no vendor can ship

Journalists call it the morgue file – the archive of everything ever published. Every quality department owns one, curated or not:

  • the 8D archive, containment photos and verification evidence attached
  • the QRQC boards, solved and unsolved
  • PFMEA revision history, including the revisions nobody signed off gracefully
  • nonconformance records tied to part, lot, shift and supplier

Vendors will happily ship you vision systems, agents and dashboards. The box contains everything except twenty years of arguments with your heat-treatment supplier – and that is precisely the part worth money.

I learned this on the floor, not in a webinar. At WITTE Automotive, the failure-cost reduction we delivered through QRQC, A3 and Q-Wall was not cleverness; it was capture discipline. Every defect photographed, contained and traced inside 24 hours, the root cause argued out on a wall where the shift could challenge it. At SNOP, building the quality function of a 900+-employee greenfield plant from nothing, we treated failure history as a product from day one rather than overhead. The numbers at SNOP: a 70% reduction in defect costs and 98% customer satisfaction. More relevant here, we left behind a machine-readable record of every way that plant knew how to fail.

Anyone can rent intelligence now. Nobody can rent your failure history.

That archive is what makes plant AI actually work: the retrieval corpus for an agent answering a line query, the ground truth for a vision classifier, the memory that distinguishes "normal for this press in August" from genuine drift. Frameworks for AI agents that operate physical equipment are now emerging. Ask what such an agent is worth without PFMEA history behind it: a brilliant trainee holding a torque wrench.

Key takeaways

  • Make the nonconformance history machine-readable before signing any AI purchase order – structured fields, linked part and lot numbers, searchable text. A model pointed at scanned PDFs is a model pointed at nothing.
  • Capture failures at source within 24 hours: photo, containment, traceability, root cause. The capture discipline is the data strategy. It was true before AI and it will be true after.
  • Contract for swapability, not loyalty – open interfaces, exportable data, the documented right to change models. If a vendor's business depends on you never leaving, you are buying the wrong asset.
  • Log dissent, from models and engineers alike. When sixty sources agree and three disagree, the minority report is where the next field escape lives.

Caterpillar became interesting to the AI market not because it bought a model but because somebody, decades before AI was a purchase category, decided the abort data was worth keeping. That decision – unglamorous, cheap, made by people whose job titles never contained the letters A and I – is the moat. So run the Monday checklist. Open the 8D archive, check whether a machine can read it, and fix that before you fixate on a benchmark leaderboard. The model is a commodity. The context is the moat – and on most shop floors it is sitting in a shared drive, unstructured and undefended.