The fee is contingent. The baseline is negotiated.

Here is an uncomfortable observation about Murata's three-year AI production-reform contract with Accenture, the one where part of the fee rides on results. The press reads accountability. A quality director reads a baseline war with a start date. I have sat on both sides of that table. As a plant quality director I signed off savings claims; through FOREAST, my lean and quality consultancy – clients included ArcelorMittal – I was the vendor, and on some engagements the invoice depended on the number. On one of them I spent the better part of a month arguing with the client's controller over whether in-line rework counted as failure cost. Roughly €80,000 of my fee sat on the answer. The AI is new. The fight is not.

Money on a result makes the result a CTQ

Tying fees to outcomes is one of the biggest attempts yet to price AI transformation as a result rather than a licence, and the instinct is sound – I sold it myself for years. If the integrator only earns when the plant gets better, the incentives finally point the same way. That is real alignment. But follow the mechanism. The fee does not follow the model; it follows a KPI, and a KPI is a definition someone wrote. What counts as a defect. Whether rework is valued at standard cost or full cost. Which lines, which shifts, which denominator. The moment money moves on a result, that result becomes a critical-to-quality characteristic – specified, negotiable, contestable, exactly like a bore tolerance on a steering knuckle.

And a CTQ that pays an invoice attracts a very particular species of attention. Alignment does not remove the risk. It relocates it – from "will the technology work" to "who owns the measurement". Quality people have been arguing that one with consultants and finance for four decades. We are about to have it again, at AI scale, with more zeros on the fee note.

Once a number pays an invoice, the number becomes the product – and somebody will manufacture it.

The baseline is where the money is

Every results-based contract contains a baseline, and that is where the money lives. Fifteen per cent improvement against a padded baseline out-earns forty per cent against an honest one. The games are ancient in cost-of-poor-quality work: pick a reference period that happens to include a disastrous quarter; fold in one-offs – a launch, a strike, a wave of customer audits; classify chronic waste as business-as-usual so it never enters the arithmetic. Then let the argument start in month eight, when the definition turns out to be worth seven figures. I know where these fights start, because as the vendor I was in them. Never at kick-off. Always once the number became real.

The 70% defect-cost reduction we delivered at SNOP is quotable for one reason: the baseline was locked and audited before the programme launched. Which cost centres counted, how rework was valued, which months formed the reference, who could check the arithmetic – a signed document, not a handshake. At WITTE our QRQC, A3 and Q-Wall work took failure costs down hard, and I still watched the savings die in finance. Engineering's spreadsheet and controlling's spreadsheet disagreed about what had been promised. Both were honest. Neither was authoritative. The improvement was real; the credit was uncollectable.

Run the integrator like a supplier, not an oracle

So stop treating the integrator as an oracle. An oracle that invoices quarterly is just a supplier with better branding. Run it under supplier-quality logic – the same IATF 16949 thinking you would apply to any tier-one source of production-critical inputs. The CTQ definitions belong in an annex drafted with the rigour of a customer specification: formula, data source, exclusion rules, dispute mechanism, named owner of the data.

Audit rights on raw data, not dashboards. Never let your supplier run your receiving inspection. Control charts on the KPI itself, because SPC separates genuine special-cause improvement from reclassification, seasonality and a quieter customer that month. And a year-one exit gate with acceptance criteria written before kick-off, not after the first invoice.

There is an AI-specific layer, and it is the part I live in. I designed and built MultiPS, a platform that runs sixty-odd models in parallel and scores their disagreement daily; in that world, measurement governance is not an annual event – it is the operating system. Model outputs drift, disagree and fail in correlated ways. If the contract KPI is computed by a system the vendor also configures, you have outsourced not just the improvement but the scoreboard. Deploy and integrate by all means. But the measurement chain belongs to the plant.

The Murata contract is a control plan in a suit – reaction plans, acceptance criteria, measurement-system analysis, drafted by lawyers instead of quality engineers. If the quality function does not sit at that table, the fee will follow whatever number is easiest to move, and the plant keeps the defects either way. I have collected on outcome-priced work and conceded on contested baselines, and from both chairs the lesson is identical: the model will be whatever the model will be; the money is decided in the annexes. Quality people should be writing them.