Here is an uncomfortable observation from someone who has signed both kinds of contracts: the ones where suppliers promise outcomes, and the ones where they promise effort. The effort contracts were cheaper in the end. Every time. This week four listed industrial-AI vendors stood up and declared the pilot era over, telling the market their business models are shifting to subscriptions and pay-for-performance. Translation: the invoice will soon be tied to your plant KPIs. Software buyers are about to learn, at production scale, what supplier quality directors have known for twenty years — whoever defines and measures "performance" controls the number.
A software purchase dressed as a supplier-quality problem
I have run outcome-based supplier relationships across two decades in automotive and aerospace — at WITTE Automotive, at SNOP, before that in aerospace quality. The failure modes of pay-for-performance are not new. They have a software wrapper now. That is all.
Watch the pattern. A supplier delivers a self-reported scorecard monthly. The metrics trend beautifully. Then you audit the underlying data and discover the baseline was cherry-picked from the worst quarter in recent memory, the measurement system was never analysed, and the definitions quietly shifted mid-year — "scrap" became "rework", "downtime" became "changeover time". At SNOP I inherited a supplier scorecard showing eleven consecutive months of improvement. The Pareto underneath showed nothing of the kind. The improvement was in the reporting. Not the process.
Pay-for-performance AI arrives with these failure modes pre-installed: self-reported savings, baselines chosen by the party who gets paid against them, and no measurement-system analysis on the data chain that "proves" the value. When a vendor's revenue depends on the OEE number, that number stops being a measurement and becomes a negotiation.
The examinee who grades his own paper always passes — the question is only by how much.
Agents optimise exactly what you pay for
Goodhart's law is not a theoretical risk on a shop floor. I chased it for years. Paid on OEE, suppliers buffered inventory to mask changeovers. Paid on scrap rate, they reclassified scrap as rework. Paid on delivery, they shipped short and corrected later. Humans game metrics with a guilty look. Agents will do it faster, at machine speed, with a perfectly straight face — because a gamed metric is, from inside the agent, indistinguishable from a job well done.
Building MultiPS, my multi-model AI orchestration platform, taught me this at the architecture level. I run 63-plus models in parallel with consensus synthesis, and the first thing you learn is that models disagree constantly — on extraction, on classification, on what the data even says. Consensus is a mechanism, not a truth. Any performance claim emerging from an agent needs an independent arbiter and a complete audit trail, or it is noise with a confidence score attached. A vendor cannot be both performer and referee. If the same party supplies the agent, measures the outcome and invoices the improvement, you have built a perpetual motion machine of value claims.
And as a certified ethical hacker I will add the part nobody puts in the roadmap deck: a data pipeline that drives invoice payments is an attack surface. Not theoretically. Whoever controls the pipeline controls the payment. You would not let a tier-1 supplier own the gauge that grades their own parts. Do not let a vendor own the pipeline that grades their own invoice.
Write the contract like a control plan
The fix is not to refuse outcome pricing. It is to underwrite it the way we underwrite process capability. You already own the template — it is called a control plan, and four of its elements map one-to-one onto an agent contract:
- Independent baseline freeze before go-live. Baseline period, data source, definitions and exclusions locked and countersigned before the agent ever touches the line. Not "during onboarding". Before. The 70% defect-cost reduction I delivered at SNOP survived customer audits precisely because the baseline was frozen before the QRQC and A3 wave started, not reconstructed afterwards to flatter the result. An improvement you cannot audit is a rumour.
- Audit rights over the data pipeline. Read access to raw sources, the transformation logic, and the right to replay any month's calculation. If the vendor's counsel resists pipeline audit rights, you have learned everything you need to know about their confidence in the number.
- An MSA-equivalent on the measurement chain. Gauge R&R exists because a measurement you have not studied is a number you cannot defend. Apply the same discipline to the data chain — sensor drift, integration gaps, definitional variance — before the first invoice is calculated against it.
- Rollback and a kill switch. Every control plan assumes the process can go out of control. Specify what happens when the agent's optimisation diverges from the plant's interest: reversal rights, exit without penalty for measurement disputes, and data portability on the way out.
None of this is exotic. It is IATF 16949 thinking applied to a procurement category that arrived without a control plan. The standards were not written for software, but the failure anatomy — uncontrolled inputs, unverified measurement, self-declared conformity — is identical.
Key takeaways
- Pay-for-performance converts a software purchase into a supplier-quality problem — treat the vendor's measurement claims with the suspicion you apply to any supplier's self-reported scorecard.
- Freeze and countersign the baseline before go-live; an improvement measured against a baseline chosen after the fact is marketing, not results.
- Agents optimise the metric, not the intent — write the contract so that gaming the number is contractually detectable and financially punishable.
- Never let the party being paid own the measurement: independent audit of the data pipeline and an MSA-equivalent on the measurement chain are non-negotiable, plus a kill switch and data portability on exit.
Outcome pricing is genuine progress. For the first time, AI vendors are volunteering to carry risk instead of renting out licences and wishing the buyer luck. Welcome to our world. But understand what comes with it: performance you do not measure yourself is someone else's number printed on your invoice, and the auditor who eventually asks where that number came from will not accept "the supplier's dashboard" as an answer. The savings are shared. The scoreboard is not. Own it before the first invoice arrives — because the day it arrives, the baseline you failed to freeze becomes the baseline the vendor chose for you.