An AI agent does not decide what “good” means. It optimizes whatever you were already measuring before it arrived.
What the research did
Supply Chain Management Review published a study on September 9 by Knut Alicke and Ida Hedman that builds a synthetic manufacturer in software: five functional agents, covering sales, supply chain, manufacturing, procurement and finance, negotiate a monthly plan on top of a deterministic planning engine. Each agent carries two independent dials. Locality governs whose objective it optimizes, from the whole enterprise to its own KPI alone. Trust governs how much it believes the signals coming from the others.
Moving all five functions from enterprise objectives to purely local KPIs cost the company roughly €400,000 of EBITDA a year, about 16%, and was by a wide margin the largest value driver anywhere in the map. Trust, on its own, barely moved the company number: where incentives were aligned, a fully skeptical organization earned essentially the same €2.55 million as a fully trusting one.
The worst-performing culture was not the cynical one. It was the corner where every function optimized its own KPI in good faith and every other function believed the inflated numbers, so nothing was filtered anywhere.
The authors are careful to say this is a laboratory calibrated to one synthetic company, not an estimate of what an average business loses. The direction and the mechanism are the point, not the euro amounts.
Why this lands differently now
None of the underlying mechanics are new. Forecast padding, allocation games and the bullwhip have been documented for decades, and every procurement leader has met them. What is new is that those scorecards are being written into software that acts on them continuously, without fatigue, and faster than anyone reviews the output.
That changes the cost of a bad metric. An agent measured on negotiated savings will produce negotiated savings, including the kind nobody can trace to an invoice twelve months later. An agent measured on sourcing cycle time will close events quickly, whether or not the award was the right one. A supplier-risk agent measured on alerts raised will raise alerts. The behavior is not a malfunction. It is the scorecard, executed precisely.
What to settle before you switch one on
- State the objective in enterprise terms. Cost, service, working capital and risk together, not one function’s number.
- Delete the metrics you would not want pursued ruthlessly. If you would be uncomfortable seeing a target hit by any available means, do not hand it to an agent.
- Keep the filter. The analyst who challenges a number that looks too good is not friction in the process. Remove the incentive to inflate and the challenge becomes unnecessary; leave it in place and you need the challenger.
- Write down decision rights. Agents can prepare analysis, draft documents and monitor renewals. Awarding business and declaring savings stay with people.
- Measure results where they land. Agent output is not value. Invoices and budgets are.
One idea to act on
Take the scorecard you use today for one procurement process and read it as if a tireless agent were being paid on it. Ask plainly what it would do to hit every number. Anything on that list you would not want done at scale is a metric to fix before you automate it, not after.
Read more: