Every AI review board I have sat in on this year has spent its time on the wrong artefact.
The model card gets read line by line. The bias audit gets a slide of its own. Refusal-rate charts get a meeting. Then the model is wrapped in a dashboard, a warning banner, a set of recommended templates, and the wrapper ships without a second review.
The head of pricing who could not name her own templates
A head of commercial pricing at a Zurich B2B SaaS firm sat down with me in April. She brought a nine-page model-safety memo. Refusal rates were tested. Bias had confidence intervals. Hallucination rates had a control chart.
I asked her to open the internal repricing tool and walk me through what her sales-ops team saw at login.
Three recommended price ladders on the landing screen. A yellow banner warning against undercutting on renewals. A toggle labelled Match Their Move.
She could not tell me who had authorised any of it. The banner had never been scoped in a review. The toggle had been added by a designer during a Friday sprint eight months earlier. The model-safety memo did not mention either.
What the paper measured
Hillenbrand, Normann, Potarca and Werner at ZEW, DICE, KIT and Maynooth ran a preregistered market experiment with 1,083 participants. Pairs of sellers designed rule-based pricing algorithms on a dashboard, playing a sequential Bertrand duopoly of 50 periods, repeated five times. Six treatments varied the design features around the tool.
The LLM advice on its own did nothing. LLM Advice Only produced +0.09 on prices, p=0.327. A clean null.
The wrappers were where the movement lived. A warning against price wars raised prices +0.70. A menu of three algorithm templates raised prices +0.83. A single tit-for-tat coordination recommendation raised prices +1.85, significantly higher than every other treatment including the plain warning (p=0.001). Adding LLM advice to the warning added nothing on top (p=0.537 vs the warning alone).
Once the human's chosen starting price is added as a control, every treatment effect goes to zero. Starting price alone explains 29.7% of market-price variation. The wrappers work by lifting the anchor humans set before the algorithm ever plays.
"We find no evidence that LLM advice adds to the effect of the nudge."
The head of pricing had spent eleven weeks scoping the model. She had spent zero minutes scoping the yellow banner or the toggle labelled Match Their Move. Those are the levers the paper measured. Ask who authorised your wrapper before you ask what your model refuses.