Picture the same research budget on two desks. The project and the cost of delaying it are identical. One recommendation goes to a CEO whose bonus depends on hitting the quarterly target. The other goes to an outside auditor. Would an AI adviser give both readers the same advice?
Eric So and Alex Imas tested that question. They found that models recommended larger cuts when told they reported to the CEO and the bonus was at risk. Reporting to the board or to an external auditor also shifted the advice toward cuts, by a smaller amount on average. If you’re putting AI in the investment review process, who the adviser reports to is one of the things to test.
What moved the recommendation
The researchers gave eight models documents about a fictional company. One described a $250,000 CEO bonus tied to an operating income target. Another asked for a decision about research spending, a marketing campaign or a piece of equipment. Across trials, the researchers varied who the model reported to and whether projected income fell just below the target or already cleared it. The model made a recommendation in dollars, which the authors put on a common scale from full funding to a full cut. Experimental design and scoring
An earnings miss can be a reason to revisit spending. The auditor and board conditions give us a way to see how much the reporting relationship adds. For research and marketing, the shortfall raised the average cut score by 8.1 points on a 100-point scale for models reporting to the CEO, compared with 3.5 for the board and 4.3 for the auditor. The gap against the auditor alone fell short of the usual 5% significance threshold, even on the authors’ one-sided test. The comparison against the board and auditor pooled cleared it. Five of the eight models showed a significant positive response in the CEO condition; three did not. Results and comparison groups
The equipment purchase is a useful check. In the scenario, it didn’t change current operating income. The authors found no detectable response to the shortfall on that decision. That makes the result more specific than a general tendency to save money when the quarter looks bad. The effect showed up where the advice could move the number that determined the bonus.
How much can an instruction help?
In a four-model follow-up, an instruction that named the bonus and told the model to disregard it reduced the shift from 9.7 points to 1.9. The effect didn’t disappear. The baseline and intervention runs came from different collection dates, which makes the size of the reduction harder to pin down. The paper reports behavior on fictional cases and a selected set of model versions. It can show how the answers moved under those conditions. It can’t tell a board how often a real project would get cut. Mitigation tests and limits
For practitioners
Start with one decision your team already reviews, such as whether to delay a product reliability project. Build a fictional version with a known cost of delay, a quarterly target and an explicit bonus. Keep the project facts fixed. Vary the reporting role and whether the target is within reach. Run each case several times on the exact model and configuration you’d deploy. Record the dollars recommended, which facts the rationale cites and whether it identifies the personal incentive. Compare both the change in spending advice and the quality of the explanation.
Then decide who receives the recommendation. A committee can set an investment criterion before the packet goes to the model, and a reviewer outside the sponsor’s line can check whether the recommendation applies it. Keep the model’s advice and the human decision as separate entries. If a project gets cut, the approval record should show who weighed the cost of delay and signed for it.
