The consultants had financial figures and interviews with people inside a company. They had to advise its chief executive about a brand. There was a plausible answer in the spreadsheet and a reason, in the interviews, to doubt it. A careful reader needed both. The researchers who devised the exercise had put the decisive evidence in the sort of document a busy executive might reasonably ask a machine to summarise. 1
Fabrizio Dell’Acqua and his coauthors recruited 758 consultants for a set of experiments with Boston Consulting Group. Participants were randomly assigned to work unaided, with OpenAI’s GPT model (version 4), or with the same model and a short introduction to prompting. On a collection of 18 tasks suited to the model, people with access to it finished 12.2% more tasks and worked 25.1% faster on average. The brand decision belonged to a different exercise. Across the two groups given AI, participants were 19 percentage points less likely than the unaided group to reach the right conclusion. Independent graders still found their writing more persuasive. The research appeared in Organization Science in March 2026. It tested a particular model and a deliberately difficult decision. 1
That last detail about the writing matters in an office. A manager rushing through approvals could have spent more time questioning the recommendation that deserved to pass. The better written document offered fewer visible reasons to open the spreadsheet again.
Put the second pair of eyes where the decision changes
Imagine a coffee roaster deciding whether to discontinue its wholesale range. A spreadsheet shows the line’s gross margin is lower than the margin on retail bags. An assistant can organise the orders, list the assumptions and produce an accurate chart. A café’s purchasing manager might know something the chart cannot show. Several of the roaster’s largest retail accounts came through people who first tasted the product in cafés. That is a hypothesis worth checking in account records before closing the range. It would be equally careless to assume it is true.
Give the person reviewing the decision one assignment while they can still do something about it. Find the strongest piece of evidence that could reverse the recommendation and trace it to a named source. Here, a sales lead could look up how the last few large accounts were introduced, then ask what volume of new retail business would justify the lower wholesale margin. The finance lead can challenge that calculation. By the time anyone starts polishing the memo, the disagreement should be visible in a sentence the owner can act on.
A study in a very different workplace helps explain why a blanket rule about review is a poor use of senior time. Erik Brynjolfsson, Danielle Li and Lindsey Raymond studied the rollout of a conversational assistant to 5,172 customer support agents at one company. With access to its suggestions, workers resolved 15% more issues per hour on average. The gains were concentrated among people with less experience or lower prior performance. The most experienced and highest performing workers saw modest speed gains and small declines in quality. The agents remained in the conversation with customers. 2
For a manager, the practical question is where experience already resides. A new account handler may benefit from seeing how a senior colleague explains a familiar charge. On the wholesale decision, a person who remembers how accounts arrived should challenge the premise before the assistant drafts an attractive argument. Buying the same software for both jobs tells you little about which arrangement will help either employee.
At the next review, ask to see one recommendation, the evidence on which it turns and the name of the person who checked that evidence. Then ask what they found. If all they can remember is changing the adjectives, the business put its most expensive reader on the wrong paragraph.


