Hiring for AI fluency is mostly hiring for scepticism

Job specifications in finance have started sprouting AI requirements, and most of them are useless. Listing tool familiarity selects for people who have used a product that will be replaced within two years, and tells you nothing about whether they can spot a plausible, confidently wrong answer — which is the actual skill.

The interview exercise that works is unglamorous. Give the candidate a piece of AI-generated finance output containing two errors: one obvious arithmetic slip and one subtle judgement error, such as a misapplied accounting treatment described fluently. Ask them to review it. Strong candidates find both and, more importantly, articulate why the second one is the dangerous kind. Weak candidates are impressed by the writing.

Internally, the same competence needs to appear in appraisal criteria or it will not develop. Review of machine output is a distinct skill from preparation and it is currently being learned accidentally by whoever happens to be curious. That is not a capability strategy; it is luck with a budget line attached.

One warning about the internal enthusiast. Every finance function has someone who has automated half their workflow with tools nobody approved. They are usually your most valuable person on this topic and simultaneously your largest ungoverned risk. Bring them inside the tent with a mandate and a control framework. Do not, whatever the temptation, simply tell them to stop; they will not, and you will lose visibility of what they are doing.