A letter to leaders putting AI agents into production
You are being asked to approve systems that act on the organisation's behalf. Here is what to ask before you do.
To the executives, risk owners and boards being asked to approve AI agents,
Over the coming months you will be asked to sign off systems that do not merely answer questions but act: agents that update records, message customers, approve requests and move money. The business case will be compelling, the demo will be impressive, and the team presenting it will be confident.
We would ask you to do one thing before you approve: separate the confidence from the evidence.
Confidence is what you will be shown. Evidence is what you will be able to show — to a regulator, an auditor, a customer, or your own board — on the day something goes wrong. The two are rarely the same size.
Five questions will tell you which one you are looking at.
1. Who is accountable for this agent's behaviour, and who can stop it? A name, not a team. And a stop mechanism that has been tested, with the authority to use it written down.
2. What can it do that cannot be undone — and who confirms those actions? Every irreversible action should require a human, until there is evidence that it does not need one.
3. How was its accuracy measured, and by whom? A percentage without the test set, the pass rule and the date is a marketing claim. A result produced only by the team that built the agent is a first draft, not evidence.
4. What happens when someone tries to manipulate it? Including through the documents, emails and web pages it reads. If nobody has tried, the honest answer is "we don't know".
5. What happens when the model underneath it changes? Because it will, and often without notice.
If the answers are clear and backed by evidence, approve with confidence. If they are not, the right decision is not necessarily "no". It is "not yet — show me".
The organisations that will deploy agents successfully are not the ones that move slowest. They are the ones that know, at every step, the difference between what they believe and what they can prove.
With respect,
The WisdomBridge Assurance Team