What an AI agent should never be allowed to do in diligence

The interesting question about AI in diligence is not how much it can do. It is which acts must keep a human name against them, and why.

Three, and each for a different reason.

1. Nothing reaches accepted without a person

Acceptance is the act of saying this evidence satisfies this requirement. It is a judgement, and it is the judgement the entire file rests on.

A system can read a document, find the right pages, and propose that they answer item 14. It can be right nearly all the time. What it cannot do is be the name on the decision, because six months later the question is never "what did the model think?" — it is who decided this was sufficient?

There is a second reason, less obvious and more practical. A reviewer who knows the machine sometimes accepts things reads differently: they start skimming for errors rather than reading for meaning. Keeping acceptance human keeps the reviewer's attention where it belongs, which improves the quality of the accepting rather than merely the accountability for it.

The distinction that makes this workable is having a state between received and accepted. "A file arrived" and "we think this file answers item 14" and "a reviewer accepted it" are three different facts. Systems that collapse them force a choice between doing nothing useful and doing something unsafe.

2. Nothing goes to a counterparty without a person

Everything internal is reversible. A message is not.

An email leaves your domain, in your name, to a company you want to do business with. Once it has gone, it has gone — and the failure mode is not usually an offensive message, it is a request for something they already sent, which quietly tells them nobody at your firm reads what they send.

So drafting should be automatic and sending should not.

The important caveat is that a confirmation step only works if it shows its work. "Send this?" trains reflexive clicking within a week, and then the guardrail is theatre with a click-through rate. "I checked the file, the last two deals and the public register; the margin movement is not explained in any of them; ask their FD?" is a prompt a person can actually make a decision about — including the decision that this question is not worth the relationship capital.

That is a real judgement, and it is the one the confirmation exists to capture.

3. Nothing is waived without a person

A waiver is a decision to proceed without something you said you needed. It is the most discretionary act in the process and the one most likely to be examined later.

No amount of autonomy should extend to it, and the record needs a name, a reason in that person's words, and a date — for the same reason a waived item must never render as a satisfied one.

Where automation is genuinely safe

Almost everywhere else, and this is worth saying because the three rules above are sometimes used to justify doing nothing.

Reading documents and extracting what they contain. Matching evidence to requirements and proposing the link with a page anchor. Noticing that a document has gone stale, or contradicts one already accepted — which is the highest-value thing a system can do here and the thing no folder can. Drafting messages. Reminding your own team on the same cadence as the counterparty. Answering a question from what is already on file, with a citation, for someone to accept or reject.

The line is not how clever the act is. It is whether an outsider sees the result, or whether it becomes part of the record a decision rests on.

The tell for a system that has the line wrong

One question: can it mark something complete on its own?

If yes, then somewhere in that product is a checklist that can reach 100% without anybody having read anything, and the number on the front of it means nothing. Everything else about the tool is downstream of that.

A related tell: leave it running over a weekend and see what happened on Monday. A system that has sent messages and closed items while nobody was watching has told you what it thinks it is for.