Every compliance team we speak to asks the same question in the first ten minutes, usually phrased quite carefully.
So does the AI make the decision?
No. But the flat answer is less interesting than the real one, because the agent goes further than most people expect before it stops.
The line is not where you would guess
An agent gathers the evidence from your systems. It writes up the case, with each finding linked to the source it came from. And then it tells you what it thinks: a proposed disposition, the reasoning, and the specific findings that reasoning rests on.
So it does form an opinion. Claiming otherwise would be a strange thing to be proud of, and it would make the product a good deal less useful.
What it never does is decide. Three things keep that boundary real rather than rhetorical.
The recommendation is an advisory. There is one per case, it refreshes as the evidence changes, and it has no authority of its own.
Your decision is terminal. Once you have ruled on a case, the system stops recommending on it. A fresh opinion after the fact is treated as noise rather than a second view worth raising. Your answer ends the conversation. The agent's never does.
It works out its own view rather than inheriting yours. Where a human decision already exists as context, the agent assesses independently instead of deferring to it. An advisory that quietly agrees with whatever was decided last time is worth nothing to anybody.
Why the line sits there and not further along
Moving it would be straightforward. A system that can write a defensible rationale can act on one, and plenty of alerts have obvious answers.
We do not, for three reasons.
Someone signs. An MLRO is personally accountable for the filings their firm makes. Not their employer, them. When a regulator asks why a transfer was cleared, the answer cannot be that a system decided and nobody looked. There is no version of that conversation that ends well, and no vendor indemnity that makes it end better.
The judgement was never the bottleneck. The work in an investigation is the assembling. Finding the records, chasing the counterparty, reconstructing the flow. That is hours. The decision, once everything is in front of you, is usually minutes. Automating the minutes and leaving the hours would be the wrong way round.
An opinion can be audited. A decision by a model is harder to examine. When an agent recommends, you can check the evidence it used and the standard it applied, because both are recorded. When an agent decides, you are examining a judgement, and the only artefact is an explanation written after the event.
What it costs us to draw it here
It would demo better the other way. "Fully automated alert disposition" is a cleaner slide than "your analyst reviews the case note", and there is a version of this market that rewards the cleaner slide.
We would rather be straight about the design now than have you find out halfway through a procurement that you had assumed something else. And the review is not a limitation we have not got round to removing. It is where the defensibility comes from, and it is the reason the record holds up years later.
The part that makes it more than a promise
A boundary described in marketing copy is not worth much. A boundary you can audit is worth something.
Every recommendation carries the version of the encoded standard behind it. Every case note carries the version of the template it came from. Every decision, and who made it, goes into a record the application itself has no permission to change: not as a matter of policy, but because the service that writes it has no rights to update or delete.
So we are not asking you to believe the agent did not decide. The record shows who did.
The short version
Agents gather, write up and advise. You decide. Everything either of you did is on the record, and the record cannot be quietly tidied afterwards.
Your MLRO signs. Cogentic never files.