Anatomy of an AI analyst

When we tell a bank that AI analysts work on the SQF platform, the first question is never about capability. It is about control. Who does this thing answer to, what can it touch, and what happens when it is wrong. Those are the right questions, and the honest answer is that an AI analyst is safe on a lending team only if it is governed the way a person is — with an identity, a mandate, a supervisor, and a hard limit on what it may do alone.

That sentence is easy to write and hard to build. This piece describes how we built it.

A named colleague, not a feature

An AI analyst on SQF is not a button labelled "AI". It has a named identity in the organisation directory, with its own role, its own permission set, and its own email address. It is onboarded like any hire — assigned to a role, granted the permissions that role carries, given a supervisor of record — and it is offboarded the same way, with its work history intact.

This matters because everything downstream depends on it. Access control only works if the analyst is a first-class subject of it: the analyst holds no master key and no special back door, only the same permission keys a human colleague in the same role would hold. When it acts, the audit record names it. When its access is wrong, the same role administration that fixes a human's access fixes its access. The analyst is inside the governance system, not beside it.

What it does all day

An AI analyst picks up work the way its human colleagues do — from the platform's own events and from its inbox. A new counterparty appears on an invoice, an application arrives for review, a document lands for screening: the event is the trigger, and the analyst takes the case.

It then investigates step by step, using a narrow set of approved capabilities — the compliance check, the credit report, the document record — and nothing else. It cannot improvise a new tool or reach a system outside its mandate. And it always finishes: every case it opens ends in a recorded conclusion with written reasoning, never a silent abandonment. A case that cannot be concluded is escalated to a person, on the record.

The loop, and the step that cannot be skipped

Every case runs the same five-step loop: Trigger, Investigate, Propose, Approve, Execute. The fourth step is human, and it cannot be skipped. The analyst proposes; a person decides. Client-facing and money-facing actions always carry a human signature — the platform enforces this in software, it is not a policy an enthusiastic operator can waive.

Each step lands on the audit log as it happens. A reviewer can replay any case months later: what triggered it, what the analyst examined, what it proposed, who approved it, and what was executed. There is no step of the loop that leaves no trace.

Confidence, and what it is allowed to mean

Every conclusion an analyst reaches carries a confidence score, and the score is not decoration — it routes the work. At 0.95 and above, a conclusion may proceed autonomously, but only in tiers of work the architect has explicitly approved for autonomy. Between 0.85 and 0.95, the work proceeds with human review by exception. Between 0.70 and 0.85, it is routed to human review as a matter of course. Below 0.70, escalation is mandatory: the analyst hands the case over, with everything it found so far attached.

These are hard rules, not guidelines. An analyst cannot argue its way past its own threshold, and a low-confidence case cannot sit in a queue pretending to be finished. Uncertainty has a defined destination, and that destination is always a person.

Autonomy is earned in stages

No analyst arrives with autonomy. It starts in shadow mode, doing the work in parallel while a human does the real thing, so its judgment can be measured against reality with no consequence attached. When its record supports it, it graduates to suggesting — its proposals now feed the human's queue. Later still, approve-by-exception: routine conclusions proceed unless a reviewer intervenes. Only at the final stage does it act autonomously, and only within bounds set in advance.

The ramp runs both ways. At every stage, rollback to full human operation is possible immediately — not as a crisis measure but as an ordinary operational control. An analyst that starts drifting is pulled back a stage, its history examined, its mandate adjusted. Autonomy is a privilege with a maintenance schedule.

Supervision that would satisfy an auditor

Every analyst reports to a named supervisor of record — a person, accountable for that analyst's book of work the way a team lead is accountable for a junior's. Any analyst can be paused instantly, and pausing loses nothing: its open cases, its history and its reasoning remain on the record for a person to pick up.

Deadlines escalate to people. If a case an analyst owns is about to breach a service commitment, the escalation goes to a human, not to another agent. And nothing — no case, no payment, no credit conclusion — is ever auto-approved. The approval step exists precisely so that the organisation's judgment, not the analyst's, is what commits the institution.

The point of all this

None of this machinery makes an AI analyst slower at the work. It reads a credit file in seconds and screens a counterparty the moment it appears; the governance sits around the conclusion, not inside the investigation. What the machinery buys is something a lending business cannot operate without: the ability to answer, for any decision on the book, who concluded what, on what evidence, and who signed it.

We think this is what putting AI on a lending team should mean. Not a model with production access, but a governed colleague — hired like a person, constrained like software, and auditable like neither has traditionally been. The keys stay where they belong: with the people who answer for the institution.