There is a specific question worth asking any vendor building an AI-assisted workflow: where, exactly, in the sequence does the AI sit? Not “do you use AI” — almost everyone will say yes — but where in the chain of steps, and what happens immediately after.

01. Where it actually sits, in our workflows

Look at the nine-stage sequence a piece of work moves through: interface, structured intake, evidence and data, workflow and knowledge, AI assistance, human review, approval, client-ready deliverable, reporting. AI assistance is stage five. Human review is stage six, immediately after it, every time. Approval — a second, separate human decision — is stage seven. The deliverable is only produced and issued at stage eight, after both of those human gates.

That ordering is not incidental. It is the specific design choice that makes AI assistance safe to have at all.

02. What “assists” actually means here

A draft is assembled — the routine composition an expert currently does by hand, built from evidence that has already been gathered and attributed. It is explicitly marked as a draft, and it is a starting point for review, never a finished output. What it produces is not client-facing at any point before a person has read it against its sources.

03. Why this is a rule, not a setting

We describe this as a design rule rather than a configuration option because the difference matters: a setting can be turned off under time pressure, when a deadline is tight and everything “probably looks fine”. A design rule cannot. The workflow is built so that assistance is structurally incapable of reaching a client without passing through review — not merely discouraged from doing so.

This commitment is worded identically wherever it appears — on the homepage, on What We Build, and on Trust — because a rule that changes depending on which page you are reading, or which week it is, is not actually a rule. It is also the substance behind what we do and do not do with your data.

04. What this looks like for the reviewer

A qualified person reads the draft against the evidence it was built from and changes what needs changing. That step cannot be skipped or configured away — it exists specifically because the assistance step exists, as its counterweight. What the reviewer produces is a reviewed piece of work and a record of what was changed, which becomes part of the audit trail. It is also why automation removes the work rather than the people.

05. Why this matters beyond any one workflow

Once an organisation is comfortable with AI assisting inside one workflow, the natural next question is whether the same principle holds everywhere else AI touches the operation. The honest answer, for anything we build, is yes — because the rule is not “AI is fine here specifically”. It is “AI assists, a person decides”, applied the same way regardless of which workflow it is inside.

Read the full governance commitments — access, review, traceability and exit — in one place: how the work is controlled →