Platform capability
Naming an AI agent is a marketing act, and the category is full of them. What makes one worth anything to a compliance buyer is being able to answer, months later, what it looked at and what it did.
When an agent has been collecting your evidence for a year, the questions that follow are not about how clever it is. They are: what did it look at, what did it change, did a person agree, and what produced that output? "The AI did it" is not an answer, and a system that cannot do better is one you have to take entirely on faith.
So every run of every agent here writes a row — what it read, what it produced, what it changed, which model, how long it took, who set it going. Including the runs that failed. Including the runs that found nothing to do, because "checked, nothing needed" and "nobody has checked since June" are different facts and only one of them is reassuring.
The second decision is the number. Competitors ship eight named agents; we ship three that do real work, and each one states plainly what it is allowed to do. An agent that can quietly change your compliance position is a liability dressed up as a feature.
Coverage
Each one is limited to a single kind of action, and says which.
Gathers and files the artefacts your controls need, from your own data: access records, policy register, audit trail, processors, consents, data-rights history. It needs no approval because it invents nothing — it exports what the platform already knows and files it against the clauses it evidences, dated and with an expiry.
Answers the questionnaire rows your answer library has never seen, grounded in your real control state and instructed to say "no" or "partial" where that is the truth. Everything it writes is a draft behind a human approval gate, and its output is deliberately excluded from bulk approval because nobody has read it yet.
Reads your policy set against the frameworks you hold and reports the drift: published documents nobody approved, reviews that have come due, control themes with no policy behind them. It writes nothing at all. It proposes; it never publishes.
The Evidence Collector and the Policy Architect are pure data work and keep running with the AI layer switched off entirely. A compliance control that stops working when an API key expires is not a control.
An agent that throws still closes its ledger row with the error attached. A run permanently stuck in "running" that nobody can explain is the failure mode the design exists to prevent.
Across all three: an agent may read anything in its own tenant, may write only what a person can undo, and may publish nothing.
Approach
From the roster, or as a side effect of the work it belongs to — running the questionnaire fill is the Analyst doing its job, and it leaves a ledger entry like any other run.
What it is about to look at, written down before it starts, so a crash leaves evidence rather than a gap.
Acts, drafts or advises — whichever it declared. Nothing else.
Items read, items produced, items changed, model, duration, outcome. Success, no-op or failure, all recorded the same way.
Deliverables
The report is the product. If it cannot be acted on by a developer and understood by a director, we have not finished.
Every run of every agent, newest first, filterable by agent.
Scope, result, counts, model and duration on each one — the four questions, answered.
What each agent may do, on the page, next to the button that runs it.
The Policy Architect's report rendered properly — overdue reviews, unapproved documents, uncovered themes — rather than as a JSON blob.
Is this for you?
If none of them are, say so on the call and we will tell you honestly whether this is the right piece of work — or point you at the one that is.
Book a scoping callHow we work
The same engagement model applies to every piece of work we take on, so you always know what happens next.
A 30-minute call, then a written scope: what is in, what is out, what we need from you and what it costs. Nothing starts before you sign it.
Rules of engagement, testing windows, escalation contacts and a signed authorisation. Out-of-hours windows where production cannot take the load.
Automated coverage first, then manual testing where judgement is required. Critical findings are reported the day we confirm them, not at the end.
One report a developer can act on and an executive can read, with evidence, reproduction steps, business impact and a fix for every finding.
A walkthrough call with your engineers. We answer questions on the fix, not just the finding.
A free retest cycle to confirm the fixes hold, and a clean summary you can hand to a customer, auditor or board.
Questions
The questions clients actually ask during scoping. If yours is not here, ask it directly.
Because three that do real work are worth more than eight that are announced. Each of these has a defined job, a stated limit on what it may do, and a ledger entry for every run. Adding a fourth is easy; making the first three accountable was the work.
Not in any way a person cannot undo, and never by publishing. The Evidence Collector files artefacts, which is additive. The Questionnaire Analyst produces drafts that require approval. The Policy Architect writes nothing at all. No agent marks a control verified, approves a policy or sends anything outside your organisation.
Two of the three keep working exactly as before, because they are data work rather than model work. The Questionnaire Analyst becomes unavailable and says so — and it still writes a ledger row recording that it did not run and why, rather than failing silently.
Which agent, what it was scoped to, a plain-English result, items read, items produced, items changed, the model if one was used, the duration, how it was triggered, who started it, and the start and finish times. Runs that found nothing are recorded as such, and failed runs carry the error.
Only for the agent that needs one, only the context that agent requires — the questions and your own control state — and only on our hosted platform. On an on-premises or private-cloud deployment you can point the AI layer at a model inside your own network, or leave it off and run the two agents that never needed it.
Next step
A 30-minute call, then a scope document with what is in, what is out and what it costs. No obligation, and no charge for the conversation.
