AI agents the audit team approved, and the controls that got them there
We build data, analytics and AI for public bodies, starting from what an auditor needs to approve rather than from a demo. Our published public-sector work: eight AI agents in production on citizen services, approved and rolled out in the week they were presented, with zero audit gaps.
- Published work in this industry
- 1 case study
- AI agents live
- 8
- Audit gaps
- 0
- Rollout
- Same week
On this pageWhat we hear in public sector
What we hear in public sector
Each of these comes from our public-sector case study or from an article we wrote for teams in it.
- 01The models work. Nobody will sign off.
The output looks good in a demo, but no one can say what the system used to reach a conclusion, or what happens when an agent gets one wrong.
- 02Data that is not allowed to leave the country
The database is in the right region. The backups, the logs and the vendor's support tooling may not be.
- 03Access rules that live in the dashboard
A row filter protects the dashboard. It stops protecting anything the moment a notebook or an export reaches the same data.
- 04A process nobody has written down
The flowchart describes the good day. The two days a week the job actually takes are spent on everything it leaves out.
- 05An app that can change records
The first write-back button is the cheapest moment to get identity, permissions, an append-only log and a way to reverse a change.
What we have delivered
One published case study, anonymised: citizen services, eight weeks, fixed scope.
Across every agent running in production.
In production on citizen services.
Approved and rolled out in the week it was presented.
Fixed scope, fixed fee, agreed before Week 1.
The controls that got it approved
Each was part of the build rather than something added before go-live.
- 01Lineage on every signal
What an agent used to reach a conclusion can be traced back to source after the fact.
- 02Policy checks
Run on the signals feeding a decision, not on the output once it has already gone out.
- 03Confidence scores
Attached to every signal, so a weak basis is visible instead of buried inside a confident answer.
- 04A kill switch per agent
Each of the eight can be stopped on its own.
- 05An on-call governance pod
So the controls stay live after go-live, which is where most AI governance quietly lapses.
Read the full account
- Case study, public sectorShipped AI agents the audit team approved on day one
We began from what the audit team would need to approve, rather than from what the model could do.
Read the case study ↗
Where to start
Four places a first engagement can begin. Week 1 tells you which one comes first.
- AIGovernance before go-live
Defensible in a board meeting, auditable on demand.
AI governance ↗ - AIKnow what is worth building
One week, and you know which two or three AI initiatives are worth building.
AI audit and roadmap ↗ - ApplicationsOff paper and shared spreadsheets
The forms, approvals and registers that still run on paper, WhatsApp and one shared spreadsheet, moved into a system with a trail.
Process digitisation ↗ - DataOne set of records departments accept
One set of records two teams both accept, with integration, quality and governance underneath.
Data integration and governance ↗
How the first weeks run
The same three phases in every industry. What changes is what we build in them.
- Week 1, fixed feeAudit
Two calls and a written one-pager. Yours to keep whether or not you continue.
- Weeks 2 to 6Build
Something real in front of real users by week three. Median six weeks to the first production-grade artefact.
- Week 7 onwardOperate
Quarterly reviews and on-call governance. The people who built it pick up the phone.
Questions we get asked
Further reading
- Data that is not allowed to leave the country changes the architecturePicking a region satisfies the questionnaire.
- Write down what the AI feature is not allowed to do before it shipsThe shorter, harder list is the one a reviewer asks to see.
- Access rules that live in the BI tool are not access rulesA row filter in a dashboard protects the dashboard, and nothing else.
Start with the Week 1 audit
Two calls and a written one-pager naming the first thing worth building, and who owns it. You keep it either way. NDA-friendly, fixed scope. Write to hello@woodfrog.tech.
