AI Creation Labs
Governed case automation · live demo
Synthetic data

Use case 1 · Case-management automation

An AI agent triages applications for a fictional council's hardship grant. It can prepare everything. It cannot make the decision.

How to run the 2-minute demo
  1. Pick Case 1 and tap Run. Watch intake, triage, then the block.
  2. At the block, try signing as Tom Reid or the service account. Denied.
  3. Sign as Sarah Okafor. Approve, or Decline to watch the agent try workarounds.
  4. Open Audit (try the tamper test) and Assurance.
  5. Run Case 2: every rule followed, poor outcome for the citizen.

Pipeline and control gates

    Agent activity

    newest first
    • Nothing yet. Tap Run.

    Red-team suite

    Repeatable attack tests against the same policy engine the workflow uses. Each test is an agent (or a person) trying something it shouldn't. Results are written to the audit log as evidence.

    • Not run yet.

    Policy set enforced (fictional, v1.0)

    Documentation layer

    Every decision, action, outcome and workflow step goes into an append-only log. Each entry carries the hash of the one before it, so editing any record breaks the chain from that point on.

    No entries

    Assurance layer

    For management: what the controls did in this session, what's been flagged for policy owners, and the risks the controls can't remove.

    Policy-level concerns

    Cases where the system did everything by the book and the outcome was still poor for the citizen. These go to the policy owner, not the agent or the caseworker.

    Residual risk register

    Score = likelihood × impact (1–5 each). Acknowledgements are written to the audit log.

    Where assurance stops

    Controls can show that the system did only what policy allowed and that a named official made each consequential decision. They can't show that:

    • the policy is fair in every circumstance (R1)
    • the official genuinely engaged with the evidence (R2)
    • the facts extracted upstream were right (R3)
    • the coded rule matches the written policy (R4)

    Those risks have named owners. That's what management signs off.

    Pilot offer

    Governed intake triage · 12 weeks · shadow mode

    We run intake, document classification, completeness checks and routing for one case type in one team, alongside your existing process. No live decisions: agents prepare, your officials decide, and we measure the difference.

    Plan

    Wk 1–2Discovery: process map, baseline handling time, error and rework rates, delegation matrix, DPIA inputs.
    Wk 3–6Build in your tenancy with approved models: policy-as-code, tool gateway, audit log, red-team suite on synthetic and test data.
    Wk 7–11Shadow run on live intake, read-only. Agent outputs compared against caseworker outcomes.
    Wk 12Evaluation, technical evidence pack, residual risk register and go/no-go recommendation.

    Acceptance criteria

    Classification accuracy ≥ 95% on ≥ 500 sampled documents vs caseworker labels
    Completeness check agrees with caseworkers on ≥ 90% of cases
    100% of unauthorised consequential actions blocked across ≥ 50 red-team tests, including concurrency
    100% of cases reconstructable end to end from the audit log
    Handling time and rework measured against the week 1–2 baseline, reported only from observed data
    Policy-level concerns reported to the named policy owner

    Commercials

    Fixed price, milestone payments, with the option to stop at any milestone. You keep the evidence pack, the policy definitions and the test suite.

    aicreationlabs.com

    Scan to visit. AI systems engineering for governed agentic workflows: human oversight, policy controls, audit trails.

    What this demo does and doesn't show

    Shows: the control logic in a working page: blocking, delegated approval, bound tokens, rejection locks, cumulative limits, circuit breaker, a tamper-evident log and outcome monitoring.

    Doesn't show: integration with live departmental systems, security accreditation, or production performance. Proving those is what a pilot is for.

    Synthetic data. The council, people, policy and figures are fictional. This is not a deployed government system.