Much of the argument about AI in government is fought over the wrong line. The question is not whether a machine should decide a citizen's case — it should not, and in our deployments it cannot, at any setting. The useful question is narrower: why does an officer spend three days assembling a case they will decide in ten minutes?
The delay in most citizen-facing work is not the decision. It is the preparation: pulling records from five systems, checking eligibility rules, reconciling the identity that appears three ways in three registers, assembling the file. The officer's judgment is the scarce resource, and it is spent mostly on retrieval.
What arrives on the desk
Under an approval gate, the agent does the assembling. The case arrives prepared: the records joined, the rules checked, the discrepancies flagged, the evidence attached to each conclusion. Not a recommendation smuggled in as a default — a preparation, with the reasoning visible, that the officer can accept, adjust or reject.
The difference between this and a decision-support dashboard is accountability. The agent's work is signed and recorded. The officer's decision is signed and recorded. When the case is reviewed two years later — and in government, cases are reviewed — the record shows what was prepared, what was decided, and by whom, reconstructed as it stood on the day.
The line that does not move
Some actions in government are gated Auto — reminders, routing, low-risk requests. Some are Advise — analysis and evidence, no action at all. Anything touching a citizen's entitlement, eligibility or penalty is Approve, and that gate is enforced in policy on the decision path, not in a guideline document a busy week can erode.
This is not a limitation we work around. It is the product. A system that prepares honestly and decides never is one a department can defend to its audit institution, its courts and its citizens — and defence after the fact is the actual requirement, whatever the brochure says about speed.
Why officers like it
The surprise in these deployments is not that management likes the throughput. It is that officers like the work. The part of the job that made it feel clerical — the copying, the chasing, the reconciling — is the part that goes away. What remains is the part they joined the service for: judgment, exercised on a case that is finally in front of them whole.
Ten minutes of deciding, on three days of preparation nobody had to do.