What AI should and should not do in a command centre
Agents are excellent at clustering, summarising and noticing. They should not be the last decision before something reaches the public.
Civiq.plus research · AI governance
The useful question about AI in civic systems is not whether to use it. It is which specific steps in a workflow it should own outright, which it should draft, and which it must not touch. Answering that per step produces a much better system than answering it once for the whole product.
Where agents earn their place
Deduplication is the clearest win. A ward generating four hundred reports about the same broken line is generating one problem, and clustering those correctly is tedious, high-volume and well suited to a model. So is transcription and classification of voice reports in regional languages, which is otherwise a hard staffing constraint on field reporting.
Anomaly detection is a second. A rolling average hides a spike concentrated in three streets; a model looking for geographic concentration does not. Narrative clustering and propagation analysis are a third — tracing how a claim moved across channels is graph work at a scale humans cannot do by reading.
Where they must stop
Agents in Civiq.plus do not publish high-impact material unless an explicit policy and approval allow it. Broadcasts, sensitive messages and high-impact recommendations pass through a human approval gate, and the gate is not a formality — the approver sees the evidence, the assumptions and the confidence, not just the output.
The registry makes this enforceable rather than aspirational. Each agent definition carries its data scopes, geographic scope, model policy, cost and latency ceilings, approval policy and safety policy. An agent cannot quietly acquire access to a district it was not scoped for.
The audit is the product
Every agent execution records the input references and data scope, the model provider and version, the tool calls and their outputs, the generated recommendation, the stated confidence and assumptions, the human approval or rejection, the final action taken, and the subsequent measured impact.
That last field is the one that changes behaviour over time. An agent whose recommendations are consistently approved and consistently fail to move the underlying condition is visible as such, which is a far more useful signal than any offline evaluation.
A non-goal worth stating
The platform does not create covert individual-level psychographic scores or hidden persuasion profiles, and does not fabricate public support or automate inauthentic engagement. These are architectural constraints rather than settings, because a capability that exists will eventually be used by someone under pressure.
