Agentic AI the public sector can audit

Government agencies manage the longest-lived assets and the most scrutinised budgets in the built environment. When AI enters that work, the questions are correctly harder: Who approved this number? What method produced it? Can we show the reasoning to an auditor, a minister, a court? Any AI that cannot answer is unusable in public work, whatever its accuracy.

Three tests for public-sector AI

• Provenance: every output traces to the source documents and the method that produced it — a named correlation, a stated measurement rule, a defined reconciliation pair.
• Approval: the AI proposes; an accountable person disposes. Approval chains are explicit, delegations are recorded, and no posting happens without a human decision.
• Audit: every action leaves a trail — what was extracted, what was proposed, who approved it, when, and what changed downstream.

Agentic does not mean autonomous

The word 'agent' suggests software acting alone. In professional work it should mean the opposite of a black box: a system that does the assembling, reconciling and drafting under rules you can read, then stops at the decision boundary. The decision belongs to the officer, the engineer, the reviewer. This is not a limitation added for the public sector; it is how the systems should work everywhere.

Where to start

The lowest-risk, highest-value entry points are workflows that are document-heavy, rule-rich and audit-critical: cost submissions reconciled against contracts; investigation data structured into reviewable models; claims lifecycles with every certification logged. These are the workflows Greenshift builds agent teams for — designed from the first line for provenance, approval and audit, because our private-sector users carry liability too.