How should a local government buy AI?

Short answer: treat AI like any other piece of public infrastructure. Require auditability, data ownership, and a clean exit; pilot on one narrow workflow before you commit; and keep a named human accountable for every consequential decision. The vendor demo is not the deliverable — the audited, in-production workflow is.

What to require in any AI buy

• Auditability — you can inspect the inputs, outputs, and basis of each decision.
• Data ownership and privacy — your data stays yours; no training on it without consent; clear retention and deletion.
• Human-in-the-loop — the system assists; named staff review and decide anything consequential.
• Exit and portability — you can leave with your data and prompts; no lock-in.
• Security — breach notification, access controls, and compliance with your records obligations.

How to pilot

1. Pick one narrow, high-volume, low-risk workflow.
2. Define what “working” means before you start.
3. Run a short, time-boxed pilot with a human reviewing every output.
4. Decide on measured results, not the demo.

Questions to ask any AI vendor

• Where does our data go, and is it used to train your models?
• Can we see and export the logs behind every decision?
• What happens to our data and workflows if we leave?
• Who is accountable when the model is wrong?

Key takeaways

• Buy AI like infrastructure: auditable, owned, and exit-able.
• Pilot narrow; scale on evidence.
• Keep a human accountable for consequential calls.