Getting AI right in the public sector
AI can genuinely lift the quality and reach of public services. But in government the order of operations matters: governance, accuracy and accountability have to come before automation — not after it.
I came to AI from the other side of the counter. Before I built an AI company, I ran public organisations that answered phones, took reports, dispatched crews and explained decisions to residents who were often having a bad day. That experience makes me both an optimist and a sceptic about AI in the public sector — optimistic about what it can do, sceptical about the way it is often sold.
The optimism is easy to justify. A large share of public-sector contact is repetitive, time-sensitive, and after-hours: a blocked drain, a lost dog, a road hazard, a question about a permit. People deserve to reach a service that responds accurately at 2am, not a recorded message telling them to call back on Monday. Done well, AI extends the reach of a service without diluting its quality. It is not about replacing the people who do this work — it is about making sure the community can always get through.
Why "move fast and break things" is the wrong model
The scepticism is about how that capability gets deployed. Consumer technology culture rewards shipping quickly and fixing problems later. That instinct is actively dangerous in government, where "breaking things" means a vulnerable person given wrong information, a false promise of a callback that never comes, or a decision no one can explain after the fact. The public does not get to opt out of its council or its agencies. That lack of an exit is exactly why the standard has to be higher.
So the question I care about is not "can AI do this?" — increasingly it can. The question is "can we stand behind what it does?" Three things determine the answer.
1. Governance before automation
Before a single call is automated, you need to know who owns the outcome, what the system is and is not permitted to do, and how a mistake is caught and corrected. In practice that means the machine owns the parts that must be reliable — what gets promised, what gets dispatched, what gets escalated — through deterministic rules, not the improvisation of a language model. The AI's freedom should be greatest where the stakes are lowest, and tightly bounded where they are highest. Governance is not paperwork you complete afterwards; it is the design.
2. Accuracy you can prove
An AI service that is usually right is not good enough when the cost of being wrong is someone's safety or someone's trust in government. Public-sector AI has to be grounded in the organisation's real information — its rules, its service areas, its escalation paths — and it has to be tested the way a serious operation tests anything: against reality, repeatedly, with the failures fed back into the design. Confident and wrong is the failure mode that erodes public trust fastest, and it is the one that generic AI produces most easily.
3. Accountability that survives scrutiny
Government decisions get audited, questioned and debated in public — and they should. Any AI operating in that environment has to leave a record: what was said, what was decided, why. If you cannot reconstruct and explain an interaction after the fact, you should not have automated it. Transparency is not a constraint on public-sector AI. It is the thing that makes it legitimate.
The standard is human, not robotic
The bar I hold for AI in public services is simple: it should meet the standard of a good human professional on their best day — accurate, warm, honest about what it can and cannot do, and never making a promise it cannot keep. That is a high bar, and most deployments do not clear it yet. But it is the right bar, because the public did not choose to be a test case.
AI belongs in the public sector. It just has to earn its place there the same way any public servant does — by being reliable, accountable, and genuinely in service of the community. Get that order right and the technology is a gift. Get it backwards and it is a liability wearing a friendly voice.