AI in public systems: the hard part is not the model

I keep seeing the same pattern in conversations about AI and government. The discussion quickly becomes about the model. Which one is better? Which one is safer? Which one can summarize, classify, translate, predict, or generate at lower cost?

Those are fair questions. They are just not the questions that usually decide whether an AI project works.

In public systems, the harder problem is the system around the model. The data may be incomplete. The workflow may be unclear. The people expected to use the tool may not trust it, or may trust it too much. Procurement may reward pilots more than maintenance. A dashboard may exist without changing a single decision.

That is why I think of most public-sector AI work as an implementation problem first.

What this changes

If AI is treated mainly as a model-selection exercise, the project starts too late. By the time the team is comparing tools, many of the important choices have already been made badly or not made at all.

The better starting point is more basic:

• What decision are we trying to improve?
• Who makes that decision now?
• What information do they already trust?
• Where does the current process break down?
• What should happen when the AI system is wrong?

That last question matters more than people like to admit. Public institutions do not get to fail in the abstract. A bad recommendation can delay a service, distort a program, exclude someone, or simply waste scarce attention.

What I would rather see

I would rather see smaller AI projects with clearer judgment than bigger pilots with vague ambition.

Use AI to help a team read a pile of reports faster. Use it to compare field notes across states. Use it to draft first-pass classifications that a human reviews. Use it to find contradictions in policy documents. These are not glamorous examples, but they are useful. They also force the institution to answer practical questions about data, review, accountability, and maintenance.

The best uses of AI in public-interest work will probably look boring from the outside. That is fine. Boring tools that survive contact with institutions are more valuable than impressive demos that die after the workshop.

The lesson I keep coming back to

AI does not remove the need for institutional judgment. It exposes whether that judgment exists.

If a team cannot explain the decision it wants to improve, AI will not fix that. If the data is politically contested, AI will not make it neutral. If nobody owns the workflow, the model will become another orphaned tool.

But where the problem is clear, the data is understood, and people know how decisions should be made, AI can help. It can reduce friction. It can make evidence easier to use. It can give busy teams a better first pass at work they already know how to judge.

That is the space worth building in. Less spectacle. More useful judgment, better supported.