On this page
Today I was reviewing the AI training project demos put on by staff at a municipal government’s research and evaluation office. I wrote my comments using speech-to-text (it’s been so long since I wrote a document by hand with no LLM at all, and it felt weird).
My one-line takeaway:
Everyone assumes the older generation and civil servants are behind the times and can’t use AI. Wrong. Every single team built something with a coding agent that they can actually use in their jobs. The experience and judgment veterans carry can deliver enormous value, given the right guidance.
Here’s the bullet summary of the ten pages of written comments I gave them.
Want AI working in your own job, not just in demos? See the 3-month 1-on-1 program →
Accountability
- In AI governance, accountability sits with people, not with AI.
- Responsibility is the capability humans keep for themselves. AI cannot take it over.
- Using logical reasoning to verify AI output is a capability AI cannot take from you.
- The instructor’s role is to run alongside. The designer is the staff member.
Preparation before adoption
- Clarifying specs and tidying up the process before adoption is where the productivity gain comes from.
- Use subtraction to find the shortest path, instead of stacking more layers on the old process.
- Automation is the end of the adoption process. Data and process preparation come first.
- Data governance is the foundation of an AI-friendly structure.
- Distillation extracts decision criteria from past human decisions, and a human confirms what was inferred.
- Fixed templates constrain AI output and prevent format drift.
Verification and risk
- Review the AI conversation logs to catch decision trees that diverge instead of converging.
- Reversible operations are low risk. Irreversible ones need a human in the decision loop.
- Verification speed has to keep up with efficiency gains, and that burden falls on the user.
- The root cause of accuracy problems hides in how the AI produced the output, not in the output itself.
- Reversibility and ease of verification are the two axes that set the priority order for adopting AI in a business.
People and organization
- Taste for soft criteria comes from expert experience. AI’s taste comes from a global average sample.
- The steam engine paradox: efficiency gains increase workload rather than reducing it.
- Learning pace is set by individual motivation, not by how fast the tools update.
- Seed personnel are selected on motivation, logical reasoning, and communication ability.
Why the public sector is actually doing better
Here’s another counterintuitive one: the best AI adoption results I’ve seen so far are in the public sector, and the biggest pushback actually comes from young frontline staff.
I think there’s another reason too. In the private sector, finish early and your boss hands you more work. In the public sector, finish early and you’re simply done early. The time really goes back to the staff, so they’re far more motivated.