Agencies are told they need an AI strategy before they can do anything. So they commission one, it recommends a governance board and a roadmap, and a year later no invoice has been processed any faster.
Start smaller. Pick one workflow that eats staff time, measure it, automate the bounded part, and measure it again.
What makes a good first workflow
- Volume you can count. Invoices, permit applications, records requests, requisitions: something with a number attached.
- Minutes you can time. Sit with the people who do it. Their managers' estimates are usually wrong in both directions.
- A bounded model task. Extracting fields, drafting a response for review, summarizing with citations. Not "decide."
- A system of record. The output lands somewhere that already has permissions, retention and an audit trail.
- A named approver. Someone whose job already includes signing off on this work.
The arithmetic
The formula is deliberately plain: annual volume × minutes per item × fully burdened labor rate × automation share. Twenty thousand invoices at ten minutes each is about 3,300 hours of work a year. If a pilot can take on 65 percent of that time, with a person still reviewing, that's roughly 2,170 hours, or about $97,500 of staff capacity at $45 an hour.
The number isn't the point. The point is that finance, IT and the department can all check it.
What the number doesn't tell you
The biggest number isn't always the right pilot. Every candidate also gets scored on data readiness, integration complexity, records and privacy sensitivity, and political risk. A smaller workflow with clean data and a willing department often beats a larger one with neither.
That's what a three-week assessment is for: five to ten workflows measured the same way, and one recommended pilot. After the pilot, you'll have a before-and-after on your own work, which is the only evidence a strategy should be built on.