How to budget for AI without guessing.

Most AI budgets are set by anchoring on a vendor quote, which means the vendor sets your budget. Work it out from your own numbers instead — it takes twenty minutes.

Start from the cost of the work, not the cost of the software

Pick the workflow. Count the people who touch it, the hours a week they spend, and their loaded hourly cost. Multiply by 52. That annual figure is the size of the prize, and every spending decision should be a fraction of it.

A sensible first-year allocation is a small slice of that number — enough to prove the workflow, not enough to hurt if it fails.

The three-part budget

Build (one time). The fixed cost of the thing itself. In our ladder that is $1,500 to $7,500 for a build, $24K–$95K for a production agent.

Run (recurring). Model usage and infrastructure. Usually modest, and falling — model prices have dropped sharply and keep dropping.

Keep (recurring). The one everyone forgets. Industry guidance puts maintenance at 15–20% of build cost per year, and for anything connected to systems that change, that is realistic rather than padding.

What to hold back

Reserve roughly a third of your first-year allocation for the second decision. The most valuable thing the first build produces is information about what to build next, and a budget with nothing left cannot act on it.

The question your CFO will ask

“What happens if it does not work?” Have an answer before the meeting. Ours is written into the terms: a refund-backed diagnosis, and if an agreed metric is missed we keep working at no additional fee for up to 60 days against the baseline set at kickoff. Whatever vendor you choose, get their version of that answer in writing.