AI token estimation for agile teams

AI token estimation is the practice of forecasting, before work begins, roughly how many AI tokens a task is likely to consume. Tokens are the unit of usage for large language models — the input context you feed the model plus the output it generates — and they map directly to the cost of AI-assisted development.

Instead of waiting for the invoice, teams put a rough number on expected usage during planning. The goal is not precision to the token; it is a shared, early signal that a ticket might be cheap or expensive in AI terms.

Once AI tools are part of delivery, token usage affects real decisions: how much a feature costs to build, which model to use, how much prompting and context a task needs, and whether a workflow is a good automation candidate. Surfacing that in planning turns a hidden assumption into a deliberate conversation.

It also helps prioritization. A ticket that is low effort but high token burn may be worth scheduling differently than its story-point score alone would suggest.

Treat it like story points: relative, not absolute. Pick a few reference tickets your team has already shipped with AI, anchor them at known token tiers, and estimate new work against those anchors. Token Poker ships with sensible default tiers — from a tiny prompt up to “call finance” — so teams can vote quickly.

When estimates diverge, that gap is the valuable part. It usually means people are assuming different amounts of context, different models, or a different implementation path. Talk through the disagreement, then lock the estimate.

In Token Poker, every ticket gets two votes: a classic story-point estimate for effort and a token estimate for expected AI usage. The reveal shows both at once, so the team leaves with a view of how hard a task is and how AI-heavy it is likely to be — in the same round, with no extra meeting.