AI coding tools changed the economics of building software. Usage is metered, so two features that take the same engineering effort can cost very different amounts to ship. The good news: you can bring that cost into sprint planning without adding a heavy new process. Here is a practical approach.
1. Make AI cost a planning signal, not an afterthought
Most teams discover AI cost after the fact, when someone looks at the monthly invoice. The fix is to estimate expected token usage at planning time, alongside story points. You are not trying to predict the exact number — you are creating an early, shared signal that a ticket is likely cheap or expensive in AI terms.
2. Estimate relatively, with reference tickets
Token estimation works best the same way story points do: relatively. Pick a few tickets you have already shipped with AI and anchor them to known token tiers. Then estimate new work against those anchors instead of trying to compute tokens from scratch. If you want a ready-made deck, the token estimation tiers explained post breaks down each tier with examples.
A simple set of tiers your team can vote on:
- Tiny prompt / quick chat — a focused, well-scoped change.
- Normal ticket — a standard feature with some iteration.
- “Let it cook” — meaningful context and back-and-forth.
- Repo archaeology — agent exploration across a large codebase.
- Call finance — the ticket nobody wants on the invoice.
3. Treat disagreement as the valuable part
When token estimates diverge across the team, do not just average them. The gap usually means people are assuming different amounts of context, different models, or a different implementation path. Talk through why, align on an approach, and then lock the estimate. That conversation is often more useful than the number itself.
4. Turn estimates into decisions
Once you have token estimates next to story points, use them:
- Re-prioritize low-effort, high-token tickets — they may not be the cheap wins they look like.
- Choose a cheaper model or tighter context for high-token work where quality allows.
- Flag context-heavy tickets for a human-led approach instead of full agent automation.
- Roll the estimates up into a rough AI budget forecast for the sprint.
5. Keep it lightweight
The whole point is to add one signal, not a new ceremony. In Token Poker, the token estimate is a second vote in the same planning round you already run — revealed alongside story points, so cost awareness becomes a natural part of how the team plans. New to running both estimates together? Start with why AI teams may need both story points and token estimates.