Story points have been the default way agile teams talk about effort for years. They work because they are relative, fast, and force a conversation about scope and risk. But for teams that build with AI coding tools, a single story-point score no longer captures everything that matters about a ticket. There is a second dimension now: expected AI token usage.
What story points describe
Story points describe implementation effort, complexity, and risk. They answer the question “how hard is this for our team to build?” That signal is still essential. It drives capacity planning, surfaces hidden complexity during the reveal, and gives the team a shared vocabulary for comparing work.
What token estimates describe
Token estimates describe expected AI usage. They answer a different question: “how much AI prompting, context, and model usage will this likely consume?” In AI-assisted development, that maps directly to cost and to how the work actually gets done.
The key insight is that effort and AI usage are not the same axis. Two tickets can share a story-point score and have wildly different token profiles:
- A small, well-scoped change the model nails in one short prompt — low effort, low tokens.
- A small-looking change buried in a large, ambiguous codebase that needs repeated context and correction — low effort, high tokens.
- A large human task where AI mostly accelerates boilerplate — high effort, moderate tokens.
Why AI teams may need both
When you track only story points, the AI-cost dimension stays invisible until the bill arrives. Tracking both makes it a planning input. A ticket that is cheap in effort but expensive in tokens can be scheduled, scoped, or approached differently — maybe with a smaller model, tighter context, or a human-led implementation.
Keeping them as two separate estimates is deliberate. It preserves the familiar story-point discussion while making AI cost explicit rather than a hidden assumption in one engineer's head. If you are new to the token side, our guide to token estimation tiers walks through a simple deck your team can vote on.
How to start
You do not need a new process. Run planning poker the way you already do, and add a second vote for expected token usage on each ticket. When the two estimates diverge, treat that gap as a signal worth discussing. For a step-by-step walkthrough, see how to estimate AI coding costs during sprint planning. That is exactly what Token Poker is built for: story points and token estimates in the same room, revealed together.