Your velocity went up. Your estimates stopped meaning anything.

Your velocity went up. Your estimates stopped meaning anything.

Live StratoAtlas Case

A development team adopted AI coding tools mid-project. Velocity jumped. Sprint after sprint, they shipped more than before. But the relationship between a story point and actual delivery time became unstable — not wrong, just unpredictable.

A feature estimated at 8 points might take 20 minutes to generate and 4 days to validate. Or 3 days from the start. The estimate wasn’t broken. It just stopped being a reliable input into any commitment.

—

The Fibonacci sequence doesn’t have a number for “2 minutes plus 3 days of cognitive-heavy review work.”

That’s the concrete form of the problem. A developer debates whether a feature is 5 or 8 points. Then prompts an AI assistant — working code appears in 2 minutes. Then come three days of code review, security validation, integration testing, and refactoring. The planning poker session optimized for the wrong variable.

The acceleration isn’t uniform, which makes it harder to see. Front-end work flows through AI easily. Architectural integrations are faster on first pass — but require extensive manual review. Some 3-point items take three days to validate. Some 8-point items are done before lunch.

—

Here’s the structural break: AI hasn’t made estimation wrong. It has made the underlying assumption untenable.

Estimation works when complexity maps to effort in a roughly stable, unimodal distribution. That’s the assumption that makes velocity a forecasting instrument — if this sprint produced X points, the next sprint will too. Imperfect but directional. Enough to anchor commitments.

AI broke the distribution, not the average. A task’s complexity hasn’t changed — architectural integration is still architecturally complex. But the execution time profile has split into two distinct economic entities: generation work (fast, cheap, scalable — often minutes) and validation work (slow, expensive, cognitively bottlenecked — often days). What was a unimodal distribution around human-hours has become bimodal. No estimation model designed for human-only development has a mechanism for this shape.

—

Teams respond by recalibrating velocity — adding more scope per sprint, adjusting point values, refining their planning poker practice. These are interventions at the calibration layer. The problem is at the assumption layer: the premise that complexity maps to effort in a predictable distribution no longer holds. Recalibrating inside the existing system is the Proxy Fix. The velocity number still comes out. It just stops being anchored to anything.

AI did not break estimation. It broke the assumption that estimation can anchor commitments.

The real cost has shifted from generation to validation. For high-variance tasks — architectural work, security-critical paths, integration-heavy features — commitments should attach to validation capacity, not generation speed. These are different budget lines, different timelines, different contractual structures. The teams discovering this aren’t doing estimation wrong. They’re discovering that the economic entities their contracts were built around no longer exist in the same form.

Full structural diagnosis: https://stratoatlas.com/cases/case-a-ai-2026-034.html

Roman Kir · StratoAtlas Research ORCID: https://orcid.org/0009-0004-2907-9522 stratoatlas.com · CC BY 4.0