Estimation
How Agile teams estimate work — story points to T-shirt sizing. Adapted for AI's inherent uncertainty and experimental nature.
Estimation in AI is not about predicting the future. It is about creating shared understanding of effort, complexity, and unknowns before a single line of model code is written.
Not this
- →Hours to completion
- →A commitment to a date
- →A performance metric
- →Something to optimize
This
- →Relative complexity vs other stories
- →Shared team understanding
- →A planning input
- →A signal when refined wrong
AI-specific inputs
- →Data readiness uncertainty
- →Model performance unknowns
- →Spike results before estimation
- →Confidence calibration (40–80%)
The Three Dimensions of a Point
Effort
How much work is required — coding, data engineering, model training, feature engineering, deployment.
Complexity
How difficult is the problem — algorithm choice, interdependencies, edge cases, model interpretability.
Uncertainty
How much is unknown — data availability, model convergence, label quality, production behavior.
Why Fibonacci?
The gaps between Fibonacci numbers grow non-linearly — mirroring how estimation uncertainty scales. The difference between a 1 and 2 is meaningful. The difference between a 20 and 21 is not.
This non-linearity is especially apt for AI work, where complexity often scales exponentially — a 13-point story is not "13 times harder than a 1" but it is genuinely in a different risk category.
The Series
Select a Fibonacci value to see what it means for AI work and when to use it.
1
Well-understood task. Minimal complexity. No meaningful unknowns.
AI Examples
Update model logging level in production config
Fix a typo in a model output label
Add a field to the monitoring dashboard
PM Signal
"If a story cannot hit 1, it should be a sub-task, not a story."
Fibonacci Quick Reference
Planning Poker is the ceremony where the team aligns on story points. Five steps — each with a common failure mode in AI teams.
PO reads the story
Read aloud, answer clarifying questions. Only proceed when the team has no more questions. For AI stories, questions often involve: what dataset, what accuracy target, what happens at low confidence.
AI-Specific Note
If questions cannot be answered in 5 minutes, the story is not ready. Do not force estimation — put it back in the backlog.
Session Rules
No story without a DoR
Definition of Ready: AC written, dependencies named, team has seen it in refinement.
5-minute max per story
If it takes longer, the story is not ready. Stop estimating, start refining.
Simultaneous reveal
Never let one person anchor before others reveal. All cards at once.
? means spike, not pressure
A team member who cannot estimate is telling you something important. Listen.
Four estimation techniques — each with a distinct job. Match the technique to the planning level.
Story points
Sprint-level planning for refined User Stories
Best For
Detailed execution — when requirements are clear enough to estimate effort, complexity, and uncertainty.
How
Fibonacci scale (1, 2, 3, 5, 8, 13, 21). Relative to a reference story the team has agreed on.
Not For
Roadmap planning, epics, or any story the team cannot confidently estimate in 10 minutes.
AI-Specific Note
In AI, uncertainty is a legitimate dimension of a point score. A story with "known effort but unknown model outcome" is legitimately an 8 even if the engineering work is only a 3.
T-shirt Size Quick Reference
1–3 pts
Fix monitoring dashboard label
5–8 pts
Retrain model on updated dataset
13–21 pts
Integrate pre-trained NLP model
34–55 pts
Build intent classifier from scratch
Epic → decompose
Build full recommendation engine
Program-level
AI-powered product line initiative
How do you run a story point estimation session?
The session works only if stories are refined before it starts. I never open Planning Poker with a story that has unresolved acceptance criteria — we review the refinement backlog in the days before the session and flag any story that is not ready. In the session: I read the story aloud, answer questions, and give the team 60 seconds to privately select a Fibonacci card. Cards reveal simultaneously. For close estimates (3, 3, 5) we take consensus and move on in under 5 minutes. For wide divergence (2 and 13), the outliers explain their reasoning — this almost always surfaces a data dependency or model uncertainty that one person factored in and another did not. If a ? card appears, the story goes back to refinement with a spike added. I aim to process 10–15 stories per hour.
Common Trap
Using Planning Poker as a design session. Estimation time should be minutes, not hours. If the team is designing during estimation, the story was not ready. Stop, refine, reschedule.
What is the difference between story points and T-shirt sizing, and when do you use each?
They operate at different levels of the planning hierarchy. T-shirt sizing runs at Feature and Epic level during roadmap planning — before stories are even written. It gives directional effort without false precision. A 3-month roadmap planning session should use T-shirts, not points. Story points run at User Story level during sprint planning. They require the story to be refined enough for the team to evaluate effort, complexity, and uncertainty. A feature sized "M" in T-shirt terms typically breaks into 15–25 story points once decomposed. The mistake is using story points for features that are not yet decomposed, or T-shirt sizing for sprint-ready stories. Match the tool to the planning level.
Common Trap
Mixing sizing levels — pointing un-decomposed features or T-shirt sizing sprint stories. It creates false precision at one level and false ambiguity at another.
How do you estimate when you do not know whether the model will hit its performance target?
Unknown model performance is the most common estimation blocker in AI work. My default is to never estimate a delivery story for an untested model — I add a spike first. A spike is time-boxed (3–5 days) with a binary deliverable: "model architecture is feasible at X accuracy" or "it is not, and here is why." This spike gives the team enough information to estimate the delivery story with meaningful confidence. If the spike is not possible before sprint planning, I use three-point estimation: optimistic (model works first try), most likely (one architecture change needed), pessimistic (data quality forces 2 additional sprints). I present this range to stakeholders rather than a single number.
Common Trap
Committing to a delivery story on an untested AI model. The sprint will blow. Add the spike, present the range, and set the expectation that commitment follows validation.
What do you do if your team consistently misses estimates?
Consistent misses are a systemic signal, not a skill gap. My first step is a blameless retrospective focused on root causes, not individuals. The three most common root causes in AI teams: stories enter the sprint underrefined (acceptance criteria unclear, data dependency unresolved); velocity is being used as a target rather than a measurement, creating pressure to over-commit; or the team has unplanned work — production incidents, model regressions, ad hoc data pulls — consuming 30–40% of sprint capacity that is not reflected in velocity. Fixes: Add a Definition of Ready gate before stories can be pulled into a sprint. Protect sprint capacity by capping unplanned work at 20%. Recalibrate velocity to reflect actual available capacity. Re-establish that velocity is a planning tool, never a performance metric.
Common Trap
Responding to missed estimates with pressure to be more precise. Precision without clarity is guessing faster. Fix the refinement process first.
How do you use velocity as an AI PM, and what is it not?
Velocity is a historical average of story points completed per sprint. I use it for one purpose: capacity planning. If the team averages 28 points over 6 sprints, I plan next sprint at 28 points — not 35 because the deadline is close. Velocity is not a performance metric. Treating it as one creates incentives to inflate estimates, which corrupts the signal and destroys its value as a planning tool. In AI teams, velocity is especially unreliable as a short-term signal because model experimentation and spike work have highly variable outcomes. I smooth over at least 6 sprints and flag any sprint with more than 30% unplanned work as an outlier in the average.
Common Trap
Setting a velocity target ("we need to hit 35 points this sprint"). This immediately shifts the team from honest estimation to gaming the number — stories get inflated, spikes get skipped, and the metric becomes meaningless within 2 quarters.
Check yourself
Four short questions, about 2 minutes. Pass mark 3 of 4.