Guide · Estimation
Software estimation methods: which to use, and when.
Planning poker, story points, t-shirt sizes, three-point and PERT estimates, and simulated 50/80/90% dates. What each one answers, where it goes wrong, and how to combine them.
In short. Software estimation methods fall into two groups: ways to size work, such as story points, t-shirt sizes and planning poker, and ways to turn sizes into dates, such as three-point estimates, PERT and simulation. For a date you can commit to, estimate each part as a best, likely and worst case and read the result as a range.
What an estimate is for
An estimate is a statement about uncertainty, not a promise. “Two weeks” sounds like an answer, but it hides the question that matters: how likely is it to be three? Every method below is a way of making that uncertainty visible, or of ignoring it.
Two different questions get mixed up under one word. How big is this? is about comparing pieces of work. When will it be done? is about dates, and needs a calendar, a team and some honesty about risk. Relative sizing answers the first. Ranges answer the second.
Estimate parts, not features. Nobody can estimate “team invitations”; people can estimate “invite email”, “accept-invite endpoint” and “seat-limit check”. Breaking the feature down first is the single biggest improvement, and it is the step the feature planning checklist feeds.
The methods, side by side
Each of these is a legitimate choice. They answer different questions.
- Analogy. “This is like the export feature we built last spring.” Fast, and as good as your memory and the similarity. Best for a first rough idea.
- T-shirt sizes (S, M, L). Sort work into a few buckets without arguing over numbers. Good for roadmap-level planning; too coarse for a sprint commitment.
- Story points. A relative size on a scale, usually Fibonacci. Points compare stories with each other and, once a team has a track record, with the team’s speed. They are not hours and not days. See story points explained.
- Planning poker. A way to produce story points (or days) as a group, with hidden votes so nobody anchors on the first number. See below.
- Three-point estimate. For each part: a best case, a likely case and a worst case. Three numbers instead of one, and the gap between them is the information. Try it in the free estimation tool.
- PERT. A weighted average of the three-point estimate: (best + 4 × likely + worst) ÷ 6. A single expected value per part.
- Simulation. Run the whole plan thousands of times, drawing each part from its range, and read off the dates you reach in 50, 80 or 90% of runs.
Three-point and PERT estimates
Ask the engineers who will build a part for three numbers, in days: best if nothing goes wrong, likely if it goes as usual, worst if the known risks happen. Not the worst imaginable, the worst you would be surprised to beat.
PERT turns the three into one expected value: (best + 4 × likely + worst) ÷ 6. If a part is 2 / 3 / 8 days, the PERT value is (2 + 12 + 8) ÷ 6 = 3.7 days. It sits above the likely case because the worst case is further away than the best case, which is how real work behaves: tasks overrun by more than they underrun.
The spread is as useful as the average. A part at 2 / 3 / 4 is well understood. A part at 2 / 3 / 12 is where to spike first, because that is where the surprise will come from. Sort your parts by spread and the riskiest ones are at the top.
Where PERT falls short is the total. Adding up the PERT values of every part gives an expected total, but it says nothing about how likely you are to finish by a given date. For that you need the next step.
Ranges and 50/80/90% dates
Add up the likely cases of ten parts and you get a number that is rarely met, because ten things going as usual is not the usual outcome. Some run over, a few run under, and the overruns are bigger. The total has to reflect that.
The way to do it is to simulate: for each part, draw a duration from its best-likely-worst range, add the parts, and repeat thousands of times. Sort the results. The 50th percentile is a date you hit half the time. The 80th is a date you hit four times out of five. The 90th is one you hit nine times out of ten.
Then use each date for what it is good for:
- 50% is for planning. It is the middle of the range, and a coin flip as a promise.
- 80% is for commitments to others: a customer, a launch partner, another team.
- 90% is for dates that are expensive to miss, such as a legal deadline or an event.
Planning poker
Planning poker is a group technique for sizing. Everyone chooses a card privately, all cards are revealed together, and the people with the highest and lowest cards explain their reasoning before the team votes again. The point is not the number. It is the conversation when two people are three cards apart, because that usually means they pictured different work.
It works best on small, well-described stories, with the whole team who will build them in the room. It works worst on large, vague features, where you will vote on something nobody has pinned down. Break the work down first.
The usual deck is Fibonacci (0, 1, 2, 3, 5, 8, 13, 21), because bigger work is harder to size precisely and nobody should argue about 13 versus 14. You can run a room for free with no sign-up at planning poker online.
Try planning poker online for free, and see how to calculate story points when your team votes in points.
Which method to use
A practical combination for most product teams:
- Rough scope: use analogy or t-shirt sizes to decide whether a feature is worth planning at all.
- Break it down into parts a single person could build in a few days.
- Give each part a best, likely and worst case from the people who will build it. Planning poker can help the team converge on the likely case, and it surfaces disagreement about scope.
- Simulate the total and read the 50, 80 and 90% dates. The free estimation tool does this from your three numbers per part, with no sign-up.
- Commit to the 80% date, plan on the 50% one, and write down which parts carry the spread.
- Re-estimate when something changes, and keep the old estimate to compare with what happened.
Common mistakes
Most bad estimates fail in one of these ways.
- Estimating the feature, not the parts. The error is in what you didn’t list.
- Estimating alone. The person who knows the code best is usually the worst at guessing how long it takes someone else.
- A best case that is really the likely case. If your best and likely are equal, ask what would make it faster. If nothing could, the range is too narrow.
- Forgetting the work around the work. Review, testing, deployment, migrations, documentation and meetings. Add it as its own line, or as overhead.
- Treating a range as a number. Someone will pick the low end and call it the date. State which confidence the date is at.
- Re-estimating mid-sprint with no decision behind it. When new work arrives during a sprint, treat it as a sprint scope change: check it against the goal, size it, then swap something out.
- Never looking back. Compare estimates with actuals on a few finished features. It is the only way a team’s estimates improve.
Doing this in SprintProof
SprintProof breaks a feature into parts, asks for a best, likely and worst case for each, and runs the simulation for you. Parts tend to overrun together, so each simulated run shares part of its overrun across the parts, and dates assume engineers work in parallel about 70% of the time. It shows the range, the dates at 50, 80 and 90%, and the parts with the widest spread. Estimates can be given blind by each person on the team and revealed together, which is planning poker without the meeting. Blind estimates cost no credits.
The same simulation powers the free estimation tool. The AI suggests ranges; people accept them, and nothing changes without a click. See pricing for the Team plan.
Plan the feature, not just the document.
Describe the feature and SprintProof interviews you, drafts the breakdown, asks the questions reviewers ask, and gives an estimate as a range with 50, 80 and 90% dates.