Estimation techniques that actually work.
Every team estimates. Most teams argue about how. This is the short, opinionated tour of the techniques that are actually in use — what each one is good at, where it breaks, and how to run the one that won.
By the Pointing team12 min readUpdated July 2026
Estimates are a conversation, not a promise.
The useful output of an estimation session is almost never the number. It's the ten seconds after the reveal, when someone who voted 2 and someone who voted 13 discover they were describing two different pieces of work.
That's the whole trick. A technique is good when it surfaces disagreement cheaply and early, and bad when it manufactures a number precise enough that nobody questions it. Judge the ones below on that, not on how scientific they sound.
An estimate is
A shared reading of size, uncertainty, and complexity — produced by the people who'll do the work, in the room, in minutes.
An estimate isn't
A deadline, a contract, a performance metric, or a number of hours. The moment it becomes any of those, teams start inflating and the signal dies.
So optimize for
Speed and disagreement. A fast technique you run every sprint beats an accurate one you run twice a year and then quietly abandon.
Ten techniques, honestly compared.
Roughly ordered from most to least common. Almost every team ends up running two: something fast and coarse for the roadmap, something sharper for the next sprint.
Planning poker
Everyone picks a card in secret, all cards flip at once, and the spread drives the discussion. Re-vote until you converge.
- Best for
- Sprint-sized items, estimated by the people who will actually build them.
- Watch out
- It decays into ritual if you never recalibrate against a story the team has already shipped.
In Pointing
Native. Hidden votes, one-tap reveal, distribution, and consensus grading on any deck.
Try it free — Planning pokerT-shirt sizing
Size items XS → XXL instead of using numbers. Deliberately coarse, deliberately un-mathematical.
- Best for
- Roadmap-level conversations, mixed audiences, and anything too vague to point.
- Watch out
- Somebody downstream will map sizes back to numbers anyway. Agree that mapping out loud, once, or don't let it happen.
In Pointing
Native — the T-shirt deck is built in. Sizes aren't numbers, so reveal stats stay off by design.
Try it free — T-shirt sizingBucket system
Fixed buckets (usually a Fibonacci row). One person drops an item into a bucket, the group moves it if they disagree. Repeat, fast.
- Best for
- A first pass over 50+ items when you need coverage more than precision.
- Watch out
- The speed comes from suppressing discussion. Expect to properly point the buckets that reach a sprint.
In Pointing
Backlog mode: load up to 500 items and work the list in order with live progress.
Try it free — Bucket systemAffinity mapping
Silently group similar items together first, then size the groups rather than the items.
- Best for
- A brand-new backlog, before the team shares any vocabulary for size.
- Watch out
- Groupings feel objective but mostly encode whoever moved the first card.
In Pointing
Backlog mode — bring your clusters in as items and point each cluster once.
Try it free — Affinity mappingWideband Delphi
The ancestor of planning poker: anonymous written estimates, a moderated discussion of the outliers, repeat until the range narrows.
- Best for
- Unfamiliar, high-risk work where you want experts to disagree in writing before anyone speaks.
- Watch out
- Slow by design. That's the point, and also the problem.
In Pointing
Planning poker is the compressed version — vote, reveal, discuss outliers, re-vote.
Try it free — Wideband DelphiThree-point / PERT
Estimate best case, most likely, and worst case, then combine them — classically (O + 4M + P) / 6 — and keep the range.
- Best for
- A handful of high-stakes items where you must express uncertainty explicitly.
- Watch out
- Three invented numbers are still invented. The range is the deliverable; the weighted mean is decoration.
In Pointing
Build a custom deck and run three quick passes — optimistic, likely, pessimistic — then export the numbers.
Try it free — Three-point / PERTRelative mass valuation
Lay the whole backlog out and sort items against each other by size before assigning any number at all.
- Best for
- Seeding a new backlog with internally consistent sizes in one sitting.
- Watch out
- Needs the entire set visible at once. It does not work item-by-item, which is how most backlogs arrive.
In Pointing
Backlog mode gives you the ordered list; point it top to bottom once the order feels right.
Try it free — Relative mass valuationIdeal days / ideal hours
Estimate uninterrupted working time, then apply a focus factor to get calendar time.
- Best for
- Teams contractually obliged to report in a time unit.
- Watch out
- The focus factor becomes a negotiation, and time estimates get read as commitments by everyone outside the room.
In Pointing
Custom deck — e.g. ½, 1, 2, 3, 5 days. Custom decks need at least one numeric card so reveal stats still work.
Try it free — Ideal days / ideal hoursDot voting
Everyone spends a fixed number of dots across items. A prioritization tool that regularly gets drafted into estimation.
- Best for
- Deciding what to estimate next — not how big anything is.
- Watch out
- 'Most wanted' and 'biggest' are different axes. Conflating them produces a roadmap nobody can schedule.
In Pointing
Not an estimation mode here. If you want a bounded vote, the Linear 1–10 deck does the job.
Try it free — Dot voting#NoEstimates / throughput
Stop sizing. Slice work to roughly uniform pieces, count what finishes each week, and forecast from your own history.
- Best for
- Teams with a genuinely disciplined slicing habit and clean historical data.
- Watch out
- The slicing discipline it depends on is usually the thing teams learned from estimating in the first place.
In Pointing
Keep pointing to build the habit, export sessions to CSV or JSON, and forecast from your real numbers.
Try it free — #NoEstimates / throughputTry it in 5 seconds
Reading about estimating is slower than estimating.
Open a room, share the link, and run one item with your team. No account, no credit card, no setup.
Why planning poker won.
It is not the most accurate technique on this page. It is the one teams keep doing after the enthusiasm wears off — and a technique you still run in month nine beats a better one you abandoned in week three.
Simultaneous reveal kills anchoring
Nobody hears “I think it's a 3” before committing. Every other technique that fixes anchoring costs more time than this one.
The disagreement is the product
A 2 and a 13 on the same item is the cheapest bug report you'll ever get: two people are describing different work.
Relative sizing plays to human strengths
People compare two things well and predict durations badly. Points ask for the comparison and skip the prediction.
It is bounded
A deck of roughly eight options makes the decision small. That's why a sprint's worth of backlog fits in an hour.
Everyone votes, so everyone owns it
An estimate the team produced is far harder to wave away than one the tech lead produced — including by the team.
It survives remote
It's the rare agile ceremony that's genuinely better online: no table, no tells, no cards to hand out.
Where it falls down
- Precision theatre — arguing 5 versus 8 usually costs more than being wrong would have.
- Points leaving the room and getting converted to hours by someone who wasn't in it.
- Discussion before the reveal, which is anchoring with extra steps.
- Estimating badly-sliced work. A 21 is a signal to split the item, not to estimate harder.
Where it came from
Planning poker descends from Wideband Delphi, the consensus estimation method Barry Boehm's team refined in the 1970s. The card-based version was described by James Grenning in 2002 as a way to escape hours-long estimation meetings, and reached the mainstream through Mike Cohn's Agile Estimating and Planning in 2005.
Planning poker, free
Hidden votes, one-tap reveal, real consensus.
Everything this section describes is the default behaviour of a Pointing room — on any deck, with up to 100 people.
The playbook: ten rules that fix most sessions.
If you adopt nothing else on this page, adopt rules 1, 3 and 9. They're responsible for most of the difference between a session people attend and a session people dodge.
Calibrate with two reference stories first
Pick a known small and a known medium the team already shipped. Every estimate is relative to those; without anchors the scale drifts every sprint and your history becomes unusable.
Read the item, then vote — no discussion first
Pre-reveal discussion is anchoring wearing a lanyard. Whatever needs saying will be said better once people have committed.
Timebox discussion to two minutes
The goal is to surface information, not to reach agreement by attrition. If two minutes doesn't do it, information is missing — see rule 5.
Outliers talk first
Highest and lowest explain their card. That is where the hidden requirement, the forgotten migration, or the misunderstanding lives.
Re-vote at most twice
A third split isn't a disagreement about size, it's missing information. Log the question, park the item, or spike it — don't grind.
Split anything above 13 (or XL)
A big number is a slicing signal. An item that can't finish inside a sprint isn't an item, it's a project wearing one.
Estimate the work, not the person
“How long would it take you” produces one scale per person. Size the story for whoever picks it up, which is how it'll actually get assigned.
Make ? and ☕ mean something
? means “I need information”, not “I abstain”. A cluster of ? is a hard stop, and it should feel like one.
Never convert points to hours
The instant a point has an hour value, the estimate is a commitment and the team starts inflating defensively. Forecast with a velocity range instead.
Re-estimate only when scope changes
Not when work runs late. Re-pointing late work retroactively destroys the historical data your forecasts depend on.
Straight to the bin
- Averaging to a decimal
- The lead votes first
- Points as a productivity metric
- Re-pointing to fix a burndown
- The whole department in the room
- Estimating work nobody starts this quarter
- One person estimating for everyone
- Comparing velocity across teams
Pick the scale that fits the decision.
The scale is not a detail — it's the whole vocabulary your team will use to argue. Pointing ships five decks and lets you build your own; here's when each one earns its place.
Fibonacci
The default, and the right default. The gaps widen as items get bigger — which is exactly how your confidence behaves.
Use this deck — FibonacciModified Fibonacci
Adds ½ and stretches to 100. Use it when one board has to hold both a typo fix and a quarter-sized bet.
Use this deck — Modified FibonacciT-shirt
No numbers, no false precision, no arithmetic. Best for roadmap conversations and rooms with non-engineers in them.
Use this deck — T-shirtPowers of 2
Every step is exactly double the last, so “twice as big” is unambiguous. Popular for platform and infrastructure work.
Use this deck — Powers of 2Linear 1–10
A flat scale everyone reads instantly. Better for risk, confidence, or priority votes than for size.
Use this deck — Linear 1–10Custom deck
Ideal days, three-point passes, or your own house scale. Needs at least two cards with unique labels, and one numeric card so reveal stats can render.
Build a deck — Custom deck?, ∞, ☕, ½ and T-shirt sizes are non-numeric: they're votes, but they never skew the average, median, mode, or spread.
Bring your own scale
Five decks built in. Or build the one your team argues in.
Pick a deck on the way into the room, or compose a custom one — labels, order, and numeric cards are yours.
How Pointing supports every technique above.
One room, no account, and the mechanics the techniques on this page assume you have.
Hidden votes, one-tap reveal
Nobody anchors anybody. Cards stay face down until the facilitator flips them all at once.
The spread, visualized
Distribution the second you reveal, plus consensus grading — full, near, or split — so you know whether to nudge or discuss.
Stats that don't lie
Average, median, mode, and spread computed from numeric cards only. ?, ∞, ☕ and ½ never skew the math.
Five decks, or build your own
Fibonacci, Modified Fibonacci, T-shirt, Powers of 2, Linear 1–10 — or a custom deck with your own labels.
One item or five hundred
Standalone for a single gut-check, Backlog mode for a whole list with live progress, Team for saved history every sprint.
Your data, exportable
Export results to CSV or JSON and forecast from your own history instead of somebody else's benchmark.
Roles that match the room
Facilitators, voters, and observers — so the people watching don't accidentally become the people estimating.
No account, no setup
Start a room, share the link, point. Anonymous rooms expire on their own; sign in only when you want history.