When is a back-of-envelope estimate good enough to decide on, and when do you insist on measurement?
answer
- precision is not the goal
- write the decision threshold first
- build a low and a high
- does the band straddle the line
- buy data only on the deciding assumption
basics
~10 sAn estimate is good enough when the decision does not flip anywhere across its plausible range. Write down the threshold that would change the call first, then check whether the low-to-high band straddles it.
solid answer
~40 sI invert the usual order and start from the decision, not the estimate. First I write down the threshold that would change the call: the volume at which a new site pays for itself, the spend above which this gets its own team. Then I build the low and high versions of the estimate from the honest range on each assumption. If the whole band sits on one side of the threshold, the estimate is decision-grade and further work is waste, however imprecise it looks. If the band straddles the threshold, I identify which single assumption moves it across and commission work on exactly that one. I also weigh reversibility: a cheap, reversible call can ride on an order of magnitude, while an expensive one-way commitment deserves measurement even when the envelope looks comfortable.
go deeper
Be ready to present an estimate as a range with its assumptions listed rather than a single confident number, and to ask what decision the number is for before deciding how hard to work on it.
An interviewer at this level expects you to state the threshold that would change the call and check whether your low and high sit on the same side of it, instead of chasing precision for its own sake.
Demonstrate that you stop work when the band cannot flip the decision, and that when it does straddle you scope information-gathering to the single assumption responsible rather than re-doing the whole estimate.
Own the norms around how estimates circulate: thresholds written before numbers, ranges and assumption sheets travelling with every figure, reversibility and downside cost weighed alongside the band, and estimates never allowed to become targets.
## Precision is not the goal; a stable decision is The usual framing of *is this estimate good enough* is unanswerable, because good enough for what is unstated. Precision has no intrinsic value. The question becomes tractable the moment you attach it to a decision, and then it has a clean criterion: **An estimate is decision-grade when the decision is the same at every value in its plausible range.** That reframes the work. You are not trying to narrow the estimate; you are trying to determine whether the estimate is already on one side of a line. ## The procedure **1. Write the threshold first.** Before estimating anything, state the number that changes the call. If we would open a second facility above 400,000 meals a year, 400,000 is the threshold. Writing it first prevents the very human failure of building an estimate and then choosing a threshold that agrees with it. **2. Build a low and a high, not a point.** Take each assumption's defensible low and high, and construct the pessimistic and optimistic versions of the chain. This is a bracketing exercise driven by assumption ranges, not a statistical statement about sampling. **3. Compare the band to the threshold.** - **Band entirely on one side.** Decide now. An estimate spanning 6 to 20 million dollars is extremely imprecise and completely sufficient if the threshold is 2 million. Further refinement is pure cost. - **Band straddles the threshold.** The estimate is not decision-grade *for this decision*. Do not average and proceed; find the assumption whose movement carries the answer across the line, and buy information about that one thing. - **Band far from any threshold in either direction.** The decision was probably never close, and the estimate's main value was to reveal that. **4. Weigh the cost of being wrong, not just the odds.** Two decisions with identical estimates deserve different treatment. A reversible choice with cheap failure can ride on an order of magnitude. A one-way, expensive commitment — a long lease, a hire, a public commitment — warrants measurement even when the band looks safe, because the loss function is asymmetric. **5. Compare the cost of the information to the size of the swing.** If the decision turns on a factor and you can resolve it for a day of work while the decision is worth millions, buy it. If resolving it costs three months and the decision must be made this week, the envelope plus a hedge is the real option, and saying so plainly is better leadership than stalling for data that will arrive too late. ## Communicating an envelope estimate How the number leaves your hands matters as much as how it was built. - **Quote a range, never a point.** *Roughly 8 million, plausibly 5 to 12* survives being repeated. *7.56 million* invites false precision and will be quoted back as a fact. - **Publish the assumption sheet.** The factors and the values you chose should travel with the number so anyone can substitute their own and re-derive it. An estimate whose inputs are hidden cannot be challenged and therefore cannot be improved. - **Name the dominant assumption in the same breath.** Say what the number hangs on, so a reader who knows that factor better than you can correct you immediately. - **Stamp it as an estimate.** Numbers migrate into slides and plans and lose their provenance in two hops. Labelling it prevents an envelope figure from hardening into a commitment nobody remembers making up. ## Failure modes at this level **Refining what cannot matter.** Teams spend weeks tightening an estimate that was already unambiguously on one side of every threshold. This is comfortable work and produces nothing. **Deciding on a straddling band.** The opposite error: the range clearly spans the threshold and the meeting proceeds on the midpoint. The midpoint is not more likely than the endpoints in any meaningful sense here, and treating it as the answer discards exactly the information the exercise produced. **Anchoring on the first number spoken.** Once an envelope figure is said out loud, later evidence gets pulled toward it. Fixing the threshold before estimating is one defence; having a second person build an independent chain is another. **Letting the estimate outlive its inputs.** A number built on assumptions that were true last year gets reused as though it were measured. Estimates should carry an expiry condition: which input, if it changed, would require rebuilding this. **Using an estimate as a target.** The most damaging outcome is an envelope figure becoming a commitment. It was a description of what you believe, not a promise about what will happen. ## The judgment being tested An interviewer asking this wants to see that you treat estimation as a tool for allocating attention. The strong answer is not *always measure* or *ship it, it is close enough*. It is: here is the threshold, here is the band, here is whether they overlap, and here is the one thing I would go find out if they do.
- Your range is wide but sits entirely above the threshold. Do you refine it?No. The decision is identical at every value in the range, so refinement cannot change the call and is pure cost. I would say exactly that in the room, record the threshold and the band so the reasoning is auditable, and redirect the effort to a decision that is actually close. Precision that cannot flip an outcome is a comfortable way to look busy.
- How do you stop an envelope figure from hardening into a commitment?Attach the range, the assumption sheet and an explicit estimate label to the number itself, so all three travel when it is copied into a plan. Name the dominant assumption alongside it, and state the condition under which it must be rebuilt. Numbers lose their provenance within a couple of hops, and a bare point figure with no visible inputs is indistinguishable from a measurement.
- The band straddles the threshold and there is no time to gather data. What then?Say plainly that the estimate does not resolve the decision, then choose on other grounds: reversibility, downside cost, and whether a smaller committing step buys time. Pick the option that fails cheaply if the pessimistic end is true. Presenting a midpoint as though it settled the question is the failure to avoid, because it discards exactly the finding the estimate produced.
saying these in an interview costs you the question
- Refines an estimate that is already on one side of every threshold
- Decides on the midpoint of a band that straddles the threshold
- Chooses the threshold after seeing the estimate
- Circulates a point figure with no range or assumptions
- Ignores that the decision is irreversible and expensive
- Lets an envelope figure become a delivery target