Three-point estimates, and what they prove
Weighting optimistic, likely and pessimistic gives a better number than asking for one date. It also produces confidence figures that are easy to over-trust.
Why the most likely case is not the estimate
Ask how long something takes and you get the version where nothing goes wrong. That is a real answer to a different question. Work has a floor and no ceiling: it can rarely finish much faster than the good case, and it can overrun almost without limit.
PERT captures that asymmetry: (o + 4m + p) ÷ 6. A task most likely to take
four days, possibly two, possibly twelve, has an expected duration of five — above the most
likely value, because the long tail has to go somewhere. Plan twenty tasks to their modes and the
project is late without any single estimate being obviously wrong.
The mistake that doubles your uncertainty
Along a path, variances add. Standard deviations do not.
Four activities each with a standard deviation of 1.67 days: adding directly gives 6.67 days of
spread. The correct figure is √(4 × 1.67²) = 3.33 — exactly half.
One number justifies a sensible contingency; the other justifies a panic. Uncertainty grows with the
square root of the number of activities, not in proportion to it.
What an 80% figure is actually worth
It is conditional on four things: that the estimates are honest, that the activities are independent, that the path stays critical, and that the optimistic-to-pessimistic range really spans about six standard deviations.
Independence is usually the weakest. Real delays are correlated — the same missing specialist, the same late supplier, the same optimistic culture affects several activities together. Correlated delays make the true spread wider than the arithmetic shows. So the honest reading of an 80% figure is "reasonably likely if things behave", not "four times in five".
Where PERT stops
It assumes the path you gave stays critical. In a real network a near-critical path can overtake the critical one once durations vary — and a path two days shorter but far more uncertain will govern the finish more often than its length suggests. Answering that needs a simulation across the whole network, not arithmetic along one chain.
That is a limitation of the method, not a missing feature, and it is worth knowing which questions the technique can actually answer.
What this tool will not do
It calculates from the figures you enter and nothing else. It cannot tell you whether your estimates are honest, whether your dependencies reflect how the work really flows, or whether the plan is achievable with the people you have. Those are judgements, and a schedule that looks precise can still be built on all three being wrong.