What is a three-point estimate?

Each of the three values answers a different question about the same task, and naming them carefully is what makes the estimate honest. Pull them from history, from the task owner, or from a reference class of similar past work — not from optimism.

PERT collapses the three points into a single expected duration using a weighted average that leans heavily on the most likely value: TE = (O + 4M + P) / 6. The most likely estimate gets four times the weight of either extreme, so the result tracks reality while still being pulled by a long pessimistic tail.

A worked example: with O = 7, M = 10, and P = 20 days, the PERT mean is (7 + 40 + 20) / 6 = 11.2 days — noticeably longer than the 10-day "most likely" figure, because the downside risk is larger than the upside. That gap is exactly the information a single-point estimate throws away.

PERT also gives a quick uncertainty measure: the standard deviation is approximately (P − O) / 6, so the example task has a spread of about 2.2 days. A wide O-to-P range flags a task that deserves attention, a buffer, or more investigation before you commit to a date.

The weighted mean is a shortcut for a richer idea: the three points describe a Beta distribution — specifically the PERT-Beta (also called the Beta-PERT) — bounded by O and P with its peak at M. Unlike a symmetric bell curve, the Beta distribution can be skewed, which matches how real tasks behave: they have a hard floor on how fast they can go, but a long tail of ways to run late.

This is why three-point estimating pairs so naturally with simulation. A single PERT mean gives you one number; the full PERT-Beta distribution gives you a shape you can sample from thousands of times to see the whole range of outcomes, not just the average. The classic (O + 4M + P) / 6 formula is just the mean of a Beta distribution under PERT's standard shape assumptions.

A deterministic Gantt asks every owner for one number per task, then sums them along the critical path to produce a finish date. The problem is that those single numbers are almost never the median — they are hopeful most-likely values — and uncertainty does not add up symmetrically. Across a chain of tasks, the late tails compound while the early finishes rarely get passed downstream, so the deterministic date is systematically optimistic.

The result is the board-deck date that no one actually computed and that slips every quarter. A single-point estimate cannot tell you the probability of hitting a date, cannot show which task is driving the risk, and cannot price a buffer. Three-point estimates fix the input; Monte Carlo turns that input into the probabilities you can defend.

In CritPath AI, three-point estimates are the native unit of a task. You enter optimistic, most likely, and pessimistic durations, and the engine builds a PERT-Beta distribution for each task as the direct input to its Monte Carlo simulation — with optional duration correlation so related tasks slip together the way they do in real programs.

From those distributions the simulation produces P50, P80, and P90 finish dates, a criticality index showing how often each task lands on the critical path, and a tornado chart ranking which tasks drive the most schedule variance. The same three points feed CCPM buffer sizing and decision-gate analysis, and a Claude- and Gemini-powered copilot — grounded in your actual dependency graph — can suggest realistic ranges and explain which estimate is pushing your P80.

It runs in a modern web app at $10 per user per month, with AI usage billed separately by metered usage — quantitative, three-point schedule risk without a legacy desktop seat.