Estimation and Scope — The Cone of Uncertainty and How to Negotiate

Estimates miss because a question about a range gets answered with a point. The cone of uncertainty, three-point estimates, why buffers don't add up the way you think, reference class forecasting, and how to turn a deadline conversation into a scope conversation.

Your GPS Lies to the Minute

Pull onto the highway and the navigation system tells you: arrival 5:38 PM. To the minute. In reality, traffic or roadwork can push that by an hour. We believe the number anyway — and if the device announced "you'll arrive sometime between 5 and 7," we would assume it was broken.

People dislike being given ranges and press for points. Half of what goes wrong with software estimates starts there. The honest answer is a range, the question demands a point, a point gets given, and that point walks out of the room as a promise.

There is a second, deeper misconception: that a blown estimate means someone was lazy or unskilled. The dominant cause is simply things you don't know yet, and the volume of unknowns doesn't shrink through willpower. It shrinks by investigating, building, and touching the thing. Which gives us the single most important property of estimation: accuracy is a function of action, not effort.

The Cone of Uncertainty — Precision Isn't Available on Day One

The picture of that property is the "cone of uncertainty." The underlying figure comes from Barry Boehm's 1981 book; Steve McConnell gave it the name that stuck. Put project progress on the horizontal axis and "estimate as a multiple of the actual" on the vertical, and you get a funnel: widest at the start, narrowing as the work proceeds. At the concept stage the range is usually drawn as 0.25× to 4× of the eventual actual — a factor of 16 between best and worst case.

That sounds theatrical until you recognize it. The feature you called "oh, about two weeks" sometimes finishes in two days and sometimes takes two months. Same mouth, same sentence.

Here is the reading you must not get wrong. The cone does not promise to narrow on its own. McConnell is explicit that it is a best-case envelope: unless you take actions that actually remove uncertainty, the width stays where it is. This is the most common accident in the field — "our numbers will be sharper next month," said while nothing about the spec or the technical risk gets settled, and a month goes by. The range holds steady while the deadline closes in. What narrows a cone is not time. It is deciding things and trying things.

Concretely there are only three moves: close the ambiguous parts of the spec one at a time, run a small experiment against the unfamiliar technology, and get one thin path working end to end. The last one earns its keep. Push a single skinny slice from the screen through to storage and all the work nobody listed shows up at once.

Answer with a Range — Three-Point Estimates

So what do you actually say? The smallest useful tool is the three-point estimate: instead of one number, give three.

• Optimistic $O$: everything you're aware of goes well
• Most likely $M$: the ordinary case, the value you'd bet on
• Pessimistic $P$: known risks materialize (not meteors — the bad things you can already name)

With three numbers you can compute a representative value and a spread. The classic formula goes by the name PERT.

$E$ is the expected value and $\sigma$ (sigma) is the standard deviation, a measure of spread. In words: $E$ is "an average that counts the most likely value four times over," and $\sigma$ is "one sixth of the gap between optimistic and pessimistic." The six comes from the rule of thumb that nearly everything falls within three sigma either side of the mean.

For example, $O=3$ days, $M=5$ days, $P=15$ days gives $E \approx 6.3$ days and $\sigma=2$ days. If you had committed to the most likely value of five days, you were going to break that promise fairly often.

The formula is a crude approximation to a beta distribution and is not precise. Its real value lies less in the arithmetic than in forcing three numbers to be said out loud. The moment you ask "what does the bad case look like?", the risk list that was sitting silently in someone's head becomes words. And if the pessimistic value is three times the most likely one, you have just learned that the task needs investigating before it needs estimating.