Estimating

Faster tools should have made estimating easier. They made it harder, for four reasons that have nothing to do with model quality.

The reference class broke. Almost all practical estimating is comparison with your own past. Two years of changing workflow means that past is no longer a guide, and the single most reliable technique in software estimating quietly stopped working. When this estimation problem needs an operational counterpart, workforce optimization software offers a useful reference point.

The variance widened even where the average did not. Some tasks collapsed to a fraction of what they took; some take longer. An estimate is a claim about a distribution, and a mean drawn from a bimodal one describes nothing that will actually happen.

Perception is no longer calibrated, which matters because estimating has always leaned on a felt sense built from experience. That sense is now known to be optimistically biased, and knowing about it does not correct it.

And the work changes shape mid-task, between generation-heavy and verification-heavy, often without warning.

What survives is a method rather than a technique: classify the task before estimating it, size producing and verifying as separate numbers, give ranges proportional to the uncertainty, and name in advance what would change the figure. Then record actuals against estimates until the ratios are yours instead of anyone's average. Atlassian publishes a practical overview of agile project-management approaches See Atlassian Agile.

One thing worth saying at the outset, because it reframes the whole section: estimates were never mostly wrong because writing was slow. They were wrong because of things nobody knew when the estimate was made — misstated requirements, systems that do not behave as documented, data that is not what it was described as. Generation speed touches none of those, so their share of the total error grew while they themselves stayed the same size.

• Estimating Review, Not Writing

  The most common way assisted work runs over is estimating the half that got cheap. How to size the half that did not.

• Estimating Assisted Work

  A method that survives unstable speed: classify the task, estimate the two halves separately, and give a range you can defend.

• When the Client Wants One Number

  Ranges are honest and clients still want a figure they can plan with. How to give one without carrying risk you cannot size.

• Padding and Buffers, Honestly

  Everyone pads and nobody admits it. Why hidden padding fails now, and how to carry uncertainty where the client can see it.

• Re-estimating Mid-Project

  Bad news early is information. Bad news late is a failure. The difference is almost entirely in when and how you say it.

• Reference-Class Estimating

  The most reliable estimating method is comparison with your own past. Here is how to rebuild that past after two years of churn.

• Unknowns That Remain

  Estimates were never wrong because writing was slow. Six sources of uncertainty that generation speed does not touch at all.

• Why Estimates Got Worse

  Faster tools should have made estimating easier. They made it harder, for four reasons that have nothing to do with the models.