Project Estimate Reality Check

Anchor your estimate in how similar projects actually went, ask for a checkable reason before you accept "ours is different", and imagine the project has failed while changing course is still cheap. About 30 minutes finding the numbers, 15 minutes comparing, and a 45-minute premortem with the team.

Research, with limits: reference-class thinking and the uniqueness finding come from Flyvbjerg's project research (Flyvbjerg & Gardner, 2023; Flyvbjerg et al., 2024, working paper): large observational datasets, largely one research group. The premortem is practitioner craft with a named origin (Klein, 2007). Every threshold on this page is a rule of thumb.

Use it when:

• You are about to commit a budget or a date upward, to a sponsor, a board, a client or a bank.
• A plan was built from a work breakdown and nobody has compared it with how similar work went.
• A project is being re-planned after slipping.

Skip it when:

• You need to define the project and its scope first: use 8.5 Project Charter Template.
• You want to find out what went wrong on a finished project: use 8.20 End-of-Project Review.

Your project is one draw from a class of similar projects, and the class's track record is a better first forecast than your plan. A plan is built from intended steps and quietly assumes each goes roughly as intended; the class already contains every surprise such projects meet. A plan is also a persuasion document, which pulls estimates toward what will be approved rather than what will happen. Optimism in estimation is systematic, not occasional, so the correction has to be structural rather than a vow to be realistic (Flyvbjerg & Gardner, 2023).

Hold the evidence with its limits: the project databases are observational and largely the work of one research group, the uniqueness result is correlational and still a working paper (Flyvbjerg et al., 2024), and the premortem (Klein, 2007) is practitioner craft with no effect study of its own in our register. The outside number comes first so it exists before anyone defends the inside one. The premortem then attacks what survives from a different angle: how this project in particular could fail.

Estimate from the outside view (what happened to projects like this) before trusting the inside view (what we plan to do). In a working paper's data on IT projects, projects whose leaders rated them as more unique had larger cost overruns (not schedule overruns), and the authors' own comparison found earlier similar projects even for the highest-rated cases (Flyvbjerg et al., 2024, working paper). Almost every project is "the first time" only for its own team.

Strong answer: (illustrative) "Cold-store extension 2022: planned 5 months, actual 8 (60 percent over); driven by foundation surprises and a late permit." Not: "Similar projects usually go a bit over."

If you cannot find three comparables: loosen the match (the same type of work elsewhere in the organization, industry experience, a colleague's project) rather than concluding the project is incomparable. If you still cannot, break the project into components: the whole may feel unique, but every part (a system migration, a training rollout, a construction phase) has a reference class.

Illustrative example: the organization and numbers are made up.

A Danish dairy cooperative plans a new packaging line: 6 months, 2.4 MDKK. The reference class (the line installed two years ago, a sister plant's cheese line, a supplier reference) shows overruns of 40 to 90 percent in time, driven by foundations, commissioning and food-safety validation every time. The plan assumed the team would beat the class because "the supplier promised a standard installation": not checkable.

Adjusted estimate: 9 months, 3.2 MDKK, with the buffer held by the sponsor. The premortem adds two failure reasons the plan missed (the one controls engineer is also needed on another site; nobody owns the validation), which move staffing and name an owner before launch, and puts one named risk in the log. The sponsor chooses to fund the realistic version. The line is running at 8.5 months, which the board experiences as on time, because the estimate was honest.