Why enterprise delivery estimates survive — or do not

Ask an engineering team for an estimate and you will usually get a number derived from how long similar work took before. That is a reasonable heuristic, and it is why so many enterprise projects slip: the previous work happened in a context that no longer exists.

Estimate the constraint, not the feature list

Delivery duration is governed by the slowest thing in the system, and in enterprise environments that is rarely typing speed. It is a security review, a data migration window, a partner’s change freeze, or a single person who must approve a schema change. Find that constraint in week one, write it down with a name and a date, and build the plan around it.

Five practices that keep estimates honest

• Size to the first production slice. Commit to the smallest end-to-end path that reaches real users, then re-estimate with evidence.
• Budget discovery separately. Discovery is a deliverable with a fixed duration, not overhead absorbed by the build.
• Name the unknowns. Anything unverified in writing is a risk line with an owner and a date to resolve it.
• Track change cost. Lead time and change failure rate are better predictors of the remaining schedule than percentage complete.
• Publish the burn. A weekly view of capacity versus consumed effort ends the “are we nearly there” conversation.

Why “two more weeks” is a systemic failure

When a team reports two more weeks indefinitely, the estimate is not wrong — the reporting is. Nobody is measuring what remains, so optimism fills the gap. Force the number: how many increments remain, what is in each, and what would have to be true for the date to hold. If that cannot be answered, the work is not estimated yet, and saying so is more useful than a date the team does not believe.

The commercial side matters too

Estimates are only credible if the engagement model supports iteration. Capacity-based arrangements with a quarterly review survive discovery findings; fixed-price contracts signed before architecture exists convert every learning into a change request. Choose the commercial shape that matches how much you actually know.