Your projects are unique. Your schedule data isn't.

The finding: uniqueness lives at the project level, not the activity level

I was a planner on defense programs. No two were alike. Yet one level down in the work breakdown structure, the same families of work kept showing up: a design review, a qualification test, an integration campaign, a long‑lead component purchase. Their durations varied, but not at random; they varied according to factors that experience made it possible to anticipate.

So the information was there. It was simply unreadable for a machine, and often for any human other than its author.

Why it happens

Take ten schedules from the same large group and try to compare them. You will find the same activity named six different ways, missing logic links the planner made up for in their head, imposed date constraints used to force a timeline, and durations padded "to be on the safe side." Each schedule makes sense to the person who built it. Side by side, they don't speak to each other.

In 2023, the SMaP (French association for the advancement of project management, IPMA France) devoted a white paper to this problem. Its conclusion matches ours: current practices for structuring project data do not allow reliable analysis, including by AI, and the working group proposes revisiting several classic concepts so that information becomes formal and connected. It targets schedule development one to two orders of magnitude faster, and the ability to compare versions and variants of the same project.

In other words: before talking about algorithms, you have to talk about data hygiene.

What changes: a two‑step method

Step one: standardize. You build a shared reference library of activities, deliverables and resources, then map your existing schedules to it. You run the historical schedules through a quality check: missing links, outlier durations, abusive constraints, negative float. Each anomaly is either corrected or documented. Lessons learned stop being a document you reread before a project; they become a database you query.

Step two: generate. On top of this cleaned‑up foundation, models can propose durations calibrated on your history, recognize recurring sequences, and generate an initial schedule or several scenarios. Skip step one and you get spectacular results in a demo and disappointing ones in production.

This sequencing has a practical consequence that AI leadership teams underestimate: the first deliverable of an AI planning project is a clean dataset. It has value in its own right, before the first model is ever built.

In the field

In the space industry, deep learning models trained on the company's own data reached a measured accuracy of more than 95%, applied to schedule defect correction and generative scheduling. That figure was only possible after in‑depth work on the quality of the schedule data.

Frequently asked questions

Further reading

• –Use case: Schedule defect detection and correction
• –Module: AI LAB
• –Glossary: Quality check, Lessons learned

Sources

• –S. Boutin, G. Turré, Structuring project data — Planning digitalization, white paper, SDP/PDS working group, Société française de management de projet (SMaP), version 18, September 2023.