What AI agents really change in a planning project.
AI agents write models, analyse variances and simulate scenarios. After a year of using them on real projects, here is what they shift, and what they do not.
The build is no longer the bottleneck
For fifteen years, the duration of a planning project was measured in build days. Adding a dimension, creating a module, recalculating a level: every business request translated into effort, and effort into delay. Agents break that equation. A model described in plain language is generated in minutes, and iteration happens during the workshop rather than between two workshops.
The consequence is not the one usually advertised. Projects do not become half as expensive. The bottleneck simply moves upstream: knowing what to build, and on which data.
Framing now carries more weight
A badly framed model is generated as fast as a good one. That is the most uncomfortable change agents bring: the error now happens at the same speed as the success, and it surfaces later, when users find that the figures do not match their reality.
Turning a vague business need into explicit rules has never been machine work. Deciding whether replenishment favours service level or inventory value remains a leadership call. Reconciling two definitions of margin between finance and merchandising takes a meeting, not a model.
Data becomes the real prerequisite
An agent placed on approximate master data produces a wrong answer, faster and with more confidence than a spreadsheet. On the projects we run, the share devoted to data quality and governance has grown, even as the share devoted to building has shrunk.
That is also what makes connecting external assistants delicate. Opening plain-language access means having decided beforehand what is exposed, to whom, and with what traceability.
What changes for the teams
- Testing changes in nature: you check reasoning, not only formulas
- Adoption accelerates, because a user can query the model without prior training
- Documentation stops being the poor relation, since the agent produces its own
- The product owner role gains value rather than losing it
Our full take on the subject is set out on the AI and planning page.