Managing AI Projects: What Traditional Project Managers Need to Unlearn

The instinct that makes someone a strong traditional project manager ; define scope precisely, lock it down, deliver against it. Is the exact instinct that causes AI projects to stall. I learned this the expensive way, running an AI-enabled reporting initiative like a standard software delivery project and watching the plan fall apart within the first month.

Traditional projects assume that once requirements are gathered, the solution space is mostly known. AI projects don't work that way. You don't know how well a model will perform on your actual data until you've tried it, and 'try it' isn't a two-week spike, it can be several iterations of data cleaning, retraining, and evaluation before you know whether the approach is even viable. Committing to a fixed scope and date before that exploration happens sets the entire project up to either miss its date or quietly reduce its ambition to fit the deadline.

The first thing to unlearn is treating model performance as a deliverable you can schedule with confidence. It's closer to a discovery outcome. I now scope AI initiatives in two distinct phases with two different planning approaches: a time-boxed feasibility phase with a go/no-go decision point, and only after that, a delivery phase planned the way you'd plan traditional software work.

The second thing to unlearn is who owns quality. In traditional delivery, QA has fairly objective pass/fail criteria. With AI systems, 'good enough' is a judgment call that has to involve the business owner directly, an 85% accurate model might be a huge win in one context and completely unacceptable in another. That threshold conversation needs to happen explicitly, in writing, before development starts, not after the first evaluation report lands on someone's desk.

The third: data readiness is a workstream, not a line item. I now budget as much planning attention to data access, quality, and governance as I do to the model or feature work itself, because in practice it's usually the actual bottleneck.

None of this means abandoning project management discipline, if anything, AI projects need more structure around decision points, not less. It means moving the structure to where the real uncertainty lives, instead of forcing certainty onto a phase of work that doesn't have it yet.