Not every AI project we have run has gone to plan. I would rather say that out loud than pretend otherwise, because the ones that stalled taught us more than the ones that sailed through.

If you have been told to "do something with AI", this is the honest version of what slows these projects down. It is also what we would do differently with the benefit of hindsight.

Three reasons projects stall

When an AI project loses momentum, the cause is rarely the AI. In our experience it comes back to three things.

Scattered data. The information the project needs lives in five places. A spreadsheet here, an old system there, a folder of PDFs that only one person understands. Before the AI can do anything useful, someone has to pull it together, and that work is almost always bigger than anyone planned for.

Unclear ownership. Nobody has decided who signs off on what the AI can see, what it is allowed to do, and who answers for its output. So every decision waits for a meeting, and the project slowly drifts.

Starting too big. This is the one we would change first. The temptation is to aim at the whole business at once, because that is where the big savings look like they sit. The result is a project with too many moving parts, too many stakeholders and no early win to keep people believing in it.

What we would change

We would start with one slice of work and prove the value there before going anywhere else.

One process. One team. One kind of task that people already find tedious and repetitive. Something small enough that the data problem is manageable and the ownership question has an obvious answer.

That narrow start does three useful things:

  • It forces you to clean up only the data that one task needs, so the scattered data problem shrinks to a size you can actually handle
  • It makes ownership simple, because one team owns the process and can make the calls
  • It gives you a result you can show people within weeks, which buys you the trust and budget for the next step

The lesson

Proving value small beats promising value big. A working result on one process is far more persuasive to a board, a business partner or your own team than a plan that covers everything.

It is also less risky. If the first slice does not deliver, you have learnt something cheaply and you can pick a different one. If it does deliver, you have a pattern you can repeat.

How it changed our approach

These days the first conversation we have about any AI project is about scope. Where is the busy work that people would happily hand off? Which one piece of it has data we can get at and a clear owner? That becomes the starting point.

Everything else waits until that first piece is working. It feels slower at the start. In practice it gets you to real results sooner, because you are not stuck untangling the whole organisation before anything useful happens.

If you are about to kick off your first AI project, resist the urge to go wide. Pick one thing, get it working, then grow from there.

Try this. List three repetitive tasks your team does every week, then circle the one with the tidiest data and the clearest owner. That is your first AI project.