The hardest part of an AI project is not the AI

There is a comfortable myth doing the rounds in the industry, the one that says the difficulty of an AI project lives inside the model. Pick the right tool, write the right prompt, and the rest is just delivery.

I believed a gentler version of that myth once too, right up until I was standing inside a four-stage engagement watching a capable organisation discover, in real time, that the technology was never the thing that was going to break.

The thing that breaks is everything around the technology. The expectations. The handoffs. The quiet assumption, held by almost everyone in the room, that an AI output is finished the moment it appears rather than the moment a human has actually checked it.

Managing an AI project, I have come to think, is mostly the work of managing the gap between what people believe AI will do and what it actually does.

A real engagement, not a thought experiment

Last year I ran a complete readiness-to-execution engagement for a client, structured deliberately in four gated stages. We began with a Readiness Assessment (an honest, sometimes uncomfortable audit of where the organisation actually stood, as opposed to where its leadership hoped it stood). From there we built a Roadmap to Readiness, and only then did we move into two distinct phases of execution. Each stage was paid, each stage was gated, and that structure was not an accident of invoicing. It was the whole point.

Here is why the gating mattered. The single most expensive mistake in AI work is committing to a build before you have honestly established whether the foundations can carry it. Everyone wants to skip to execution, because execution feels like progress and assessment feels like delay. It is the project-management equivalent of refusing to read the map because you are already in the car.

The assessment stage is precisely where you find the landmines, the data that does not exist in the form people swear it does, the workflow that nobody actually owns, the approval process that quietly adds three weeks nobody budgeted for.

By the time we reached execution, twice, the genuinely hard decisions had already been made and paid for in the cheap currency of planning rather than the ruinously expensive currency of rework. That is not a coincidence. That is design.

Five things I now refuse to start a project without

What follows is not theory I read somewhere. It is the residue of having done the work, and of having occasionally done it badly enough to learn something.

First, treat the assessment as the deliverable, not the throat-clearing before the deliverable.

In a traditional project you can sometimes muddle through a thin discovery phase and recover later. In an AI project, a thin assessment is how you end up three-quarters of the way through a build that was never viable, holding a beautifully produced asset that solves a problem the client did not have. The honest audit is the work. Everything after it is just consequence.

Second, build the quality-control layer before you build anything else.

Generative tools are extraordinary at producing plausible output at a speed that flatters everyone involved. Plausible is not the same as correct, and correct is not the same as on-brand, and on-brand is not the same as legally clear. Someone has to own the checking, by name, with the time for it written into the plan rather than assumed to happen in the cracks. If your project plan does not have a human standing between the AI and the client, you do not have a project plan. You have a hope.

Third, manage the AI workstream and the human workstream as two different things that have to meet.

They move at different speeds, they fail in different ways, and the coordination between them is where most of the real project risk actually lives. The model does not get tired, but it also does not know when it is confidently wrong. The humans get tired, but they can smell when something is off. A good plan respects both of those truths and schedules around them deliberately.

Fourth, name the AI-specific risks out loud, early, in front of the client.

Not the generic risks every project carries, but the ones peculiar to this kind of work. The output that drifts subtly off-brief across a hundred variations. The hallucinated detail that nobody catches because it reads so smoothly. The model update mid-project that changes the texture of everything you have already approved. Putting these on the table early is not pessimism. It is the thing that lets a client trust you when something does eventually wobble, because you told them it might.

Fifth, protect the margin like the project depends on it, because it does.

AI work seduces people into believing it is cheap, and so the budgets get set with a kind of breezy optimism that the rework phase then quietly demolishes. The discipline of tracking scope, profitability, and change against the plan is not the boring administrative tail of the project. On AI engagements it is frequently the difference between a piece of work that builds a reputation and one that silently bleeds.

The uncomfortable truth underneath all of it
The reason these projects are hard has very little to do with the cleverness of the technology and almost everything to do with the human systems the technology lands inside. Dealerships, agencies, enterprises, the sector barely matters. People do not adopt new ways of working because something is impressive. They adopt them when the friction of the old way finally outweighs the discomfort of the new one, and a project manager's real job is to shepherd an organisation across that uncomfortable threshold without anyone losing their nerve halfway.

That is the part the tooling will never do for you. The model will generate. It will not reassure a nervous stakeholder, or hold a timeline together when two workstreams collide, or tell a client a truth they would rather not hear in the assessment phase so that they are not blindsided by it in execution. That work is stubbornly, irreducibly human.

Which is, when I think about it, rather a hopeful thing. The better these tools get, the more valuable the people who can actually manage what the tools produce become.

The orchestration is the craft now. The judgement is the moat.

And no model has yet worked out how to do the genuinely hard part for us, which is to look another human being in the eye and say, with the quiet confidence that only comes from having done the work, here is where we actually stand, and here is how we get from here to where you want to be.

Structure before AI, always.