There is an order to adopting AI, and almost every company breaks it the same way: they start by building something.

Alex Lieberman spent 14 months working on AI rollouts with executive teams and published the progression he sees most: 14 stages, from the initial audit to the jump from deterministic workflows to self-guided agents. The list is useful, but what really matters is the trap it reveals, and that is the thesis of this post: the value, and the failures, live in the first ten stages, exactly the ones everyone skips to reach the flashy part sooner. If your AI plan starts at “let’s build an agent,” you have already skipped the work that decides whether that agent is any good. And that earlier work is exactly what we do.

The 14 stages, in four phases

It is not worth repeating the whole list; it is in Lieberman’s thread, linked above. What is worth doing is seeing the shape, because when you group it into phases the order becomes obvious.

The first three stages are diagnosing: a company-wide audit that maps processes, interviews leadership and surveys employees, a readout that prioritizes by ROI and risk, and the uncomfortable moment when you discover your data is not in order.

The next five are enabling your people: coding agents for engineering, access to an enterprise LLM first for a few champions and then for everyone, workshops for leadership and training across the company.

The two after that are surfacing ideas: an internal hackathon where employees propose solutions, and a leadership decision on which to take from prototype to production.

And only the last four are building and scaling: the first project with clear ROI, cost optimization once the budget starts to balloon, the repeatable production cycle, and finally the move from deterministic workflows to agents that guide themselves.

Fourteen stages. The part most people think of as “doing AI” is the last four. The ten below them are the ones almost nobody does.

Where almost everyone starts

Here is the problem. Almost no company enters at stage 1. It enters at stage 11: building the first project, the one you can show.

It is understandable. Stage 11 is the one you can see, the one that gives you a demo for the board, the one that feels like progress. The ten before it are boring. Understanding your real processes does not produce a nice screenshot. Getting your data in order impresses no one. So they get skipped, and people go straight to what shines.

One of the thread’s commenters, Louis Amira, caught it well: a lot of people do two or three stages and jump straight to 12. And the order is not decorative. Each stage enables the next. You cannot prioritize well (stage 2) without having diagnosed (stage 1). You cannot run a useful hackathon (stage 9) if your people do not know how to use the tools (stages 5 to 8). Skipping the order does not make you faster, it makes you build the third floor without the first.

The stage that kills you is stage 3

If I had to point to a single one, it is the third: realizing your data is not in order. Another commenter on the thread, Ilman Shazhaev, called it the killer, and he is right. If you do not stop to clean and order the data layer, the only thing you automate is producing garbage faster.

And there is a dangerous version of this stage: turning “order the data” into the “company brain” project, that cure-all that becomes an endless effort where the organization hides to avoid building anything. The point is not to perfect your data before touching AI. It is to know which processes have data in good enough order to build on, and which do not. That is a prioritization decision, not an excuse to stall. I argued this in AI is not the problem, your company is not ready: the company that does not describe itself honestly ends up automating its own mess.

What IQ Source does about it

Look at Lieberman’s list again and notice where the first ten stages fall. They are diagnosis, process mapping, data, enablement and prioritization. They are exactly the work most people skip, and they are exactly what AI Maestro is.

That is not a coincidence. We structured AI Maestro as two months of discovery because these are the stages that decide everything else and the ones nobody wants to do alone. We map your real processes, not the ones in the manual. We produce an AI Opportunity Score that prioritizes where to start with a number, not with the hunch of whoever spoke loudest. And we reach a Go/No-Go gate that decides what gets built and what does not. Stage 11, building, comes after, and only if it passes the gate. That is why this is not the same as adoption, which is not transformation: plugging in the tool is the easy part, doing the ten stages below it is the part that changes the result.

If you are about to approve your first AI project today, ask one question before you sign. What stage are you really at? If the honest answer is “we jumped straight to building because we wanted something to show,” you are not at stage 11. You are at stage 1, with a project on top of it that has no foundation yet. Better to know that now than in six months, when the pilot stalls and no one knows why.

Start at the right stage, not the flashy one