Phase 0: Diagnose

Phase 0 builds nothing. No agents, no pipelines, no evals. That's why most companies skip it, and it's why most of them are still arguing about their AI strategy a year later.

This phase is where you stop telling yourself a flattering story. You place every function on the ladder, honestly. You write down every reason the company has given for why this can't happen here, and you put each one on trial. You find out whether the deadline you were handed can hold. Then you say all of it out loud, to the board and to the people doing the work, before anyone has had a chance to fill the silence with something worse.

It's the cheapest phase in the whole transition. It's also the one that decides whether the rest of it is honest.

What happens

Intake, place every function on the ladder, open the Constraint Court docket. The honest timeline goes to the CEO and board now, not in the last month.

The gate

Deadline verdict accepted, an owner named for each function

What changes for the people doing the work

Nobody's job changes yet. Leaders say out loud what's coming and what hasn't been decided.

What happens

Three pieces of work, in this order. None of them can be delegated to a consultant, because the point of each is that you see it for yourself.

Intake

Start at intake. Headcount, the functions in scope, the deadline and who set it. The last one matters more than it looks. A deadline set by the board is a different negotiation from one set by the CEO at an offsite, and a different one again from a date you picked yourself to sound ambitious. You'll be going back to whoever set it before this phase is over, so write their name down now.

Then edit the core workflow list for each function. The starter list is a draft. Your finance team may not do collections. Your legal function may be one person and an outside firm. Fix the list so it describes your company, but be careful what you remove. Every workflow you take off is one you're claiming doesn't need to reach rung 5, and that claim gets a hearing like any other.

Place every function honestly

Now place each function on the ladder, one core workflow at a time. The placement questions are yes or no. There's no "partly", on purpose. "Partly" is where a rung 2 workflow goes to call itself rung 3. A workflow sits on the first rung it hasn't fully cleared, and a function sits at the rung of its weakest workflow.

Don't take the answers from a survey. People are bad at judging their own speed with these tools, and the evidence is embarrassing. Self-reported progress is a feeling. Go and look.

Looking means sitting next to the people doing the work for an hour and saying almost nothing. Watch where they stop trusting the output and start checking by hand. Watch what they paste, and where from. The one-week diagnosis covers engineering in detail, and the same hour works in support or finance. You'll learn more from watching one person reconcile an invoice than from any adoption dashboard.

Expect the reading to hurt. Engineering at 3, support at 2, finance at 1 is a typical first reading, and plenty of companies come in lower. It's a starting line, and it's only useful if nobody has shaded it.

Open the docket

Every company arrives with reasons. The codebase is too old. Security won't sign off. The seniors won't use it. We need to pay down tech debt first. Let's wait for the models to get better.

Write every one of them down. That list is the Constraint Court docket, and in Phase 0 your job is to open it, not to win it. Each constraint will get a verdict: real, negotiable or false. A real one gets designed around. A negotiable one gets a price and an owner. A false one gets a date after which nobody is allowed to say it in a meeting.

The first question at every hearing is the one that clears half the docket on its own: who owns this constraint? Not a team. A name, and where it's written down. "Security won't approve it" is a feeling until someone can tell you which security reviewer said no, to which proposal, on what date. If nobody can name the owner, the constraint is treated as false until someone proves otherwise.

A constraint nobody owns is a rumour with a veto.

State them at the Court as you hear them. Don't filter. The ones you're most tempted to leave off are usually the ones the room believes most.

The gate

Phase 0 has a gate, and a function can't cross into Phase 1 without it. Two conditions. Both are simple to state and uncomfortable to meet.

The deadline verdict is accepted

Run the timeline check against your own inputs. It comes back fits, tight or won't hold, with what to cut or re-sequence. The guide publishes rules, never durations, so the verdict is built from your headcount, your rungs and your deadline, not from somebody else's case study.

"Accepted" is the hard word. It doesn't mean you read the verdict. It means you and whoever set the deadline agree on what the deadline can hold. If the verdict is "won't hold", accepting it means one of two conversations has happened: the date moved, or the scope did. Push back on scope rather than on the date where you can. A date is usually a promise somebody else already made upstairs. Scope is where you have room to negotiate.

What you can't do is sit on a "won't hold" and hope. Hope doesn't pass the gate.

An owner is named for each function

Every function in scope gets an owner. One person, by name. That person will write its intent, answer for its rung, and handle what escalates when the intent runs out.

"The support leadership team" is not an owner. Neither is "a working group." Ownership is the atomic unit, and the test is the same one you'd use for a service: at two in the morning, when this breaks, whose phone rings? If the answer is a discussion, you don't have an owner. You have a committee, and committees are where AI programmes go to wait.

A name, not a team. If you can't name the owner, the function isn't ready.

Two more rules. Functions that move side by side need separate owners, because one leader can't write intent for two functions in the same stretch. And the owner should be someone who knows the work, not someone new you plan to hire for it. The people who know your systems and your customers are the ones who can write the intent. A new "Head of AI" can't.

If a function has no one willing, that's information. Leave it out of scope for now and say so. Don't fill the slot with a name that hasn't agreed.

Tell the board the honest timeline now

Whatever the verdict says, the CEO and the board hear it now. Not in the last month of the deadline, when it's a confession. Now, when it's a plan.

The timeline chapter walks through the worked example: 200 engineers, SOX, a board mandate of "AI-native in 90 days", and a verdict of won't hold. The lesson in it is about timing as much as scope. Say it to the board in week one, not week eleven.

The same news lands very differently depending on when it arrives. A board can absorb almost any bad news that comes early with a plan attached. What it can't absorb is finding out you knew and didn't say. In week one, "this won't hold, here's what will" is leadership. In week eleven, it's a failure you hid for ten weeks.

Put it in writing, in the monthly note, in plain words. The verdict. What the deadline can hold. What you're cutting or moving, and why. The asks: the owners you need, the security review you need scheduled, the budget line for evals. A board that sees the whole reading, low rungs included, can help you. A board that sees a green dashboard can only be surprised later.

This is also where most CTOs flinch, because the honest reading makes the company look further behind than last quarter's story did. It was always that far behind. The only thing you're choosing is whether the board learns it from you, early, or from the calendar.

What changes for the people doing the work

In Phase 0, nobody's job changes. Nobody is reorganised, and no role is redesigned. That's true, and you should say it.

But saying only that is a mistake. People can read. They've seen the board mandate, or they've guessed it. They know the company just started scoring their function on a ladder. If all they hear is "nothing is changing," they'll hear the part you didn't say, and they'll fill it in with the worst explanation on offer, which is their jobs.

So leaders say two things out loud, in the same breath.

What's coming. The company is moving the hands-on work to machines, one function at a time, starting with engineering. The human seat moves from doing the work to writing intent, making the calls and handling escalations. Roles will be redesigned around judgment and escalation, function by function, later. The headcount conversation will happen then, in the open, for each function. Say when that is in the sequence, not on the calendar.

What hasn't been decided. Which roles change and how. Whether any function ends up with fewer people. What the new standing roles look like. Say plainly that these are open, and that you won't pretend otherwise. You can't promise what you don't control, and a promise you break in a year will cost you every true thing you say after it.

Then say what you do control. What the numbers you're collecting are for (deciding what to build next, not anyone's review). That nobody's standing depends on the tools looking good. That people will hear about changes to their team from you, early, the same way the board does. And tell them the pressure behind the timeline, not just the timeline. People who know what you're working against can reason about the plan. People handed only the outcome can only receive it.

Write it down. A short note, the same one everywhere, repeated unchanged every time it comes up. A message delivered out loud gets retold, and every retelling drifts toward whatever the listener is afraid of.

And use the tools yourself, in the open, while you say it. Nothing you announce carries as much weight as the leadership team visibly doing its own work this way.