Phase 3: Spread

Phases 1 and 2 were the friendly rooms. Engineering already ran on pipelines. The second function had an owner who raised a hand. Everything that has worked so far worked with people who wanted it to.

Phase 3 is the rest of the building. Finance, legal, people, the sales team that hit its number last quarter and sees no reason to change anything. Owners who didn't volunteer. Workflows nobody ever wrote down, because the person who runs them has run them for years and never needed to.

It's also where the question everyone has been too polite to ask finally gets asked, whether you ask it or not. What happens to the jobs? You answer it here, one function at a time, before the work moves, and in writing.

Spreading is a capacity problem and an honesty problem. Get both right and the tooling takes care of itself.

What happens

The remaining functions move in waves, one owner each, sized to review capacity. Constraint Court rulings from earlier phases become the playbook.

The gate

Core workflows at rung 5 in every function in scope

What changes for the people doing the work

Roles are redesigned function by function around judgment and escalation. The headcount conversation happens here, in the open, for each function.

Waves, not a launch

The tempting version is a launch. An all-hands, a date, every function at once, a slide with nine logos on it. Two of the timeline rules kill that plan before it starts. Review capacity is the ceiling: how fast humans can review sets the pace, never how fast machines can generate. And one leader can't write intent for two functions in the same stretch.

So a wave is exactly as wide as the number of owners who can each hold one function, and exactly as wide as your reviewers can absorb. Not one function more.

If that sounds cautious, remember what happened in engineering when generation outran review: output went up and the constraint moved to review, where nobody had resized it. A company-wide launch is that story with the month-end close and the contract queue in it.

Pick the order with the same three signals you used for the second function: an owner who wants it, enough volume to matter, output that can be measured. The function where all three are yes goes in the next wave. The function where nobody will put their name down goes last, and "nobody owns it" gets a Court hearing like any other claim. No owner, no constraint: it's treated as false until someone proves otherwise.

Inside each function, move the checkable, high-volume tier first and keep people on the hard tier. Klarna learned this in public. Its CEO told Bloomberg "we went too far… focused too much on cost. The result was lower quality" and restarted hiring human agents. The AI stayed on the front line, doing work he put at the equivalent of 853 FTEs. The mistake was never automating support. It was automating the tier that needed judgment. The starter core workflow lists at intake already draw that line for you: the escalation column is the tier that stays human.

The playbook you already paid for

Every Constraint Court ruling from phases 0 to 2 has an owner and a date on it. That's the playbook for this phase. You paid for those hearings once. Don't pay for them again.

Security got negotiated for engineering: scoped credentials, isolated execution, zero-retention vendor terms, a named security reviewer. That package walks into the finance wave without a second hearing. The customer contract clause legal read for support covers sales, because it's the same clause. The regulated-work ruling that mapped controls to the ladder and put the audit trail in the pipeline applies to every function that faces the audit.

So the rule for this phase: a constraint that has been heard isn't heard again unless something changed. When a wave owner arrives with "security won't approve it," hand them the ruling, the owner and the date. Pull the full Court transcript from your takeaways and give it to every wave owner before their first workflow moves. Retell it in their language, the way phase 2 taught you to. A finance lead doesn't need to hear about sandboxes. They need to hear that the agent can't post a journal entry the controller didn't define.

New constraints will show up, and some will be real. Legal is the obvious one. The independent benchmark in Phase 2 had lawyers beating the tools on redlining, while some tools beat lawyers on extraction. That doesn't stop legal from moving. It tells you why anything off-playbook escalates to a lawyer at rung 5, which is how the core workflow is already written. Take new claims to the Court and let it rule. The ruling gets designed around, priced, or given a date to stop believing it.

The gate

Core workflows at rung 5 in every function in scope. Every one.

A function with four workflows at rung 5 and one at rung 3 has a person sitting in the middle of that fifth workflow, interpreting, prompting, pasting. That person is the speed limit for the whole function. The bar is strict because a partial pass hides exactly that person.

Rung 5 also needs a full drift cycle. A function's intent has to survive at least one scheduled review before it counts: monthly for operational functions, quarterly for strategic ones. A function that climbed to rung 5 last week hasn't passed. It has started the clock.

Watch for the quiet exit. The week before the gate, the hardest workflow in a function tends to fall off its list. Maybe renewals get declared "too relationship-driven." Maybe the close gets declared "special." Taking a workflow off the list goes to Court, like any other claim. Who owns the exception, by name? What would have to be true for it to move? What does it cost to leave it where it is? If the answers are good, the ruling says so, on the record. If they aren't, the workflow stays.

What doesn't count against you: the permanent escalations. Hiring decisions, performance ratings and pay stay human forever. A people function at rung 5 still has a person making every hire and no-hire call. That's the design. Rung 5 means the machine does the hands-on work and hands the call to a person when intent runs out, or when the call was never the machine's to make.

The headcount conversation

It happens here, in the open, for each function.

Not in phase 0, when nobody's job changes yet and any number would be a guess. Not afterwards, when people learn it from a new org chart. Here, because phase 3 is the first moment you can actually see it: which workflows moved, what's left, how heavy the escalation load runs. Before that you were speculating. After it, you'd be announcing.

Start with what you can't promise. You don't control headcount, so "nobody is losing their job over this" is a promise made with somebody else's authority, and everyone in the room hears the footnote. What you can commit to is smaller and survives a bad quarter: how the decision would get made if it came to that, that nobody's standing depends on the tooling looking good, notice from you early if the team is going to change, and real investment in what the job becomes.

Then refuse the spreadsheet trick. "One agent equals half an analyst" is the math a lot of plans are running, and it's wrong on both sides. Agents are not headcount: a person absorbs ambiguity and carries context forward, an agent consumes attention and returns none. What a function needs after rung 5 is people who can hold its judgment calls and handle its escalations. Count that, not fractions of a seat.

The public record shows both ways to do this. Andy Jassy put it plainly: "We will need fewer people doing some of the jobs that are being done today, and more people doing other types of jobs." You can dislike the news and still respect that it was said. Duolingo went the other way. An "AI-first" memo drew heavy backlash, the CEO later said "I did not give enough context", and in April 2026 the company dropped AI usage as a performance-review metric. Salesforce cut support from about 9,000 to about 5,000. Gartner predicts half of companies that cut customer service staff because of AI will rehire by 2027, and found only 20% of the 321 leaders it asked had actually reduced staffing. A prediction, not a measurement. It still tells you which way the early cuts are breaking.

You can't promise headcount. You can promise nobody hears it second-hand.

So here is what "in the open" means. For each function, before its first workflow moves past rung 3, you and that function's owner stand in front of its people with a short written note. Which workflows move. What stays with people. What you know about numbers and what you don't. The pressure behind the decision, whether it's cost, the board, cash, or a competitor, stated with the same weight as the decision itself. Then you say the same thing, unchanged, every time it comes up.

One more line belongs in that note: what the numbers are for. They're for deciding whether the change is worth it. They aren't an input to anyone's review. A frightened function reports what looks good, hand-writes whatever the agent struggled with, and hands you a beautifully positive rollout you can't evaluate. Usage was never a result anyway. Duolingo found that out a year after the memo.

What changes for the people doing the work

Roles get redesigned function by function, around judgment and escalation.

The new job description is already written. Read the escalation column of each function's core workflows. In support: legal threats, key accounts, the customer who disputes the answer twice. In finance: variances over the threshold, new vendors, mismatched invoices. In sales: non-standard discounts, renewals that show risk. Those were always the hard part of the job. Everything else was the cost of getting to them.

Three kinds of work form in every function. The best operator becomes the intent owner, as in phase 2. The people who used to do the hands-on work handle escalations, which is where their years of judgment finally get spent on something only they can do. Someone reads the intent against what actually happened, which becomes a standing role in phase 4. Write those roles down before the workflow moves, with names, so nobody finds out their job changed by watching it happen.

Keep a way in. Employment for 22 to 25 year olds in the most AI-exposed jobs is now about 19% below trend, and it's happening through fewer hires rather than layoffs. That's the cheapest cut to make and the most expensive one to regret. AWS's CEO called replacing juniors with AI "one of the dumbest things I've ever heard". Anthropic's own study names the trap underneath: supervising the machine takes the very skills that wither with overuse. An escalation handler who never learned the work can't judge the escalation. So each function decides, in writing, where its next people learn the job. Hire the juniors. They find the gaps in your intent months before your agents do.

The escalation column is the new job description. It was always the part that mattered.