The 90-Day Plan

The Playbook so far has been the framework. This chapter is the execution sequence. Ninety days, twelve weeks, organized around the four tracks. The plan is intentionally small at the start. Most failed AI adoption efforts don’t fail because they were too modest. They fail because they tried to do everything at once, ran out of energy, and shipped nothing.

What follows is the sequence I’d run if I were embedded with a typical team starting from Level 1. Adjust to your context — but don’t skip the parts you don’t like.

Week 0: The honest conversation

Before the ninety days begin. One meeting, ninety minutes, the right people in the room — usually the head of product, the head of engineering, and one or two senior ICs whose work would be most affected.

The meeting has three jobs.

The first is the honest diagnosis. Where are we today, on which workflow, against the maturity model from Chapter 2? The honest answer is almost always Level 1. Get to that answer explicitly.

The second is the workflow choice. Pick the one workflow that will be the focus of the first thirty days. The criteria: it’s important, the variance per operator is high, the cost of inconsistency is meaningful, and the originator (the person who does it best) is available and willing.

The third is the owner choice. Pick one person to own the work over the next ninety days. This is not a side project. They need 20–30% of their time, sustained. If you can’t free that time, either pick a different owner or accept that the plan will slip; do not pretend the work will fit into the seams.

Write down the diagnosis, the workflow, the owner. Move on.

Weeks 1–4: One process, encoded (Track 1)

The first month is entirely process encoding on the chosen workflow.

Week 1. The owner sits with the originator and watches them run the process twice, on real work, taking detailed notes. The output of week 1 is a draft skill — one to two pages — capturing trigger, inputs, steps, outputs, failure modes (Chapter 3).

Week 2. The originator reads the draft, marks it up. The owner revises. A second person — not the originator — runs the workflow using the skill on real work. The output is a more accurate skill and a list of gaps the second operator hit.

Week 3. A third operator uses the revised skill. The owner captures the corrections. The skill is updated. By the end of the week, three different people have produced output using the encoded process, and the output is visibly more consistent than the pre-encoding baseline.

Week 4. The skill is published in a shared, version-controlled location. An owner is named (often the same person as the workflow owner, but not necessarily). A review cadence is set — quarterly is fine. The team is told this is now the way the workflow runs.

By the end of week 4, you have crossed Level 2 on one workflow. Don’t underestimate this. Most organizations never cross it intentionally on any workflow.

Weeks 5–8: Compiled context (Track 2)

The second month introduces knowledge architecture, scoped tightly.

Week 5. Pick three topics. The criteria: each is referenced often when the team makes decisions, and the current source-of-truth is unclear. For most product orgs, the first three are some flavor of customer segmentation, product surface, and pricing model.

Week 6. For each topic, draft a compiled topic page using the structure from Chapter 4: current position, evidence, decisions with dates, dependencies, owner, review cadence. Two to five pages per topic. The owner of each topic is the person closest to the work, not the person highest in the hierarchy.

Week 7. Wire the encoded process from weeks 1–4 to read from the compiled topic pages instead of from raw source material or operator memory. Have the same operators run the workflow again and compare the output to weeks 3–4. The quality should visibly improve.

Week 8. Publish the compiled topic pages. Designate them, in writing, as the authoritative source for those topics. Archive or mark-as-superseded the parallel sources that now contradict the compiled pages. This is the governance act that makes the compiled pages real, and it is the one most teams flinch from.

By the end of week 8, you have an encoded process running against a compiled context, both governed by named owners. This is a real Level 2 foothold across two reinforcing tracks.

Weeks 9–12: Governance and a second loop (Tracks 3 and 4)

The third month institutionalizes the practice and starts the next iteration.

Week 9. Stand up the minimum-viable governance discipline from Chapter 5. Owner per artifact. Review cadence per artifact. Staleness check (a CI job, a calendar reminder, a monthly all-hands review — pick one). One named overall governance owner, two to four hours a week.

Week 10. Move the encoded workflow up the autonomy stack from Level A (suggestion) to Level B (augmentation), as defined in Chapter 6. The agent now produces the full first draft from the encoded process and the compiled context. The reviewer’s job is to review and finalize, not redo. Capture every meaningful correction.

Week 11. Pick the second workflow. Run the same week-1-through-week-4 process on it, in parallel with the augmentation loop on the first workflow. By now, the team knows the rhythm — the second workflow goes faster than the first.

Week 12. End-of-quarter retrospective. Review the encoded skills, the compiled topic pages, the augmentation loop, the corrections captured. Update the skills with the corrections. Plan the next quarter — usually two more workflows encoded, two more topic pages compiled, governance discipline tightened where it slipped, automation promoted where it has earned.

That’s the first ninety days.

What you should have at day 90

If the plan ran cleanly:

• One workflow at Level 2 (Augmented), running against an encoded skill and a compiled context, with active correction loops.
• A second workflow encoded and running at Level B in suggestion mode, on its way to augmentation.
• Three compiled topic pages, owned, reviewed, and treated as authoritative.
• A working governance discipline: owners assigned, cadences set, staleness flagged, an overall governance owner spending two to four hours a week keeping the system honest.
• A team that has lived through one full cycle of the methodology and can run the next one largely on its own.

This is a small thing measured in artifacts. It is a large thing measured in capability. The team that has done this work has, in ninety days, become the kind of team that can keep doing this work — which is the only durable thing the methodology produces.

What to do at day 91

The temptation at this point is to scale aggressively. Encode every workflow. Compile every topic page. Promote every workflow to Level C. Resist this. The same instinct that wants to do everything at once is the instinct that produces failed AI initiatives.

The right pace from day 91 forward is: two new workflows per quarter, two to four new topic pages per quarter, one autonomy promotion per quarter, sustained governance review on everything that exists. At that pace, eighteen months from now your organization is meaningfully Level 3 across the workflows that matter — and the compounding from Chapter 7 has had a year and a half to run.

The competitive position you want is not “we shipped a transformation.” It is “we have institutionalized a discipline that compounds, and we’re a year ahead of the competitor that hasn’t started.” The 90-day plan is the front edge of that discipline, not its endpoint.

When to reach out

If you’ve read this far, you have everything you need to start. The plan is not secret. The framework is not proprietary. Most teams that read it can execute most of it themselves.

The places I tend to be useful are on the edges:

• You want a defensible read on whether your team is actually positioned to execute this — what’s missing, what’s strong, where the realistic risks are. That’s an AI Decision Review.
• You’re a 50–500 person organization and you want help running this 90-day plan against your actual operations, including a pilot that ships one or two real automations with the team that owns them. That’s an Operations AI Engagement.
• Your AI is already live and isn’t getting better, and you suspect the iteration loop is broken in ways you can’t fully diagnose from inside. That’s a Production AI Review.
• You’re at the front of a major AI product commitment — strategy, roadmap, scope — and you want a structured outside engagement to pressure-test the plan before you commit. That’s a Strategy Sprint.

Otherwise: start. The 90-day plan above is not theoretical. It works. It just has to be done.