Your First 90 Days
You got the job nobody quite defined: make AI useful here.
Or you are about to ask for that job. Either way, the clock started before you noticed. Someone said "we need an AI strategy" in a meeting, heads turned toward you, and now you are the strategy.
Most leaders make the same two mistakes in opposite order. First they buy something. Then they announce something. Buying before listening. Announcing before proving.
There is a better sequence, and it fits in 90 days.
The wrong 90 days: Announce a vision. Buy licenses. Run a pilot with fake data. Present a deck about the pilot. Wonder why nobody uses it.
The right 90 days: Listen for 10 days. Prove one thing in 30. Template it by 60. Then make a decision you can defend, including the decision to kill it.
Every phase ends with evidence. No phase requires a purchase you cannot explain to a skeptic.
The 60-second version
- Days 1-10: Interview 8 to 10 people across functions. Find the repeatable work that eats weeks, including the AI tools people already use quietly. Buy nothing. Announce nothing.
- Days 11-30: Put one process live with one willing team. Run it on real work for two weeks. Measure before and after.
- Days 31-60: Turn the win into 3 to 5 reusable skills. Set baseline metrics for the next two bets. Start the correction loop as a team habit.
- Days 61-90: Pick a door. EXPAND, FORMALIZE, or KILL. Each door needs one number to justify it.
- Throughout: The human owns every outcome. AI never sends, spends, decides about people, or publishes numbers on its own.
Days 1-10: Listen and map
You are not allowed to buy anything yet. You are not allowed to announce anything yet. The temptation to do both on day one is exactly why this rule exists.
Interview 8 to 10 people across functions: the person who closes the books, the one who writes the proposals, someone in operations, someone in support. Ask one question, several ways: what do you do every week that a careful intern could learn in a day?
That is your inventory. Repeatable, teachable, currently expensive.
Now ask the awkward question: what AI tools are you already using? People will admit to things they never put in a policy. Treat this as data, not a violation. Shadow use is the most honest signal in your company. It marks the pain sharp enough that people solved it themselves.
Map the results. Rank processes by time spent multiplied by frequency. Your day 10 deliverable is a one page map: the five most painful repeatable processes, who owns each, their hours per month, and which ones already have shadow AI in play.
You now know more about your company's real work than most of the people in it.
Days 11-30: One team, one process, real numbers
Pick one process with three traits: real pain, a team lead who volunteered, and a baseline you can measure. No measurable before, no provable after. Do not build it.
Build the project directory and one skill for the process, the way Part 2: Run AI Like a Project, Not a Chat describes. Then run it on real work for two weeks. Real invoices, real drafts, real tickets. A pilot with fake data proves the AI works in a demo. A pilot with real work proves it works in your business.
Measure before and after the way Part 5: Prove It Pays describes. Hours per cycle. Error rate. Turnaround time. Write the numbers down while they are fresh. Memory rounds up.
At day 30 you have exactly one of two things: a measured win, or a measured miss. Both are useful. Only the win continues.
Days 31-60: Template the win
Assuming the numbers held, turn the win into 3 to 5 reusable skills. Small skills, each doing one job well, each with its own correction history.
Set baseline metrics for the next two bets. That is how adoption compounds without depending on you personally.
And start the correction loop as a team habit, the way Part 3: Memory That Travels With You describes. Every correction gets recorded and reviewed weekly. This is the moment AI adoption stops being your project and starts being their habit. A team that corrects the system owns the system.
By day 60, a new team should get productive from the skill library and the changelog in days, not weeks. That is the difference between a demo and an operation.
Days 61-90: Pick a door
Three doors. Each one needs one crisp criterion. Vagueness here is how zombie pilots live forever.
EXPAND. The numbers justify more teams. The criterion: a measured improvement you can defend in a meeting, sustained across two full weeks of real work. Hours saved, errors cut, or turnaround shortened. With the receipts.
FORMALIZE. One process becomes standard operating procedure. The criterion: the team runs it without you prompting for two weeks straight, with a named owner and written review gates. It has an owner, a standard, and a review rhythm.
KILL. The numbers do not justify it. The criterion: no measured improvement after two real weeks, or the cost of human review exceeds the time saved. Kill it publicly. Say what you learned. Reallocate the time. A program that cannot kill things cannot earn trust, because everyone knows the next proposal will be funded no matter what the numbers say.
Notice that two of the three doors are successes. Killing is not failure. Killing is the proof that the measurement was real.
The rules for review
The AI drafts. A human approves. Every external send needs a human click. Every number the AI produces gets checked before it leaves the building.
Here is what the AI may never do alone: send anything outside the company, make decisions about people, spend money, or publish numbers.
If something goes wrong, "the AI did it" is not an explanation. It is a confession. The human owns the outcome, always. Write that on the first page of the project directory. The week you need it will not be the week you remember it.
The one page version for a CEO or board
Boards fund clarity, not enthusiasm. When you present this plan, bring one page:
THE BET
One sentence: which repeatable process we will speed up, and by how much.
THE COST
Staff hours and any tool licenses. Nothing else.
AI adoption at this stage needs no new headcount.
EVIDENCE DATE
The date by which real before and after numbers exist.
Not a demo date. A measurement date.
KILL CRITERIA
The number that must appear or the work stops.
Written before we start, not after we see the results.
WHAT WE NEED FROM YOU
One willing team. Permission to measure honestly.
Agreement to kill it if the numbers fail.If the CEO reads that page and says yes, you have a mandate. If the CEO wants to skip the kill criteria, you do not have a mandate. You have a hobby with funding, and hobbies do not survive budget season.
What this creates for a business
Ninety days later, one of three true things is on the record.
Either you have a measured win, a skills library, and a team running the correction loop on its own. Or one process is standard operating procedure, with an owner and review gates. Or you have an honest kill, a public explanation, and reallocated time.
All three beat the usual outcome: licenses purchased, pilots demoed, nobody quite sure if any of it paid.
What compounds from here:
- A skills library the next team inherits instead of reinventing.
- A correction habit that makes the system smarter every week without your involvement.
- A playbook for the next 90 days: the same four phases, run faster, because the method is written down.
- Credibility. You never have to say "trust me, it works." You have the numbers, including the ones that said stop.
That is AI adoption with a backbone.
Frequently asked questions
What if I inherit a team whose last AI attempt already failed?
That is an asset, not a setback. Interview the failure before you plan anything: what broke, who owned it, how it was measured. Most failures come down to missing ownership or missing measurement, both fixable. Starting from the autopsy also earns you credibility with a skeptical team.
Leadership wants results in 30 days, not 90. Now what?
Days 11 to 30 are your 30-day win: one workflow live with one team and real before-and-after numbers. The 90-day frame builds the system around that win. Give leadership the day-30 number loudly, and keep building quietly.
How do I handle the team's fear that AI will replace them?
Name it in week one, directly. Frame the work as removing the tasks nobody wants: status formatting, first-draft drudgery, meeting-note cleanup. The people who learn to run the skills become the most valuable people in the room, because they own the system everyone depends on.
What if my first workflow pick flops?
Kill it fast with honest math. That is a days 61 to 90 decision made early, which is fine. Pick the next workflow on the list. Two fast failures with clean autopsies teach more than one slow zombie project that nobody will admit is dead.
Do I need budget approval to start?
No. You need a folder, the AI tools the company already licenses, and one willing team. The budget conversation happens around day 61, and by then you walk in with measured numbers instead of a slide deck of promises.
Final checklist
Before day one, confirm:
- 8 to 10 interviewees identified, across at least three functions.
- One process candidate with a measurable baseline and a volunteered team lead.
- Shadow AI use inventoried by asking, not assumed or punished.
- Kill criteria written down before any work starts.
- Review gates defined for every external send, published number, and decision about people.
- The one page plan dated, the evidence date on someone's calendar.
- The project directory exists: skill, baseline, changelog, owner.
Then start the clock.
One question for you
If your first 10 days started Monday, whose desk would you visit first?
Write the name down. That person is your inventory, your first team, and your hardest skeptic, all in one chair.