AI Adoption Plan: The First 90 Days, and How to Measure It

Most AI rollouts do not fail on day one. They stall somewhere around week six. The licences are bought, the workshop has happened, three enthusiasts use AI every day and everyone else has drifted back to the old way of working. Nobody planned what happens between the first session and the day AI is simply part of how work gets done.

This is the 90-day plan we run with 10-200 person companies, in the order we run it, plus the numbers that show whether it is working. It assumes no data scientists, no new platform and no spare full-time person. It does assume one person in management who cares.

What does a 90-day AI adoption plan look like?

The short version:

• Before day 1: one sponsor, one primary tool, a short survey and three tasks timed as a baseline.
• Days 1-14, foundations: one workshop for everyone, one page of data rules, one shared way of prompting.
• Days 15-60, build: each function builds one working assistant for a task that recurs at least weekly, with a named owner.
• Days 61-90, put it to work: mentoring, a review ritual, a written summary for management and a licence decision based on usage.

Three rules hold it together: one primary tool, one owner per workflow, one baseline measured before you start.

Before day 1: four decisions that save you a month

1. One sponsor in management. Not a steering group. One person who can decide on tools, time and licences, and who gives it about half an hour a week. In a 30-person company that is usually the CEO or the COO.

2. One primary tool. Start with the assistant you already pay for: Copilot if you run on Microsoft 365, Gemini if you run on Google Workspace. If people already use ChatGPT or Claude, move them onto a business plan, where the vendor does not train its models on your data by default. We compare the four in a separate guide.

3. A short survey. Ten questions, ten minutes: which AI tools people use today and on which accounts, which tasks eat their week, what worries them. It tells you where to start, and it usually shows how much company work already runs through personal accounts. The answers are personal data, so handle them like any other.

4. Three tasks with a baseline. Pick three tasks that recur at least weekly and time them now, before anyone is trained. Good candidates: answering a standard customer enquiry, drafting an offer from a template, preparing the monthly management report, summarising a supplier contract for a colleague. Without this number, every later result is an opinion.

Days 1-14: shared foundations for everyone

• One workshop for everyone. Half a day, built on the survey answers, with more than half of the time spent on the team's own document types, in the tool you license. Not a tour of twenty tools. We wrote separately about how to tell a good AI training from a forgettable one.
• The traffic light rule. Green is public or anonymised information and can go into any account, a personal one included once model training is switched off. Yellow is client and business correspondence: company account only, and a person reads the output before it is sent. Red is passwords, bank access and employee personal data, which go nowhere. One page, pinned where people work. Here is a one-page template you can copy.
• A shared way of prompting. We teach three levels: just ask (about ten seconds), a structured prompt using ROCSRS for anything that goes to a client or a manager (two to five minutes), and letting the AI write the prompt for recurring tasks. The more important the result, the more time the prompt deserves.
• A personal master prompt. One standing instruction, set once: language, tone, length, plus two lines that prevent most embarrassments. "Tell me first what is missing or contradictory." "Do not invent prices, deadlines or quantities."

By day 14, everyone has used the company tool on a real piece of their own work, the data rules exist on one page and the three baseline tasks are confirmed.

Days 15-60: each function builds one working assistant

This is where most plans stop, and it is where adoption is actually decided. Training makes people capable. Building something for a real weekly task makes them use it.

Split the company into function groups: sales, finance, customer service, operations, HR, whatever matches your organisation chart. Each group gets a 2-3 hour working session and leaves with one assistant (a Copilot agent, a ChatGPT project or a Claude project) for one task that comes back at least every week. A task that happens once a quarter does not deserve an assistant.

"Working" has a definition:

Element	What it is	Example: customer service, delivery-time enquiries
Instructions	What the assistant does, in what tone, and what it never does	Reply in the customer's language, quote only confirmed delivery dates, write [CHECK DATE] when unsure
Starter prompts	Two or three ready-made openings	"Draft a reply to this enquiry", "Shorten this for chat"
Test data	Fictional but realistic inputs	Five made-up enquiries, including an angry one and one without an order number
Pass criteria	What a good answer must do	Right language, no invented date, under 120 words, polite close
Owner	One named person who edits the instructions	The team lead who answers these enquiries most often

A few rules from our programmes. One assistant, one job. Build version 2 next to version 1 instead of overwriting it. When the assistant gives a confident wrong answer, add one rule to the instructions and rerun the tests. The owner edits the instructions; everyone else reports what broke. And in our experience with Copilot agents, instructions of 3,000-4,000 characters are followed more precisely than ones that use the full 8,000-character limit.

By day 60, each group has one assistant that passes its tests, is used every week and is documented on one page: input, instructions, output, owner. That is the smallest version of what we call a shipped AI solution.

Days 61-90: put it to work and decide what comes next

• Mentoring per group. Real use breaks things the tests did not catch. Short sessions with each owner fix them.
• A review ritual. Every two weeks, half an hour per owner: is it used, what broke, what changed in the instructions. Without this, quality decays quietly.
• Re-time the baseline at day 60 and day 90.
• A written summary for management. What is live, who owns it, how much time it saves, what comes next. Two pages, no slides.
• A licence decision based on usage. Who uses the paid tool every week, who does not, who needs more. Look at the admin usage report, not at enthusiasm. One Microsoft 365 trap: removing a person's Copilot licence orphans the agents that person built, so hand over ownership first.

The 90 days on one page

When	Focus	Who	Done when
Week 0	Sponsor, tool, survey, baseline	Sponsor and one coordinator	Tool chosen, three tasks timed
Weeks 1-2	Workshop, data rules, prompting method	Everyone	Everyone has used the tool on real work; rules fit on one page
Weeks 3-5	Function sessions, version 1 of each assistant	Each group and its owner	One assistant per group, with test data and pass criteria
Weeks 6-8	Testing, fixes, weekly use	Owners	Assistant passes its tests, is used weekly and is documented
Week 9 (day 60)	First re-timing	Owners and sponsor	Same tasks compared against the baseline
Weeks 10-13	Mentoring, reviews, management summary	Sponsor and owners	Licence decision, next three workflows chosen, second re-timing at day 90

How do you measure AI adoption?

Measure four layers, in this order. Each one only makes sense once the one before it is healthy.

Layer	What to measure	How	When
Adoption	Weekly active users of the company tool	Admin usage report (in Microsoft 365: Copilot Analytics)	Every two weeks from day 14
Workflow	Tasks done with each assistant per week; time per task against the baseline	Owner's tally; re-timing	Day 60 and day 90
Quality	Owner sign-off against the pass criteria; errors caught in checks	Owner reviews a sample of outputs	At every review
Business	Hours returned per week; response times to customers	Time saved per task multiplied by weekly volume; timestamps in email or ticketing	Day 90, then quarterly

What to ignore: the number of prompts sent, the number of licences bought, generic productivity percentages from vendor surveys and how much people enjoyed the workshop. They feel like progress and tell you nothing about your own work.

The baseline method

1. Pick three recurring tasks before day 1.
2. Time each one several times with the people who normally do it. Note the type of input, because a short enquiry and a long complaint thread are not the same task.
3. Re-time at day 60 and day 90 with the same people and the same kind of input. Count the corrections needed before the output was usable.
4. Hours returned per week = (minutes before minus minutes after) × tasks per week ÷ 60.

It is unglamorous and takes about an hour. It is also the only number your CFO will trust.

Where do 90-day plans usually break?

• No owner. The assistant belongs to "the team", so nobody fixes it, and by month four nobody uses it. We wrote more on why AI adoption fails.
• Too many tools. Three tools mean three sets of rules, three logins and three times the questions. Pick one.
• Pilots with no test. Without test data and pass criteria, every judgement is a mood, and the first wrong answer kills trust.
• Measuring ROI in week 2. People are still learning and early numbers are noise. Measure adoption first, time per task at day 60, business effect at day 90.
• Training without a build phase. A workshop on its own changes how people talk about AI, not how they work.

What happens after day 90?

There are two sensible paths. Run it yourselves: the owners keep their review ritual, the sponsor picks the next three workflows each quarter, and new joiners get the one-page rules and the prompting method. Or bring in an external lead: as a Fractional AI Operations Lead we ship one or two new solutions a month on a rolling basis, each with an internal owner. Either way, the first 90 days should leave you with owners and documentation, not dependence on an outsider.

What to do this week

• Next 15 minutes: name the sponsor and write down which AI tool you already pay for.
• Next 7 days: send the ten-question survey and time your three candidate tasks.
• Next 30 days: run the foundations workshop and book the first function group session.

Sources: Microsoft 365 Copilot plans and Copilot Analytics; OpenAI: how business data is used; Anthropic: consumer and commercial terms. Product and legal facts checked on 29.10.2026.