An AI strategy for a company where nobody owns it yet
You don't need a strategy document yet. You need one named owner, three jobs that cost real hours every week, and a 90-day loop that proves something. The strategy is what you write down afterwards.

Why the first version has to be small
Most of the AI strategies we see in companies of 20 to 200 people are borrowed from much bigger ones. A steering group, a roadmap, a maturity model, a committee to pick a platform. When the department heads are already full, that plan gets agreed in a meeting and then nothing happens – every hour it needs is an hour nobody has.
Nobody owns it right now, and that's fine for the moment. What matters is that one person owns it by the end of this month, and that they're given something small enough to finish.
Pick one owner and give them a day a week
Not the busiest department head, and not whoever's most excited about the technology. In our experience the right owner is curious, writes clearly, is trusted across the business, and can be given one day a week without a department falling over. Often it's an operations or finance lead, sometimes a sales manager, and occasionally the owner themselves, if they're realistic about the diary.
Put the day in the calendar as a recurring block and treat it like a client meeting. The moment it becomes "when things are quiet" it's gone. Who makes the best AI lead goes into what to look for.
For the first quarter the owner's job is narrow: run the loop below, keep the list of jobs, and report once a month on a page.
Which three jobs to start with
Ask each department head one question, 20 minutes each: what does your team do every week that's repetitive, takes a kind of judgement you could write down, and nobody would miss doing? You'll get a list. The usual answers in the businesses we work with:
- Writing up meeting notes and getting them into the CRM.
- First drafts of replies to routine enquiries and quote requests.
- Pulling the weekly numbers together from three systems.
- Researching a company before a first call, which can easily take 20 or 30 minutes a time.
- Reformatting supplier documents into the internal template.
Choose three. The rule is that each one has to have a number you can measure before you start – hours a week, days to turn around, or how often it comes back wrong. If nobody can put a number on it, it isn't a first job.
The 90-day loop
Weeks 1 and 2 are for measuring, not building. Count the hours, time the turnaround, note the error rate. This is the step everyone skips and then regrets, because without it there's no way to say whether the thing worked.
Weeks 3 to 8: build the smallest version that does the job, with the people who do it today in the room. Not a platform, and not a pilot of everything. One job, done start to finish for one team, even if parts of it are still manual.
Weeks 9 to 12: run it alongside the old way. Keep the numbers. Watch where it's wrong, and what the people using it start doing with the time instead.
Week 13: decide, and write the decision down. Keep it and give it an owner, drop it and say why, or widen it to the next team. Then choose the next three jobs.
What to write down at the end
That's the strategy, and it fits on a page:
| Section | What it says |
|---|---|
| What we use AI for | The jobs that passed, and the next three in the queue. |
| What stays with a person | The judgement calls you've decided not to automate, and why. |
| Who owns it | The owner, the day a week, and who they report to. |
| The rules | What data it may see, who checks its output, who may build. |
| The money | What this quarter cost and what the next one's allowed to. |
A page is enough, because the strategy will change in three months anyway. A 30-page document won't get read, and it'll be wrong by the time it is.