Write your AI strategy second. Not never, and not carelessly: second. A strategy is a genuinely useful document, and the version you write after running one job for three months beats the version you write before, because it holds things you know rather than things you have read.

What an AI strategy is genuinely good for

A good strategy does four jobs that nothing else does. First, it sequences the work, so people know what comes now and what waits. Second, it sets boundaries, naming what this business will not do whatever the tool can manage. Third, it puts a figure on what you will spend before you see a result. Fourth, it hands everybody the same answer when a customer, an insurer or a new starter asks what you are doing about AI.

Those four are worth the effort. A business without them tends to buy tools in a scatter, argue the same argument every quarter and discover its boundaries only after crossing one. So the case for writing an AI strategy is strong.

The trouble is never the writing. The trouble is spending months drafting, circulating and redrafting while nothing has actually run. Parts of the document depend on evidence, and until you have some, those parts get filled in with assumption.

The parts you can only write from evidence

Those four strengths all rest on knowing your own business. Read every guide published this year and there are still five things you cannot know until a job has run.

Which jobs suit your material comes first, and that depends on how your documents are stored and how consistent they are. Next is how long checking takes, because checking is the cost that decides whether a job pays for itself. Third is how a tool behaves on your worst input rather than on a clean sample. Fourth is how much of your process exists in writing rather than in people’s heads. Fifth is a realistic timeline, because nobody else’s published timeline is yours.

An AI strategy written first has to guess at those five. The guesses then harden into commitments, targets and budgets, and a year later you are managing against numbers that nobody was in a position to know.

Month one: choose the job and name the decision

Those five are what the next three months are for, and month one goes on choosing the job. Pick an internal one with regular volume and a clear right answer. Starting where nobody sees it gives you room to be wrong safely.

Write five lines about that job before anything else happens. What starts it. What goes in. What comes out. Who reads the output before it goes anywhere. What decision this job will settle for you.

That last line matters most and gets left out most. “We want to see if AI helps” settles nothing. “We want to know whether a person can review this output in under 5 minutes” settles a great deal, and agreeing what good looks like at this stage keeps month three honest.

Month two: run it inside the real work

With the job written down, month two is for running it where the work actually happens. Not a side project, not a trial folder, not a demo on tidy examples. Real input, real volume, real deadlines and the people who normally do it.

Resist changing the job halfway. Adjust the instructions and the wording as you go, by all means, but leave the trigger, the reviewer and the output alone. Change those and you lose the comparison you spent month one setting up.

Keep a simple note through the month: what needed correcting, how long the check took, and anything the tool got confidently wrong. Three lines a day is plenty, and a shared document beats a notebook because other people can add to it.

That note becomes the evidence for everything you write later. Without it you reach month three with impressions, and impressions lose arguments to whoever speaks with most confidence.

Month three: keep it, change it or stop it

By month three you have enough to decide, and there are only three honest outcomes. Keep it as it is, because it clears the standard you set and the checking cost is acceptable. Change it, because one part works and another does not. Stop it, because the saving never appeared or the checking swallowed the saving whole.

Stopping is a proper result, so treat it as one. You have paid for an answer, and you now know something about your material that no amount of reading would have told you. That knowledge belongs in the document you are about to write.

Whichever outcome you land on, say it out loud to the people involved and say why. A decision that nobody announces gets remembered as a project that fizzled, which makes the next one harder to start.

What to write once the three months are up

Now the AI strategy has something to say. Write it short and write it specific, because the value now sits in detail that three months bought you.

Name the kinds of work that suit your business and the kinds that do not, with a reason beside each. Give a real cost per job, checking time included. State your review capacity honestly, since that is the true limit on how many jobs you can run at once.

Then set the rule for what information may go into a tool and what may not, name an owner for each job, and put a spending limit against the next 12 months. A strategy built this way is usually a few pages long, and people follow it because they recognise their own business in it.

The conditions that say a job is not ready

Your strategy should also say when to leave something alone, and these conditions are the most useful thing in it. A job is not ready when the process exists only in somebody’s head, or when nobody can point to the authoritative copy of the source material.

Leave it alone, too, when no one can describe a good output in a sentence, or when nobody has the time to review what comes back. The same applies when volume is so low that a month of running teaches you nothing.

Written down, those conditions save you from the enthusiastic suggestion that arrives every few weeks. They give you a reason to say not yet that is about readiness rather than appetite.

You could not have written those conditions in month one. Writing the AI strategy second costs you a quarter. What it buys is a document where every line answers to something you have watched happen.