A practical 90-day AI plan for a small business

Most AI advice aimed at small businesses is borrowed from large ones. It talks about transformation programmes, centres of excellence and multi-year roadmaps, none of which fit an organisation of fifteen or fifty people with no spare capacity and no patience for theatre. A smaller business does not need an abstract programme. It needs a short, controlled sequence that produces real evidence, builds a little capability, and creates enough momentum to justify the next step.

Ninety days is the right unit. It is long enough to change how one or two pieces of work actually get done, and short enough that nobody loses interest or budget before there is something to show. What follows is a plan built around that window: understand, then build, then adopt.

Days 1 to 30: Understand

The temptation is to start with a tool. Resist it. The first month is for getting an honest picture, because almost every failed adoption we see skipped this step and built on a guess.

Start with leadership objectives. What is this actually for? More capacity without more hiring, faster turnaround, fewer errors in a risky process. Write it down in a sentence, because everything later is judged against it. Then map the existing use, including the quiet, unofficial kind. People are already using AI; you simply may not know where. Surface it without blame, because it is your best evidence of what helps.

Next, do a light risk discovery: what client or personal data must never be pasted into a public tool, and where could a confident, wrong answer cause real harm. This does not need a consultant or a committee. It needs a clear-eyed hour and a short written rule. Then list candidate workflows, the frequent, worthwhile, checkable tasks where AI is likely to help. The field experiment behind the "jagged frontier" is a useful warning here. In it, AI helped knowledge workers on tasks suited to it but actively hurt them on a task just outside its range. Choosing the right task matters more than choosing the right tool.

Finally, take a baseline measure. If you cannot say how long the chosen work takes today, or how often it goes wrong, you will never be able to show that anything improved. The government's AI Opportunities Action Plan frames adoption as "scan, pilot, scale" for the same reason: look honestly before you leap.

Days 31 to 60: Build

The second month is where one or two pieces of work actually change. Select no more than two use cases from the candidates, the ones with the best balance of value and low risk. Two is a deliberate ceiling. A small business that tries to change five workflows at once changes none of them.

For each, create reusable instructions and context: the standing prompt, the house style, the examples, the definitions, so the work does not start from scratch every time and does not depend on one person's private knack. This is the most valuable single move a small team can make, and it is what turns a clever individual habit into something the business owns.

Test with a small group, and deliberately include someone sceptical, not just the enthusiasts. Record the failures honestly, because the failures are the most valuable output of the month: they tell you where the human checks need to sit. The research on customer-support teams by Brynjolfsson, Li and Raymond found AI raised productivity by around 14% on average and more for less experienced staff, but those gains came from a well-defined, repeated task with good examples, exactly the conditions you are trying to build here.

Then establish human approval. Decide, before anything goes live, which outputs a person must check and who that person is. The NIST AI Risk Management Framework treats this as designing your checks in, not hoping for them later.

Ninety honest days beat a three-year programme nobody finishes.

Days 61 to 90: Adopt

The final month decides whether any of this survives contact with normal working life. Train the people who will actually do the work, using your own real examples from the build phase rather than a generic course. Formalise the workflow: write the short version of how it now runs, so it does not live only in one head.

Measure the results against the baseline you took on day one, and be honest if the answer is disappointing. Assign ownership, a named person responsible for keeping the workflow healthy as the tools change. Then make the decision the whole plan exists to enable: for each use case, stop it, improve it, or scale it. Some will not have worked, and saying so plainly is a feature of the plan, not a failure of it.

The honest limits

Two qualifications. First, ninety days is a cycle, not a finish line. The output is not "AI, done"; it is a couple of workflows that genuinely changed, a team that learned how to do this, and the evidence to choose the next two. The value compounds across cycles, which is exactly the lag the productivity J-curve describes.

Second, not every business should start today. If there is no leadership attention to give it, or no single task worth improving, a rushed plan will just produce a tidy failure. The plan rewards a real commitment of attention, modest but genuine. Without that, the honest move is to wait until it exists.

What to do next

Block the first week. Get leadership to agree the one-sentence objective, surface the AI already in use, and write the short data rule. That alone puts you ahead of most organisations, who skip straight to a tool and wonder why nothing sticks. The rest of the plan follows from that honest start. And resist the urge to widen the scope mid-cycle: the discipline of finishing two workflows beats the comfort of starting six and completing none.

The tool

To run the full ninety days without losing the thread, we have built the 90-Day Small-Business AI Adoption Planner: a week-by-week planner across the understand, build and adopt phases, with named owners, review points, the measures to capture, and the stop-or-scale decision built in at the end.

Download the 90-Day Small-Business AI Adoption Planner (PDF)

Running this with a team, rather than alone, is the 90-Day Adoption Build: the core engagement where an embedded, fractional adviser turns scattered AI use into a couple of real, owned workflows. It follows directly from choosing the right use case and from treating experiments as the raw material of a strategy, and it is the practical answer to why so many pilots never scale.

Sources and further reading

• UK Government, AI Opportunities Action Plan, January 2025. Source for the "scan, pilot, scale" sequencing of adoption.
• Dell'Acqua et al., Navigating the Jagged Technological Frontier, Harvard Business School / BCG working paper, 2023. Independent field experiment. Source for the importance of task selection.
• Brynjolfsson, Li and Raymond, Generative AI at Work, NBER Working Paper 31161, 2023. Independent. Source for where real productivity gains come from.
• NIST AI Risk Management Framework. Independent US standards body. Source for designing human checks in by intent.
• Brynjolfsson, Rock and Syverson, The Productivity J-Curve, NBER Working Paper 25148, 2018. Independent. Source for value compounding across cycles.