A 90-Day AI Adoption Plan

Ninety days is enough to move from scattered tool use to one working, measured AI workflow. It is not enough to transform a business, and any plan that claims otherwise will end with a deck rather than a result.

This guide sets out what to do in each phase, what to deliver, and how to know whether the program is working. It is the companion to How to Adopt AI in a Small Business.

The objective of ninety days is one workflow that runs without prompting, with a check that happens and a result that can be shown.

The shape of the ninety days

Five phases, each with a deliverable rather than an activity.

The sequence is deliberate: rules and workflow before live work, training before scale, measurement before continuation.

Days 1–15: assess and choose

• Complete the readiness assessment across data, process, people, policy and measurement.
• Publish a one-page policy if the policy dimension scores low, which it usually does. This is a one-afternoon task and it removes the most immediate risk.
• Shortlist three use cases and score them on frequency, cost, verifiability and containment.
• Select one, and name its owner. One person, by name, accountable for the output.
• Capture the baseline: hours per cycle, cycles per year, correction rate and where corrections appear.

See The AI Readiness Assessment and Choosing Your First AI Use Case.

Days 16–30: design

• Document the current workflow, including the review step, so the change is visible.
• Design the new workflow using the six-step pattern: input, draft, check, correct, approve, record.
• Define the check: what is verified, by whom, and how it is recorded.
• Agree the data rule for this use case specifically, and confirm it against the policy.
• Build the first prompt playbook entry for the task, with a worked example.

The design phase produces two artefacts — a workflow document and a playbook — and both should be short enough to be read in a sitting. See Designing a Human-in-the-Loop Workflow and Designing Prompt Playbooks.

Days 31–60: pilot

• Run the workflow on live work, at least two full cycles.
• Log every correction, with its cause: source error, prompt issue, check missed, training gap.
• Time the cycle honestly, including the verification step.
• Hold a weekly fifteen-minute review with the owner to look at the corrections, not the progress.
• Do not scale yet. The pilot exists to find the problems while they are cheap.

Two cycles is the minimum for an honest read, because the first cycle always includes setup noise.

Days 61–75: train and formalize

• Deliver role-specific training on the workflow, not the tool.
• Practise the verification step on outputs containing seeded errors — a fabricated citation and a plausible wrong figure.
• Circulate the policy and confirm receipt.
• Add the workflow to onboarding, so it survives staff changes.
• Record the workflow and the playbook where the team can find them.

See Training Your Team to Use AI Well and the Prompt Playbook.

Days 76–90: review and decide

• Compare against the baseline on all four measures: usage, time, quality, risk.
• Report the result honestly, including the parts that did not improve.
• Decide: scale, adjust or stop — and record the reason.
• Write down the lessons, because the second use case is built from them.
• Update the roadmap with the next use case, its owner and its gate requirements.

The output of day 90 is a decision and a documented method. The decision matters to the business; the method is what makes the next ninety days cheaper.

What to measure

Four measures, captured monthly and reported in one page.

• Usage — cycles completed by the people the workflow is for.
• Time — cycle time against the baseline, including verification.
• Quality — corrections per cycle, and the stage at which they are found.
• Risk — verification steps recorded, incidents and their causes.

If a measure cannot be captured, say so rather than substituting a proxy. An honest gap is more useful than a comfortable number. See Measuring AI Adoption and Return.

What can go wrong in ninety days

Four things derail the ninety days, and all four are visible early.

• The baseline is skipped. It feels like a delay, and it is the thing the review depends on. Capture it in the first week or the program cannot prove anything.
• The use case expands. A stakeholder adds a second audience or a second data source, and the check that made the use case suitable disappears. Hold the scope; add the second audience in the next ninety days.
• The pilot runs on synthetic work. Testing on invented examples hides the edge cases that real work produces. Run the pilot on live cycles.
• Training is deferred to the end. People adopt their own workarounds in the gap, and the workflow arrives after the habits. Train as soon as the pilot has stabilised the process.

Where one of these has already happened, it is usually recoverable within the ninety days: restart the baseline, re-scope and re-pilot the narrowed output, move to live work, or bring the training forward.

Common mistakes

• Starting the pilot before the rules exist. Confidential data ends up in an unapproved tool.
• Skipping the baseline. The day-90 review becomes a debate about impressions.
• Training after the rollout. People invent their own approach in the gap.
• Scaling a workflow that has run once. One cycle is not evidence.
• Measuring usage only. Usage shows activity, not value.
• No decision at day 90. The program continues by default rather than by choice.

Frequently asked questions

What should an AI adoption plan cover in the first 90 days?

Assessment and one chosen use case with a named owner; the workflow design and data rule; two live pilot cycles; role-specific training; and a measured decision at day 90.

Is 90 days enough to adopt AI?

Enough for one contained, measured workflow. Business-wide adoption is a multi-quarter program built by repeating the same ninety-day shape.

What is the most important thing to get right first?

The owner and the baseline. Without a named owner nothing is scheduled, and without a baseline nothing can be proved.

Should we train before or after the pilot?

Training on the workflow happens after the pilot has found the problems, and before any scale. Training before the pilot teaches a process that is about to change.

How do we know if we should continue?

Compare usage, time, quality and risk against the baseline. Where the workflow is used, faster and no less accurate, continue; where a measure has not moved, understand why before extending the scope.

What if we miss the 90-day target?

Check whether the scope was contained. Slippage is almost always a scope problem rather than a speed problem, and the fix is to narrow the output rather than to add people.

Should the ninety days run alongside normal work?

Yes — the pilot runs on live cycles rather than a separate project, which is what makes the result credible. The cost is the owner’s time on design and review, not a parallel workstream.

Can several teams run the plan at once?

They can, but the first one should run alone. Running one workflow to a measured result teaches the organization the method; running three at once multiplies the coordination and dilutes the evidence.

Next step

Book the five phases into the calendar, name the owner this week, and capture the baseline before anything changes. See Building an AI Adoption Roadmap for the year-long view, or book an AI adoption call to run the ninety days with you.

Sources

• Plan structure reflects standard ninety-day program delivery applied to AI adoption: assess, design, pilot, train, review, with the baseline captured before change.

No statistic in this article is invented; where figures appear in the linked guides, they are cited there with their source and date.