When a leadership team asks me for an AI roadmap, I want to understand the work before discussing the tools. Where does the company wait? Which tasks require employees to move information between systems? Which decisions arrive late because nobody trusts the underlying data?
Those questions produce a more useful starting point than a list of AI products. A mid-market company has limited management attention, uneven data, and a business that must keep running while changes are made. The roadmap needs to respect all three.
Here is how I would structure the first 90 days. The dates are a planning framework, subject to the complexity of the company and the selected workflow.
Days 1-30: understand the work and choose a problem
Start with the people doing the work. Watch how a quote, support request, account update, or management report actually gets completed. Include the steps that never appear in the process diagram: opening a spreadsheet, asking a colleague, correcting an inconsistent record, or checking the output twice.
Build a small inventory of candidate workflows. For each, record volume, time spent, error consequences, data access, current owner, and the measure that would indicate improvement. Find out whether the problem could be solved with an existing system feature or a straightforward process change. AI should earn its place in the solution.
I would prioritize one bounded workflow with a meaningful operating benefit, accessible data, and an owner willing to change the process. A lower-risk internal task is often a sensible first test. Customer-facing commitments and irreversible financial actions require a stronger control design.
By day 30, the team should have a baseline, a named owner, approved data handling, and a written test plan. It should also know which ideas are being deferred and why.
Days 31-60: test against real cases
Build the smallest version that can answer the business question. A pilot does not need to solve every exception, but it does need to encounter representative ones.
Use a test set drawn from the actual workflow, with sensitive information handled appropriately. Include missing fields, conflicting information, and cases where the right answer is to escalate. Compare the proposed process with the current one using the same cases.
Measure the total work. If AI drafts an answer in seconds but someone spends ten minutes validating it, those ten minutes belong in the result. Set expectations for accuracy, turnaround, review effort, and acceptable failure before looking at the pilot results.
Employees need a simple way to report a problem and continue through the existing process. A pilot without a usable fallback can consume more attention than it saves.
Days 61-90: make an operating decision
The final month should produce a decision to expand, revise, or stop. A promising demonstration earns further testing; a repeatable business result earns a broader deployment.
If the workflow is ready, document how it will run: who owns it, who reviews exceptions, who can change the configuration, and what happens during an outage. Train the team around the revised process. Confirm that the data permissions and ongoing costs remain appropriate at the intended volume.
Then decide how to use the released capacity. That decision might involve service levels, growth, an avoided hire, or a reduction in outside spend. Keep the benefit category explicit and have finance validate the measurement.
What I would put in the board update
A useful update fits on a page: the business problem, baseline, test results, full cost, remaining risks, operating owner, and next decision. Show realized benefits separately from estimates. Explain what has changed since the prior review.
For a PE-backed business, connect the work to the value creation plan. If the initiative does not support a clear commercial or operating priority, ask whether it deserves leadership attention this quarter.
At the end of 90 days, I would rather see one workflow running reliably, with evidence of its value, than ten pilots waiting for somebody to decide what happens next. That first operating result creates a useful foundation for the next investment.
Related reading: Measuring AI ROI and Diagnosing operating constraints.
For help choosing the first workflow and building an accountable plan, explore an operating diagnostic.