90-day roadmap to adopt AI in your service SMB (with measurable ROI)

What does a 90-day AI plan that actually leaves a defensible return look like? It is three blocks of 30 days. The first one gets things in order: which problem, which starting numbers, who takes part. The second runs the trial and measures every week. The third closes with a three-way decision, scale, adjust or stop, and with the return calculated on time saved and errors avoided. The difference between a useful plan and a decorative one is that every milestone has a deliverable and a person by name.

Why 90 days and not a long program?

Because a large-company plan opens too many fronts for an SMB, and because 90 days is long enough to learn something and short enough that the cost of a mistake stays tolerable. A bounded cycle makes three things visible before committing to more: what was spent, what was produced and the decision that comes next.

This is for someone who has already decided to start and needs the plan in blocks. The use case changes by industry, proposals at an agency, document intake at an accounting firm, searching internal records at a law firm, but the three blocks are the same.

The 30-60-90 path supports progress in stages, early learning, and adjustment before scaling.

What happens in each 30-day block?

Each block has a single goal and ends with something you can show. What follows are the guidelines for each month, not a closed schedule: the fine detail comes from your own operation.

Days 1 to 30: get organized

The first month ends with the problem chosen, the starting numbers measured and the team defined. In practice that is four decisions. Which process hurts, with the pain put into numbers. Who backs the project from the leadership side and who runs it day to day. Which tools and access it will run on. And which criteria will define the day 90 decision, written before starting and not after. At some point in these four weeks you measure the starting point of the six indicators, and it closes with a kickoff session on real cases from the company, with a test team of three to five people.

If any of that is missing on day 30, you do not move on to day 31. S&P Global Market Intelligence reports that 42% of companies abandoned most of their AI initiatives in 2025, and this barrier exists precisely so that nobody moves forward without a foundation.

Days 31 to 60: run it and measure every week

The second month runs the trial under real conditions and measures it every week. The use happens inside the workflow that already exists, not in a separate exercise; the six indicators are taken every week; and adjustments to the instruction or to the workflow are made on what the measurement shows, not on impressions. At the end of the month a 30-minute meeting is enough to review what worked and what gets adjusted.

The number that matters most in this block is how much of what the AI produces gets used without correction. As an operating threshold for this plan, and not as a universal benchmark, if a person has to correct more than 40% of the output on a sustained basis, it is worth adjusting before putting in more hours.

Days 61 to 90: close with a decision

The third month ends with the decision made, the return calculated and the learning written down. You collect the same six indicators, compare them against the first month's starting point, calculate the return on time saved and errors avoided, and leadership chooses among the three exits: scale, adjust or stop. What was learned is documented in a way that lets another person reuse it without asking.

Closing the trial is not delivering a report. It is making the decision and putting a date on the next step.

How do you calculate the return?

Before the math, what ROI is. It stands for return on investment, and it answers a simple question: is what you gained from the change worth more than what it cost you to make it? If you invested 100 and gained 150, the return is 50%. If you invested 100 and gained 100, you came out even. Nothing more than that.

There are several ways to calculate it and none is mandatory. This is the most common one and it is usually enough to defend a 90-day trial:

Estimated return (%) = ((estimated benefit − total investment) ÷ total investment) × 100

The estimated benefit is the hours saved multiplied by what an hour of that role costs, plus any other benefit you can put a number on with evidence in the same period. The hours saved come from subtracting the process time before and after, multiplied by the measured volume. The total investment includes licenses, team hours, training and outside support if there was any.

An example, with made-up numbers and no currency, just to see the mechanics. An accounting firm puts together 20 proposals a month. Before the trial, each one took 3 hours: 60 hours a month. Afterward it takes 2 hours: 40 hours a month. That is 20 fewer hours per month, 60 for the quarter. If an hour of that role costs 25, the benefit for the quarter is 1,500. The investment was 400 in licenses plus 24 team hours spent preparing and reviewing, which at 25 an hour is 600: 1,000 in total. The math comes out to ((1,500 − 1,000) ÷ 1,000) × 100, that is, a 50% return. Your numbers come from the first month's measurement and the third month's close, not from an example.

Three cautions that make the difference between a defensible number and one that does not survive a single question. Do not mix periods: if the trial ran three months, calculate benefit and investment for those three months and present the annual projection separately, as a scenario. Subtract what the change cost, because training and team time are real investment. And do not slip in benefits without a number: the team being happier matters, and it does not belong in this calculation.

If the return comes out positive and the main indicator moved, you have enough to evaluate expanding it. If it comes out zero or negative and the indicator did not move, stopping is a result too: it saves you from scaling something that was not working.

Return becomes measurable when benefits and investment are evaluated using the same criteria.

Where should you put your attention?

On the people and the process, more than on the tool. BCG estimates that around 10% of the value comes from algorithms, 20% from technology and 70% from the human component. That split does not say where to spend the budget; it says where to look when the trial is not moving.

Applied to these 90 days, it turns into three very uneven fronts of attention. On the model side, it is enough to pick the minimum capability that solves the case and nothing more. On the technology and data side, the work is securing access, sufficient quality, integration with what you already use and the controls on use. And the bulk of the effort goes to the third front: designing the workflow, training whoever is going to use it, reviewing the output and making clear who answers for the result.

That is why a proposal that only talks about licenses leaves unresolved exactly the human and organizational conditions where BCG places most of the value.

How do you know your plan is no good?

There are four signals that are easy to recognize. It is too long: a two-hundred-line schedule that nobody is going to read or follow. It is a shopping list: a list of technologies to evaluate, under another name. It is vague: "days 30 to 60: run pilot" passes for a plan when it is really a heading. It is fixed: it never gets touched, when it should be adjusted on day 30 and on day 60 based on what was measured.

The plans that do reach day 90 look alike in three ways. They measure every week, without skipping any, even when the week goes badly. They actually make the day 90 decision, instead of leaving it at "let's keep exploring." And they write down what they learned, which over the months ends up being worth more than what the tool produced during the trial.

Do you want to build your own?

VegasiO's Discovery web takes 5 minutes and returns the plan adapted to your case, with the problem prioritized and the indicators that apply to you. At the end it tells you whether the next step is a Diagnóstico de Adopción IA, a guided Implementación de IA, or consolidating before moving ahead.