Oluwaseun runs a forty-employee staffing agency in Houston that places administrative and light industrial workers. Last year he bought subscriptions to three AI tools after reading about them online: one for recruiting, one for writing job descriptions, and one he is still not entirely sure what it does. He spent about $4,800 over the course of the year. When his accountant asked what return those tools had produced, he went quiet. He knew the job description tool had saved his team some time. He had a vague sense the recruiting tool was useful. But he could not point to a single number that had changed as a result. He had spent $4,800 on intentions rather than outcomes. This year he wanted to decide in advance what AI he would invest in, why, and how he would know it was working.

Why Random AI Buying Does Not Work

Most small business owners discover AI tools the same way Oluwaseun did: a podcast, a newsletter, a recommendation from a peer. They buy the tool that sounds most compelling in the moment and add it to the stack. Sometimes it sticks. More often it joins a graveyard of subscriptions that get charged every month without being opened, and the charge is small enough individually that nobody has a reason to look at the total until an accountant asks.

The problem is not the tools. The problem is the absence of a plan. An AI investment roadmap is simply a written plan that states four things: here are the business problems I am trying to solve, here is the order I will tackle them, here is how much I am willing to spend per problem, and here is how I will know when a tool is working. Written down, in one place, before any money moves.

Building that plan does not require a consultant or a strategy workshop. It requires about three hours and honest answers to a few questions. The three hours are not the hard part either. The hard part is that the questions force you to name what your business is actually bad at, in numbers, which is uncomfortable in a way that browsing tool comparisons never is.

The mechanics of the graveyard are worth naming, because they are not carelessness. A subscription is bought with real intent, used for a fortnight, then displaced by a busier week. Nothing cancels it, because cancelling requires somebody to decide it has failed, and nobody owns that decision. The charge is too small to trigger a review on its own, and it recurs quietly alongside charges that are earning their keep. What Oluwaseun discovered was not that he had wasted money on bad tools. It was that he had no moment in the year at which the question was asked.

Step 1: Start with Problems, Not Tools

The most common mistake in AI investment is starting with a tool and working backward to find a use for it. Start instead with a list of your most painful, most time-consuming, or most expensive business problems. Write them as they occur in your operations, using the language your staff would use, rather than in the vocabulary of software categories.

Oluwaseun made a list of five:

1. Writing job postings takes his recruiters forty-five minutes each. They post thirty a week.
2. Candidate screening, meaning reading resumes and matching them to job requirements, takes two hours per open role.
3. Client check-in emails go out inconsistently. Some clients go two weeks without hearing from anyone.
4. Timesheet errors from workers cost about $800 a month in corrections and disputes.
5. Training new recruiters takes six weeks and relies on whoever has time to mentor them.

Notice that these are not "AI opportunities." They are business problems that happen to have AI-enabled solutions. That framing matters more than it appears to. A business problem has clear stakes and a natural owner, and it stays on the list whether or not a tool exists to address it. An "AI opportunity" is just a hypothesis wearing the clothes of a plan, and it disappears the moment the tool that inspired it turns out to be unsuitable.

Step 2: Size Each Problem

Estimate the annual cost or lost value of each problem before picking any solution. This sizing does not need to be precise; within 30% is fine. It just needs to be honest. The purpose is not accounting, it is ranking, and a rough figure ranks a list just as well as an exact one while taking a fraction of the time to produce.

For Oluwaseun's job postings, the sum is straightforward: 30 postings per week, at 45 minutes each, at a recruiter cost of $25 per hour, across 50 weeks, gives roughly $28,000 per year of recruiter time spent on job descriptions alone. Every input in that chain is a number he already had. That is a problem worth solving at meaningful cost, and knowing the figure changes what a reasonable price for a solution looks like.

Client check-in consistency is harder to quantify directly, and it shows how to size a problem when no timesheet captures it. He estimated that one client per quarter was not renewing because of communication gaps. At an average contract value of $15,000, that is $60,000 a year in preventable churn. The estimate rests on a judgement, which is why it is labelled as an estimate, but a labelled estimate that can be argued with is far more useful than an unmeasured intuition that cannot.

Sizing each problem lets you rank them by value. You will almost always find that two or three problems account for the majority of the cost. Start there. The remaining items stay on the list, sized, waiting for a later quarter, which is a very different thing from being forgotten.

Step 3: Pick One First Initiative

The instinct is to solve everything at once. Resist it. AI adoption in a small business works best when you go deep on one problem, get real results, build internal confidence, and then expand. Confidence is not a soft consideration here. The staff who have to change how they work are the same staff who watched the last three subscriptions arrive and achieve nothing.

Oluwaseun picked job postings first. The logic was clear. It was the largest provable time cost, and mature AI tools designed for exactly this purpose already existed, including one already sitting in his stack that he had barely used. He did not need to buy anything new. He needed to actually implement what he already owned, which is the cheapest first initiative available to most businesses in this position.

There is a lesson in which problem he did not pick. Client check-in gaps were sized at $60,000 a year, more than double the job posting figure, and he still started with job postings. The reason is that the posting figure was provable from records he already kept, while the churn figure rested on a judgement about renewals. For a first initiative, a provable number is worth more than a larger estimated one, because the whole point of the first initiative is to produce evidence that the method works. Start where the measurement is trustworthy, then spend that credibility on the harder problem.

He set a specific goal: reduce time per job posting from forty-five minutes to fifteen minutes, with equal or better posting quality as measured by candidate response rates. Note that the goal has a quality guard attached to it. Without one, a time-savings target can be met by producing worse postings faster, and the cost of that shows up somewhere the metric was not looking. He gave the initiative a ninety-day window.

Step 4: Set a Total Budget and a Per-Problem Limit

A simple rule for small businesses with fewer than fifty employees: total AI tool spend should not exceed 1% to 2% of annual revenue until you have proven ROI on at least two use cases. Before you have that proof, keep the number small enough that you can cancel anything that is not working without it hurting. The cap is not there because AI is expensive. It is there because cancelling is psychologically hard once a tool is embedded, and the easiest way to keep cancelling easy is to never let any single subscription get large.

Oluwaseun set three conditions of his own. No new AI tool would be added to the stack unless it could be tied to one of the five problems on his list, had a specific success metric defined before purchase, and cost under $200 per month. He then applied the same test backwards to what he already had, and cancelled the subscription he could not explain the purpose of. That alone saved $120 a month, and it made the point to his team more effectively than any policy would have: a tool that cannot be tied to a problem does not get to stay just because it is already being paid for.

Step 5: Review on a Quarterly Cadence

A roadmap is not a purchase list. It is a living plan. Every quarter, spend thirty minutes asking three questions about each active AI investment:

• Is the tool being used by the people it was intended for?
• Has the target metric moved?
• Is the tool worth its cost at the current level of use?

If the answer to any of those is no, you have an action rather than a verdict: increase adoption, adjust the use case, or cancel. The three questions are deliberately ordered, because a tool nobody uses cannot have moved a metric, and asking about cost before asking about adoption produces the wrong conclusion. A quarterly review is what prevents the graveyard from building back up, and thirty minutes is short enough that it actually happens.

Put the review dates in a calendar rather than in your intentions. A roadmap fails far more often through a review that never happens than through a wrong answer to one of the three questions, and the failure looks identical to success from the outside: nothing is cancelled, nothing is escalated, the tools continue. Scheduling the review at the moment you write the row costs nothing and is the single cheapest protection the roadmap has.

What Oluwaseun's Roadmap Looked Like

After three hours of work, his roadmap was a single spreadsheet with five rows, one per business problem. No consultant, no strategy deck, just a written plan that made each investment accountable to a specific result. Each row carried six columns.

At the ninety-day mark on his first initiative, time per job posting was down to seventeen minutes. That is short of the fifteen he had set, and worth recording as such rather than rounding into a success. Candidate response rates were up 12%, which meant the quality guard had held: the postings were faster and they were performing better, not faster and worse. He moved to his second priority, client check-in automation, with a sized problem already waiting for him on row three.

Anti-Patterns

• Buying before sizing. A subscription bought without an annual cost estimate for the problem it addresses cannot be judged later, because there is nothing to judge it against. This is how you end up spending a year on tools and going quiet when the accountant asks.
• Sizing only the problems that are easy to size. The job posting figure came straight from timesheet arithmetic. The churn figure required a judgement about renewals. Skipping the second kind means your roadmap systematically ignores the expensive problems that no system happens to log.
• Setting a success metric after purchase. A metric chosen once the tool is in use tends to be the metric the tool happens to move. Define it before you buy, when you are still describing the problem rather than defending the decision.
• Time-savings targets without a quality guard. Any drafting task can be made faster by accepting worse output. Pair the time target with a measure of whether the output still works, as the candidate response rate did here.
• Buying new when you already own the answer. Oluwaseun's first initiative used a tool that was already in his stack and barely used. Check what you are already paying for before adding a line to the bill.

Practice Prompts

Use these to draft the roadmap, then replace every figure the output produces with one from your own records. The AI is useful for structuring the plan and for asking you the right questions; it has no access to your numbers.

• Problem listing prompt: "I run a [type of business] with [number] employees doing [describe the work]. Ask me questions to help me list my five most painful, time-consuming, or expensive operational problems. Write each one in operational language, describing what happens and how often, not what software might fix it."
• Sizing prompt: "Here is a business problem: [describe it]. List the inputs I would need to estimate what it costs me per year. Ask me for each input one at a time and do not estimate any of them yourself. When I have supplied them all, show the calculation step by step so I can check it."
• Success metric prompt: "I am about to implement an AI tool to address this problem: [describe it]. Propose a success metric with a starting value and a target for me to fill in, plus a quality guard metric that would reveal if the improvement came at the cost of worse output."

Reflection

• What are you currently paying for that you could not tie to a specific business problem if your accountant asked today?
• Which of your five most expensive problems has no system logging it, and therefore has never been sized?
• For your most recent AI purchase, was the success metric defined before or after the money moved?
• Is there a tool already in your stack that would address your highest-value problem if somebody actually implemented it?

Glossary

• AI investment roadmap: a written plan naming the problems to solve, the order to tackle them, the spend allowed per problem, and how success will be recognised.
• Problem sizing: estimating the annual cost or lost value of a problem, accurate to within 30%, before any solution is selected.
• Success metric: the specific number a tool is bought to move, defined before purchase rather than after implementation.
• Quality guard: a second metric paired with an efficiency target to detect improvements achieved by accepting worse output.
• Per-problem limit: the maximum monthly spend allowed against a single sized problem, set so that cancelling stays painless.
• Quarterly review: a scheduled thirty-minute check of adoption, metric movement, and value for cost across every active AI investment.
• Subscription graveyard: the accumulated tools still being charged monthly that nobody opens, which a roadmap exists to prevent.

Related Lessons

• Aligning AI Strategy with Business Goals covers the one-page version of this thinking for a single goal, and is the natural precursor to a full roadmap.
• Cost Optimization: AI Subscription Budgeting goes deeper on the spend side, including how to audit a stack you have already accumulated.
• When AI Isn't Working: Recognizing Negative ROI is what to read when the quarterly review answers no to all three questions.
• Tracking Time Savings and Productivity Gains covers how to measure the kind of result the job posting initiative produced.
• Evaluating AI Vendors and Partnerships addresses the selection step once a problem has been sized and a budget limit set.

Closing

The difference between Oluwaseun's two years was not the amount of money involved. It was that the second year began with five sized problems, one chosen initiative, a metric defined before purchase, and a date in the calendar to check. Three hours of spreadsheet work bought him something the previous $4,800 had not: the ability to answer his accountant's question with a number, including the honest answer that the first initiative landed at seventeen minutes rather than fifteen.

Key Takeaways

• Start with your five most painful business problems, not with AI tools. Tools bought without a problem to solve become unused subscriptions that nobody has a reason to review.
• Size each problem in dollars before picking a solution. Accuracy within 30% is enough to rank problems by value and to avoid spending $200 a month against a $500 a year problem.
• Solve one problem at a time, deeply. Real results from one use case make the next one easier to fund and easier to get adopted by staff who have seen tools fail before.
• Set a success metric before you buy or implement any tool, and pair efficiency targets with a quality guard. "We will use it" is not a metric; "time per task drops from X to Y without response rates falling" is.
• Cap total AI spend at 1% to 2% of revenue until you have proven ROI on two use cases. Keep spend small enough that cancelling a tool that is not working stays painless.
• Review every active AI investment quarterly, in this order: is it being used, has the metric moved, is it worth its cost. Each no is an action rather than a verdict.
• Cancel subscriptions you cannot tie to a specific problem. If you cannot explain what business outcome a tool produces, it is probably not producing one.

Frequently Asked Questions

How precise does the sizing need to be?

Within 30% is stated as good enough, and the reason is that the figure is being used to rank problems against each other rather than to file accounts. A problem sized at roughly $28,000 and one sized at roughly $60,000 will sort the same way whether or not the estimates are exact. Precision beyond that costs time and changes no decision.

What if I cannot estimate a problem at all?

Look at how the client check-in problem was handled. No system logged it, so the estimate came from a judgement about how many renewals were being lost and what a contract is worth. Label it as an estimate, note the assumption, and let it be argued with at the next quarterly review. A labelled estimate can be corrected; an unmeasured intuition cannot.

Should I cancel a tool at the first quarterly review if the metric has not moved?

Not necessarily, and the order of the three questions is why. If the tool is not being used by the people it was meant for, the metric could not have moved and the action is adoption rather than cancellation. Cancel when the tool is genuinely in use and still not worth its cost, or when it cannot be tied to a problem on your list at all.

How does this differ from a one-page AI strategy?

A one-page strategy handles a single goal: one bottleneck, one capability, one ninety-day target. A roadmap handles the portfolio, adding the ordering between problems, the budget limits, and the review cadence. In practice you write the strategy for whichever initiative you pick first, and the roadmap is what tells you which one that should be.