How to find your first AI use case (and how to tell whether it pays off)

Your first artificial intelligence (AI) use case should be a task your business already does many times a week, that follows clear rules but still needs some judgment, and where slow responses, errors or handoffs cost you money you can count. Good candidates look ordinary: taking in and scheduling new requests, following up on quotes, processing incoming documents, or putting together a recurring report.

To find yours, list the candidates, measure your own baseline for each (volume, time per case, error rate, revenue at stake), check that the data and systems are in place, and pick the one with the best mix of value and feasibility that you can test in a few weeks and undo if it fails. Below is that method, with a scoring table you can use.

If you have not started yet, you are not behind. The U.S. Census Bureau's Business Trends and Outlook Survey (BTOS), a recurring survey of U.S. businesses, found that AI use among businesses stayed between 17% and 20% from December 2025 to May 2026, and was higher among larger firms (37% of those with 250 or more employees). The question is not whether to use AI but where to start.

1. 01LookRepetitive work where delays and errors cost money
2. 02SizeVolume, time per case, errors, revenue at stake
3. 03CheckData, systems, risk if wrong, who reviews
4. 04PickValuable, feasible, reversible, measurable in weeks

MeasureAgainst the baseline, then choose the next one

Where to look: repetitive work where delays and errors cost money

We always start with the work, not the technology. The best early candidates share three traits.

They repeat at volume. The task happens daily or weekly, many times, in roughly the same way. A twice-a-year task will not pay back the setup.

They mix rules with judgment. Pure rules (copy this field into that system) are a job for ordinary automation, not AI. Pure judgment (negotiating a partnership) stays with people. The best fit is mostly routine work that requires reading something messy, such as an email, a scanned form or a voicemail, and deciding what it means.

Delays, errors or handoffs cost money. Look for places where a lead waits hours for a reply, where someone retypes information between systems, or where mistakes cause rework, refunds or compliance exposure.

To surface them, ask the people who do the work: "What do you do every week that you wish you never had to do again?" and "Where does work sit waiting for someone?" Their answers beat any vendor demo.

Typical patterns that fit (common shapes of work, not results from any particular company):

• Intake and scheduling. Requests arrive by web form, email and phone. AI can read each one, collect missing details, propose times and hand a clean record to a person.
• Follow-up on quotes. Follow-up depends on who remembers. AI can draft a timely message from the quote and conversation history, and flag replies that need a human.
• Document handling. Invoices, applications or forms arrive in different formats and get keyed in by hand. AI can extract the fields, check them against your rules and route only the exceptions to staff.
• Reporting. Someone spends hours each week pulling numbers from several tools into a summary. AI can assemble the draft and point out what changed, and a manager reviews it.

How to size the value with your own numbers

Skip the industry benchmarks. The numbers that matter are yours, and you can collect them in a week or two by tallying what already happens. For each candidate, measure:

1. Volume: cases per week or month.
2. Time per case: staff minutes from start to finish, including waiting and handoffs.
3. Error or rework rate: how often a case is corrected, redone or escalated.
4. Revenue at stake: the leads, quotes or renewals that pass through this step, their typical value, and how many are lost to slow response or no follow-up.

Then do simple math. Monthly hours equal monthly volume times time per case; multiply by your loaded hourly cost (wages plus benefits and overhead) for the cost side. For revenue, multiply the opportunities that stall at this step by your usual close rate and average deal value. For risk, write down what a single error costs you.

That maps to the three reasons a business invests: grow revenue, reduce costs, reduce risk. A good first use case visibly moves at least one of them.

Use conservative ranges. If you cannot measure a candidate at all, that is a finding too: track the work before you automate it.

How to check feasibility before you commit

A valuable idea can still be a poor first project. Four questions sort that out.

Is the data available? Past emails, forms, price lists, policies, records. Are they digital, reasonably consistent and in one or two places, or scattered across inboxes and paper?

Which systems does it have to touch? List the tools it must read from or write to: your CRM (customer relationship management system, where contacts and deals live), calendar, accounting software, email. Check whether each offers an API (application programming interface, the standard way one piece of software exchanges data with another). If a key system has no way in, the project gets slower and more fragile.

What happens if it is wrong? A misfiled document is cheap to fix. A wrong price sent to a customer, or a wrong answer to a legal, medical or financial question, is not. The higher the cost of an error, the more review you need and the worse it fits as a first project.

Who reviews the output? Name the role that checks results, at what point, and how much of their time it takes. If no one owns the review, the project stalls or, worse, runs unchecked. For a more formal structure later, the National Institute of Standards and Technology (NIST), the federal standards agency, publishes a free, voluntary AI Risk Management Framework.

How to pick the first one: a simple scoring table

Score each of your three to six candidates from 1 (low) to 5 (high). This is the same logic we apply when we assess candidates ourselves.

Rank candidates by value times feasibility. Then drop anything that scores below 3 on reversibility or time to proof, however attractive it looks. Your first project has to be safe to try and quick to judge.

Before you start, write down the number you expect to move, where it stands today, and when you will check it. Run the pilot with a person reviewing every output, then compare against your baseline.

What not to start with

We advise against these as first steps, though some make sense later.

• A company-wide chatbot or "AI platform." No single owner, no baseline, no obvious number to move.
• High-stakes answers to customers. Pricing commitments, or legal, medical or financial guidance. Start internally, or with a person approving every response.
• Broken or undocumented processes. AI speeds up whatever is there, including the confusion. Fix the process first.
• Projects that need data you do not have yet. If the history lives in personal spreadsheets or people's heads, collecting it is the real first project.
• Buying a tool before choosing a problem. Licenses are easy to buy and hard to justify when nobody can say which task they improve.

What this means for your business

Your first AI use case is less about AI than about choosing work you already understand, can measure and care about. Pick one, measure it, run it for a few weeks with review, and compare.

If it pays, you have evidence and a repeatable way to choose the next one. If not, you spent little, you can switch it off, and you know where your data and processes need work. Either outcome beats a large project nobody can evaluate.

If you want a second opinion on where to start, book a meeting with us.