How to choose your first AI use case

Why the first use case matters so much

The first AI project sets expectations inside the business. If it works, staff are more willing to try the next one and owners are more willing to fund it. If it fails noisily, perhaps with an embarrassing email to a client, it can set adoption back a year.

So the first choice should favor learning and confidence over ambition. You want something that will visibly help within a few weeks, where mistakes are easy to catch, and where the people involved are keen rather than reluctant.

A good first project also teaches the business how to run AI projects: how to baseline, how to review AI output, how to collect corrections and how to make a keep-or-stop decision. Those habits matter more over time than the first saving.

Building a candidate list

Start by collecting candidates from the people who do the work. Ask each team lead a simple question: which recurring task takes the most time, causes the most delay, or creates the most rework? You will usually get a list like this:

Do not filter the list yet. Ideas that look unsuitable for AI often reveal a process problem worth fixing anyway. Add a short note next to each with who does it today, how often, and roughly how long it takes.

• Replying to new enquiries quickly, especially after hours.
• Turning sales call notes into a proposal or quote.
• Chasing clients for documents during onboarding or tax season.
• Retyping information from PDFs, forms and emails into a system.
• Following up on open quotes and unpaid invoices.
• Assembling recurring client reports from several sources.
• Answering the same internal questions about procedures and policies.

A simple scoring method

Score each candidate from 1 to 5 on six factors. Add them up. The point is not precision but forcing a structured conversation.

Weight the factors if your situation calls for it. A firm under financial pressure might double the value score. A regulated practice might double the low-risk score. Agree the weighting before you score, so it is not adjusted to favor someone's preferred idea.

• Value: how much time, revenue or margin is at stake if this improves?
• Frequency: does it happen daily, weekly or monthly? Daily scores highest.
• Measurability: can you time it or count it today, before any change?
• Data readiness: are the inputs already in a system and reasonably consistent?
• Low risk: if the AI gets something wrong, will a person catch it before it matters?
• Ownership: is there one person who wants this to work and has time to test it?

Strong first choices, and why

Some use cases score well for most service businesses. They are frequent, the AI prepares rather than decides, and the result is easy to check.

Drafting follow-up emails from call notes is a good example. It happens daily, the person who made the call reviews the draft, and the time saving is easy to see. Document request checklists for client onboarding work similarly: AI drafts a tailored list from last year's file or the engagement type, and a staff member confirms it. Summarizing long maintenance or support threads for a coordinator, or drafting monthly report commentary for an account manager to edit, are other common strong picks.

Notice what these share. A person who already does the work stays in control, the input is text they already have, and a bad output costs a few minutes rather than a client.

Another reliable pick is internal: drafting first versions of job descriptions, procedure updates or training checklists for a manager to edit. Nothing reaches a client, the reviewer knows the subject well, and the output is easy to judge. It is a gentle way for a skeptical team to see what the tools can and cannot do.

Weak first choices, and why

Some ideas are worth doing eventually but make poor first projects:

These are not bad ideas. They are better tackled once the team knows how to review AI output, the data is tidier and you have a governance routine in place.

• Anything that scores, ranks or screens applicants, tenants or borrowers. These carry fairness and legal questions, including Colorado's SB 26-189 from January 1, 2027.
• Fully automated client-facing replies with no review, especially on pricing or advice.
• Projects that need three or more systems integrated before anything works.
• Workflows that happen a few times a year, such as an annual budget process.
• Anything where the data lives mostly in people's heads or personal inboxes.
• A firm-wide chatbot over all your documents before permissions and file structure are tidy.

An illustrative scoring example

Imagine a 25-person insurance agency weighing three candidates. The scores are illustrative.

Renewal reminder emails score high on frequency, data readiness and low risk, because policy dates are already in the agency management system and an account manager approves each message. Value is moderate. Total: 25 out of 30. Quote comparison summaries for clients score high on value but lower on risk, because errors in coverage descriptions matter, and data arrives in inconsistent carrier PDFs. Total: 19. An AI chatbot on the website answering coverage questions scores lowest: high risk of a wrong answer, unclear ownership and hard to measure. Total: 13.

The agency starts with renewal reminders, learns how its team reviews AI drafts, and revisits quote summaries six months later with better templates and cleaner data.

From choice to plan

Once you have a winner, write it up as a pilot: the problem in one sentence, the baseline, the acceptance criteria, the review step and the owner. Then resist the temptation to add a second use case before the first is finished.

If the top-scoring idea turns out to be fixable with a settings change or a template, do that and move to the next one. That still counts as progress. Picking the first use case is also the core of InTheory's first 30 days with a client: rank the opportunities, agree one pilot, and set a baseline.

Write down the runners-up and their scores too. When the first pilot ends, you will already have a ranked list to return to, and you can rescore it with what you learned about your data and your team's review habits.