When a business owner says “admin is a nightmare,” that’s a starting point, not a service you can sell yet. To find a real AI opportunity, pick one specific task that keeps eating people’s time. Then find out how they cope with it today, and whether fixing it matters enough for them to act.

Your own experience helps you spot a messy process. But the next step is learning how it actually works right now. AI can help you organize notes you’re allowed to use and suggest questions. The people doing the work still have to supply the facts.

Pick a role you understand and could actually reach, like an office administrator, a trades business owner, or a small-team manager. “Small business” is way too broad. A specific role gives you a clearer task, the right lingo, and an easier way into a conversation.

Watch for work that gets copied, chased, fixed, or re-explained over and over. Think of pulling a quote together from scattered info, answering the same onboarding questions, or handing off a job without a full checklist. These are just examples. Not every business has these problems, and not every one wants an AI fix.

Jot down a short list of candidate tasks. Note what you saw yourself, what someone told you, and what you’re only guessing. Leave how often it happens, how much effort it takes, and what it costs as “unknown” until someone shows you. A task can be annoying without being worth real money.

The workaround usually tells you more than the complaint. Is someone keeping a private spreadsheet because the main system can’t give them what they need? Do they keep chasing people for missing info? Does a senior person double-check everything because mistakes are expensive?

Also ask what happens before and after the task. A slow quote might come from missing customer details, not slow writing. Fancy polished text won’t fix that. A better intake checklist might, with AI helping only where you can check its accuracy.

You’re investigating a process, not hunting for a spot to drop in a chatbot. Sometimes the best answer is a simpler form, a better template, or clearer ownership. If AI creates more checking work than it saves, it’s the wrong tool.

Ask an open question about the most recent time the task came up. Then walk through:

Skip leading questions like “Would an AI assistant save you hours?” That plants a solution and a benefit in their head before you understand the problem. Nielsen Norman Group suggests asking about specific experiences, then following up for detail.

Keep in mind that interviews only capture what people say they do. Memories have gaps, and people sometimes give the answer that sounds nice. If you have permission, compare their story to a non-confidential sample, or watch the task happen. Never ask for customer records, and never put confidential material into an AI tool without approval.

Anonymize any notes you’re allowed to use. Strip out names and sensitive business details, and give each note an ID so you can trace AI output back to its source.

Here’s a prompt you can try: “Using only these notes, group recurring task problems. For each group, identify the affected role, task, reported or observed friction, current workaround and supporting note IDs. Separate observations, reported experiences and assumptions. Mark missing frequency, effort, buyer and priority unknown. Suggest questions for the gaps. Do not invent costs, quotations, demand or earnings.”

Check every group yourself. If two people are describing different tasks, don’t lump them together just because the wording sounds alike. Treat the output as a research helper. This prompt is an APW suggestion, not a proven scoring method.

The person doing the task and the person who approves the spending might not be the same. Ask who owns the process, who would sign off on a change, and what limits apply. Existing tools, security rules, and other priorities can make an “easy” fix a bad fit.

The U.S. Small Business Administration recommends researching demand, alternatives, and pricing, including talking directly to customers. So don’t stop at “they dislike this task.” Find out what other options exist and why they do it the current way.

A friendly chat isn’t an order. One excited person doesn’t prove there’s a market. If you can’t reach a decision maker, can’t verify accuracy, or the fix is bigger than you can handle, set it aside.

Once several independent examples point to the same problem, describe one narrow deliverable. For example, an intake checklist plus a reusable quote template built from approved info. Spell out what goes in, what’s left out, who reviews it, and how many revisions are included.

Before you set a price, estimate your time for discovery, tool costs, building it, selling it, and quality checks. A paid pilot can show whether someone will actually commit, but one purchase doesn’t prove there’s steady demand. Afterward, compare the real workload and results to the problem you started with.

The best starting point isn’t the loudest complaint. It’s a specific task, believable evidence, a buyer you can reach, and a small fix you can deliver responsibly. Find the friction first. Let AI help you organize what you learn. Then let real decisions tell you whether it’s worth pursuing.