“AI” is a category, not a use case. The businesses that get value out of it don't buy AI, they solve specific, often unglamorous problems with it. Here's a grounded tour of the places it tends to pay off in an established business, how to spot the ones that fit you, and the ones that quietly don't.
First, what makes a use case actually pay off
Before the list, the filter. After enough proposals, the use cases that return real money share four traits, and the ones that disappoint are usually missing one of them. If you only remember one thing from this piece, remember these four:
- Frequent. It happens many times a day, not a few times a quarter. Small time savings only add up when the task is constant.
- Bounded. The task has a clear shape and a clear right answer. “Categorize this invoice” is bounded; “run our strategy” is not.
- Forgiving. A wrong answer is cheap to catch and cheap to fix, because a person reviews it before it matters. The unforgiving cases (where a quiet mistake is expensive) come later, if at all.
- Already in your data. You can do it because the history or documents it needs already exist. If the data isn't there, that's a data project first, not an AI one.
Score any idea against those four. The best first projects are boring on purpose: high-volume, well-defined, low-stakes, and sitting on data you already have. The exciting, headline-grabbing ideas almost always fail at least two of the tests, which is exactly why they stall.
Answering routine questions
An assistant trained on your own documents and policies can handle the repetitive questions, whether from customers or from your own staff, around the clock, and hand the genuinely hard ones to a person. A distributor might point it at order status and returns policy; a services firm might point it at “how do we handle X” questions new hires ask for their first six months.
Done well, it doesn't replace your team, it gives them their attention back for the work that needs it. Where it goes wrong: pointing it at questions your own documents don't actually answer, or letting it respond on high-stakes topics without a clear handoff to a person. Keep it to what you've written down, and make the handoff obvious.
Getting through paperwork
Pulling the details out of invoices, forms, purchase orders, and contracts is exactly the kind of high-volume, rules-light work AI is good at. Less manual entry, fewer transcription errors, and hours back every week, with a person reviewing anything that matters before it posts.
This is one of the most reliable starting points for an operations-heavy business, because the volume is high and the task is bounded. The honest catch: your documents are messier than you think (scans, odd formats, a supplier who changes their invoice layout twice a year), so budget for the edge cases and keep a human in the loop on anything that touches money.
Turning your own history into decisions
Two flavours here, both valuable, both built on data you already keep:
- Forecasting: turning your sales, demand, or staffing history into clearer plans, so reorder and scheduling decisions rest on patterns instead of hunches.
- Catching problems early: flagging unusual transactions, errors, or risks the moment they appear, before they become expensive.
These pay off most where the cost of being wrong is real and recurring: stockouts on a fast mover, cash tied up in a slow one, an error that snowballs if nobody notices for a week. They need cleaner history than the paperwork use cases do, so they often come second, after a Pilot has shown the team what good looks like.
Finding answers buried in your documents
Most established businesses sit on years of documents nobody can search well: contracts, SOPs, past projects, technical manuals. AI can give a direct answer, with its sources, in seconds, instead of someone digging through folders or interrupting the one person who remembers. It's quietly one of the highest-value, lowest-risk places to start, because it is frequent, bounded, forgiving (you can see the source), and built on documents you already have.
Drafting the first version
Reports, proposals, summaries, replies: a competent first draft in your voice beats a blank page, as long as a person still does the final pass. The win is speed to a good starting point, not hands-off automation. Treat anything it writes as a draft from a fast junior colleague, useful, and never the last word.
Where AI quietly doesn't pay off (yet)
The honest half of this list. Plenty of ideas sound like great AI use cases and aren't, and a good partner will talk you out of them. Be skeptical when an idea looks like one of these:
- Rare, high-judgment decisions. The big annual call that happens twice a year and depends on context only your people hold. Low frequency, high stakes, hard to check. Leave it to the humans.
- Anything unforgiving and unreviewable. If a wrong answer goes straight to a customer or a regulator with no human in between, the risk usually outweighs the saving until you've earned a lot of trust in the system.
- Tasks with no usable data. If the information lives only in people's heads or in formats nothing can read, you have a data problem to solve first. AI on top of bad or missing data just produces confident nonsense faster.
- “Replace the whole team” fantasies. The pitch that one tool will run a department. Real value comes from taking the repetitive 20% off many people's plates, not the whole job off one.
Saying no to these isn't caution for its own sake. It's what keeps your first AI project from becoming the cautionary tale that sets you back two years.
One business, several of these at once
These rarely live in isolation. In our one hands-on engagement so far, an online retailer used AI in three of these areas at once: answering routine customer questions, speeding up product and catalog work, and grounding stock decisions in data. That is a good reminder that the question isn't “which one AI thing should we do” but “where, across the business, does it pay off first, and what's the sensible order.”
The bottom line
Don't shop for “AI.” Look at where time and money actually leak in your business (routine questions, paperwork, planning, searching, drafting), run each candidate through the four tests, and match a tool to the leak. The boring, measurable use case is almost always the right place to begin, and the quickest way to find yours is to look at what your team complains about doing every single day.