AI Audit

What an AI audit actually involves

Marcus Olsson, Founder, Lemonstone AI

·7 min read

An AI audit is a structured look at how a business actually runs, done before anything gets built, to find where AI would create the most value. Done properly it produces a diagnosis and a ranked recommendation rather than a proposal. Most vendors skip it, because skipping it closes deals faster.

Why the audit gets skipped

Selling AI is hard, and an audit makes it harder. It puts weeks between the first conversation and any talk of money, and it occasionally ends with the honest answer that there isn't much worth building yet. That last outcome is a fair answer to the readiness question rather than a failed sale.

So plenty of vendors skip it. The conversation runs straight from what do you do to here's what we can build. Thirty minutes, no diagnosis, then a proposal.

You wouldn't accept that from a doctor. Nobody lets a surgeon operate after a ten minute chat in which they examined nothing. In AI it's close to standard, and owners go along with it because they assume the vendor knows something they don't.

Usually the vendor is guessing. The guess is what you're paying for.

Three questions to put to any vendor

Where does it actually hurt, and how do you know?

A doctor doesn't take your word that the knee is the problem. They make you stand up, walk, point at it. They look for themselves, because the place that hurts and the place that's broken often aren't the same place.

The equivalent is a vendor who sits with your team and watches the work happen before recommending anything. Not a questionnaire. Not a workshop where you list your own problems. Watching the actual work, because the friction costing you most is usually the friction you stopped noticing years ago.

Anyone prescribing AI before observing how you work is guessing, in a proposal's clothes.

What's the procedure, and what does the Monday after look like?

A surgeon can tell you what they're going to do, how long it takes, and what your life looks like a month later.

A vendor should manage the same. Which screen your team opens. What's different about their Tuesday. Which hour comes back, and to whom. If you can't picture the specific Monday morning after this thing goes live, you've been given a category rather than a procedure.

How many of these have you done?

You wouldn't want to be your surgeon's first knee replacement.

Ask for one specific project. Not a category, not "companies like yours". An actual build with a shape, a timeline and a result. If nobody can name one, that's your answer.

Those three together are the audit. The exam, the plan, and the track record. By the end of one you should know whether to carry on, and a good audit is just as willing to tell you not to.

That's the shape of the audit we run, and what follows from it depends entirely on what the exam turns up.

The question that opens it up

Owners almost always arrive with a list. AI for sales. AI for support. Something they saw a competitor announce. It's rarely a bad list, but more often than not the thing worth building first isn't on it.

People describe what they've heard about rather than what's costing them. So the exam runs on questions built to get past that.

The one that works best is the simplest.

If you could wave a magic wand and make one problem in your business disappear tomorrow, which one would it be?

Don't overthink the answer. Whatever comes to mind first is the one hurting most, and the fact that it surfaced instantly is the signal.

I've watched that question do more work than any pitch I've given. In one meeting I stopped presenting about twenty minutes in and asked it. The broker thought for roughly three seconds, then talked for fifteen minutes without stopping. At the end of it he said, almost to himself, that if he could just have something that listened to those calls, sorted his clients into boxes and pinged him when a new unit matched, he'd have done it years ago. He wrote the spec. I didn't.

I'd been spending hours sharpening pitches. That time should have gone into sharpening questions.

A second version works when the first one stalls.

What are you doing manually right now that a computer should be doing for you?

The answer is almost always something boring. Invoice data typed in every Friday. Numbers copied between five tools just to see where the month stands. The same report rebuilt from scratch every week.

And one more, this one worth asking the team rather than the owner.

What's the shortcut you use every day that you wish you didn't need?

Don't ask people what they want automated. That question makes them freeze. Ask what they've already built a workaround for. A complicated formula somebody wrote themselves, or three keystrokes that shuttle data between two tools, isn't a sign of a skilled employee. It's proof that the process is broken and they've been quietly absorbing it.

The three patterns that show up every time

After enough of these, the same three keep appearing regardless of industry.

Same data, typed twice. Somebody copies information out of one tool and into another. Every day, sometimes every hour. A lead from the website typed into the CRM. A closed deal typed into the bookkeeping system, then typed again into whatever tracks what's still owed on it. The person doing it has usually got so good they can do it without thinking, and that's the giveaway. Work a human can do without thinking is work a computer should be doing instead.

The handoff that runs through email. Two systems that ought to talk to each other don't, so a person bridges the gap. Forwards a PDF. Copies a number across. Updates a status because of a message somebody sent. It looks like work and it feels like work. It's plumbing that was never connected.

The report nobody reads. Someone spends ninety minutes a week compiling numbers into a document and sends it to people who scan it for thirty seconds and go back to their inbox. Or it's a dashboard built once that nobody trusts anymore, so everyone quietly returns to the spreadsheet they had before. The report isn't the problem. The data not already being where decisions get made is the problem.

Worth saying plainly. Two of those three often don't need AI at all. They need two systems connected, or one screen that shows the truth. I ran an audit for a tattoo studio owner who came in asking about AI, and what he actually needed was a dashboard he could update in ten seconds a day so he'd finally know how the month was going. That was the entire build.

An audit that always concludes you need AI isn't an audit. It's a sales process with a longer run-up.

Why it ends with one build, not twenty

When we started doing this we made the obvious mistake. We audited the whole company. Every process, every bottleneck, every opportunity. Then walked out with a list of twenty things to build, all at the same time.

It sounded ambitious. It was a bad idea.

Each of those problems needs its own attention, and most of them are tangled up in each other. Behind every one there are people who have to adapt to working a different way. No team absorbs twenty changes at once. It overwhelms them, and it overwhelms whoever is building it.

So the audit now ends somewhere much narrower. One problem, chosen against two tests.

Can AI handle this with confidence, and would solving it make a real difference to the business? If both are true, that's the first build, and nothing else gets started. Each one after that stacks on the last, which is how a business ends up with something closer to an operating system than a pile of tools.

That buys three things.

  • All the attention goes into one build, done properly.
  • The team adapts to one change instead of twenty.
  • The most painful problem goes first, so the value lands in weeks rather than after a three month transformation.

The real reason though is trust. When two companies have never worked together, "let's transform your whole business" is an enormous ask. They don't know you and you don't know them. Maybe the fit isn't right, maybe the way each side works doesn't line up. Better to find that out on something small than three months into something locked.

And when the first thing works, the conversation changes completely. The trust is sitting on something delivered rather than promised, and you stop having to explain what's possible, because they've seen it happen inside their own business.

The next thought usually arrives on its own. What else have we been carrying for years that we don't actually have to?

That list is normally the good one. It's also the list no audit could have produced on day one, because nobody knew to ask.

Frequently asked questions

What is an AI audit?
A structured look at how a business actually runs, carried out before anything is built, to find where AI would create the most value. It means observing the real work rather than collecting a wish list, and it should produce a diagnosis and a ranked recommendation rather than a proposal. A good one is also willing to conclude that little is worth building yet.
What should an AI vendor do before proposing anything?
Three things, borrowed from how you'd judge a surgeon. Examine the work directly by sitting with the team and watching how the job gets done. Describe the specific procedure, including what changes for staff on the Monday after it goes live. Name an actual comparable project with a shape, a timeline and a result. A recommendation made without the first of those is a guess.
How do you decide which process to automate first?
Two tests. Can AI handle it with confidence, and would solving it make a real difference to the business. Beyond that, the strongest signals come from asking the owner which single problem they'd remove tomorrow if they could, and asking the team which daily shortcut they wish they didn't need, because an existing workaround is proof that a process is broken.
Does every AI audit end with an AI recommendation?
It shouldn't. Two of the three patterns that turn up most often, the same data typed into two systems and the handoff that runs through email, are usually fixed by connecting tools rather than by adding AI. Sometimes the answer is one dashboard that shows the truth. An audit that always concludes AI is needed is working as a sales process.

The time to move is now.

AI is moving fast, and it's already changing how businesses like yours run.
You don't need to figure it out alone. One call is enough to see where it could fit your company.