A couple of weeks ago I spent 90 minutes in a room with twelve food and drink founders and operators. Brands you'd know from the shelves at Waitrose, Tesco and your local farm shop.
Before we started I asked what they were already using AI for. Writing social captions. Spotting patterns in a SKU spreadsheet. Reconciling accounting and inventory numbers. A handful of things each, most of them that they had come up with going about their work every day. Which is a great way to start identifying AI use cases… but…
By the end they had a list of more than a hundred.
That gap wasn't created by lack of effort. Everyone in that room was already using AI and already getting something out of it. They just had no way of seeing the rest of what was sitting in their own business.
After 200-odd conversations with founders and operators, I keep seeing the same four mistakes behind it. I know they're mistakes because I keep seeing the symptoms:
Someone tries to automate a job, spends longer fighting the tool than the job would have taken by hand, and gives up
A team finds a handful of use cases and leaves the rest of the business untouched
Something's clearly working but nobody can make the case for it, because the only number anyone counts is hours saved
AI gets something wrong and nobody notices until it's expensive
Each symptom has a mistake behind it. Here they are, in the same order.
Mistake one: starting with the tool, not the job
Someone on the team finds Lovable, or the founder mandates Claude Cowork, and the question becomes "right, what can I do with this?"
That's backwards. It's like finding a hacksaw in the kitchen drawer and deciding to cut your steak with it. It does the job. Badly, slowly, and with a lot more effort than a steak knife would have taken.
Nothing was wrong with the steak. You picked the wrong tool, then blamed the job when it went slowly.
Mistake two: only looking where AI is obvious
Captions, product descriptions, first drafts of buyer emails. That's where nearly everyone lands, because it's the most visible thing AI does.
Those are the bread rolls at the buffet. There's a whole spread and you've grabbed the first thing you saw.
It happens because there's no systematic way to go looking. Without one you'll only ever find what's already in front of you, which means the obvious stuff, which means content.
Mistake three: no structured approach to measuring value
Hours saved is the easiest number to reach for, and it's the one that kills the most good projects. There are four other kinds of value and none of them show up on a stopwatch.
Enhanced capability. Things you couldn't do at all before. A team of four hasn't got budget for an analyst or a copywriter, and this gets you a rough version of one. It needs checking every time, but it's still more than you had before.
Improved consistency. Every new line form filled in the same way. Every order processed the same way. No more "somebody forgot to update the price list."
Refined thinking. Better buyer meeting prep, because you stress-tested your own assumptions against something that argued back.
Reduced risk. Catching things before they cost you. One brand lost £4,000 because they took a retailer's form POs rather than checking it against their own agreed pricing, and nobody had twenty minutes spare to look.
None of those show up in a time-saved calculation. Which is how you end up with AI projects that look like failures on a spreadsheet and successes in the actual business.
Mistake four: only weighing one risk
When I ask founders what worries them about AI, nearly every answer is "data".
Data risk is real and worth thinking about properly. What you're feeding an LLM, whether the training controls are switched off, where customer data ends up.
But only weighing data risk is like watching your step for spilt milk while the kitchen's on fire behind you.
Two other risks usually matter a lot more for food and drink brands:
Consequence risk. What actually happens if it's wrong? A clumsy caption is embarrassing. A wrong allergen statement is a recall. An incorrect sales forecast means wastage or a stock-out. Same technology, wildly different tolerance for error in the use cases.
Capability risk. If you're using AI to do something you don't know how to do yourself, how would you know when the output is wrong? Get Claude to build you a financial model and you've got a choice: sense-check the assumptions, or trust it. Most people trust it.
What these four add up to
Every one of them is invisible from the inside. Nobody knows they're looking 10% of the opportunity, because there's nothing telling them the other 90% exist.
That's why the twelve people at the workshop were surprised by their own list. None of it was new information. It was their own business, sorted properly for the first time.
So over the next four issues I'm taking one mistake at a time and giving each one the actual fix, rather than a paragraph telling you a fix exists.
One thing to do this week
Don't go shopping for a tool. Open a note and write down every job you or your team currently use AI for. All of them, including the small ones nobody would call a use case.
Next issue I'll show you how twelve founders got from a number like that to more than a hundred, and how to run the same thing on your own business in an afternoon.
Sanjay co-founded Pesto, AI-native ops software for UK food and drink.