Everybody wants to know where to start with AI. Here’s one real answer, from a client, with the real number in it.
A hundred-million-dollar fruit processor. Two production lines. An ERP that hadn’t worked right for five years. They brought us in as a fractional CTO, and AI wasn’t on the list. Not even close.
So we didn’t start with AI. We started with the gaps between the software vendor and the people running the floor (that’s where most of the pain was hiding). We got the system working the way it was supposed to across every segment, and we got the right reports into the CEO’s hands. That took time. Two years in, we were also carrying systems administration and the rest of the corporate tech load.
Then we ran a four-week AI sprint on Claude Enterprise, and kept going.
What came out of it:
- A daily production planning tool across both lines and packaging
- Board-level financial reporting, plus the presentation deck that goes with it
- A quality control document management system
- A sugar resource planning tool, replacing four spreadsheets and the PLC interfaces (the controllers on the line) that someone used to stitch together by hand
It started small: the planning tool gave about two hours a week back to the person who used to build that plan by hand. Today production planning gives back about 17 hours a week: those two hours, plus about three hours a day.
But here’s the thing.
Seventeen hours back isn’t the point. Knowing what else the system can see is.
Why it worked
- The data was already clean, because we’d spent the time fixing the ERP first. AI on top of broken data just gets you a confident wrong answer faster.
- One person owns the tool. Building the plan was their job before and it’s still their job. The tool drafts, they decide.
- We started with a job that happens every day. Small savings on a daily task add up, and you find out fast whether it works. Then we moved to the bigger ones.
- We measured it. If it hadn’t given time back, we’d have switched it off.
What’s next
Point the same tools at a read-only copy of the database and see what else the data already knows. Four tools in, each one made the next one faster to build, because the data underneath was already in shape. That’s where I’d expect the bigger number to come from. We haven’t measured it yet, so I’m not going to print one.
Where AI fits
- Start with a job that happens every day, not a big one that happens once a year.
- Fix the data under it first. Every time.
- One named person owns the tool and checks what it produces.
- Read-only access first. Let it look before it touches anything.
- Measure it against the hours or errors it’s supposed to remove, and switch it off if it doesn’t.
List the jobs someone in your company does every day by copying information from one system into another, or by rebuilding the same spreadsheet. Next to each, write who does it and roughly how long it takes. That list is where your AI work starts.
If you want a second set of eyes on that list, book a 30-minute call and I’ll tell you which one I’d start with.
Elliot Swift, MurraySwift & Associates