The last three issues were about work that quietly paid off, or AI that quietly didn't. This one breaks the pattern: it's the first time in this newsletter that AI just plain worked.

I built it on a plane. Six hours round trip to Austin, Texas, for an unrelated business trip. Three hours out, three hours back — and by the time we landed the shape of the thing was done.

Every month, someone brought in a USB drive full of competitive pricing research they'd spent the prior weeks gathering by hand. That was one input. A pricing coordinator was separately tracking individual parts. The sales team was building its own reports and spreadsheets on top of that, drawn from dealer visits, customer feedback, and their own gut sense of the market.

Different people, different methods, feeding one decision: whether a part's price needed to go up or down. None of it was wrong. It was just slow, it didn’t refresh often, in a business where pricing needs to move faster than that. And that’s also to say “Who was right?”

I got bored on the plane and had internet (never a good combination) and started with a spreadsheet: all 11,000+ active parts, what they cost us, what we charge dealers for it. I walked through that spreadsheet with Claude and set it up to use Firecrawl (a tool that lets an AI read a live web page directly, the way you or I would), instead of guessing from old information, to pull competitor catalog listings and pricing pages off the open web.

The AI connects to Firecrawl through what's called an MCP server, which is just a plug-in that lets Claude reach out and use a specific tool on its own, the same way you'd hand someone a specific piece of equipment instead of asking them to build it from scratch. Claude pulls the spreadsheet, uses that connection to search and cross-check pricing against what's actually listed online, and compiles the result into a set of recommendations: raise this part, lower that one, and why, based on margin.

There's a wrinkle, though: our part numbers and our competitors' don't match, so there's no one-to-one lookup between the two.  The first pass classifies which of our parts are proprietary — nothing else on the market is really comparable, so those get set aside.  Everything else runs through a semantic search: matching a part to a competitor's based on how similar its description and specs are, not on any shared numbering, using a catalog that lists our parts and competitors' parts side by side.  That gets a high-probability match, not a certainty.

It runs on demand, not on a schedule, because we change pricing often enough that a monthly cadence was already too slow.  Someone kicks it off when they need an answer, and it produces one output for three different audiences at once: the executive team, the sales team, and finance — each seeing the same recommendation, framed for what they need to decide.

Nobody lost their job over this. And that was important to me. What changed is what they spend their time on.

The people who used to search, compile, and collate that information by hand are now spending that time on harder judgment calls — the kind of analysis a monthly USB drive never left room for.  For the pricing coordinator, that means reviewing the matches the semantic search was least confident about: is this actually the same part, or did the match miss?  That's not taking the AI's recommendation at face value, it's teaching the tool and every correction sharpens what it uses to judge similarity next time.  The tool didn't just save time.  It gave them the room to be effective instead of just being busy. Instead of spending weeks researching individual parts, the bulk and majority of the work is done by the AI, completing the initial batch of research in just 90 minutes.

This one isn't finished, either, and that's the part I'd point to first.  I built the foundation.  The Finance team owns it now, and they keep extending it as their own needs change. New parts, new checks, new logic, new questions the original version never had to answer.  Every other project in this newsletter so far ended: shipped, closed, or quietly defunded.  This is the first one that's still being built on, by someone else, right now. And that’s because AI is allowing everyone to be a developer now!

(I'm keeping this section deliberately light on specifics.  The mechanics of how we price parts are the one part of this story that stays out of a public newsletter — everything above is shaped to explain how the tool works and what it changed, not what any given part actually costs.)

Every AI story in this newsletter so far has been about the gap between what AI was promised to do and what it actually did, or about a human doing the boring, unglamorous thing that quietly beat it.  This one has a bigger headline than that, and it's the one I keep coming back to on stage: nobody extending this tool today is a developer.  I'm not one either, not in the traditional sense — I built the foundation by walking through a spreadsheet with Claude, not by writing code from scratch.

That's the shift that actually matters here.  For decades, building software had a gate on it: a computer science degree, or a job title with "engineer" in it, stood between an idea and a working tool.  AI is taking that gate off its hinges.  The skill that used to take years of formal training, describing a problem precisely enough that a machine can act on it, is something anyone willing to sit down and try can pick up now.  The wall between "technical" and "non-technical" people was never really about ability.  It was about who'd had the chance to learn the syntax.  AI has now made the syntax optional.

The Finance team proves it every time they extend this tool without opening an IT ticket.  They don't wait on a developer's backlog — they just make the change themselves, because the tool is theirs to shape now, not something handed down finished.  That's a bigger claim than "it saved time."  It's proof the tool belongs to whoever's willing to build with it, not just to whoever has "engineer" in their title.

Before you evaluate the AI platform with the ambitious roadmap, ask yourself “Is there a spreadsheet already sitting on someone's desk, refreshed by hand once a month, that a much smaller tool (built by whoever's stuck doing that work, not a developer) could just read instead?”  What would it take to find out? An afternoon?  Or a plane ride?

What's the most boring, least strategic AI use you've actually gotten value from?  Reply, I read every one.

Plant Floor to Cloud: The Newsletter goes out every Tuesday.  A monthly companion podcast arrives later in the month. Same charter, more room to think out loud.  You’ll find the link on the website when its live.

Views are my own and do not represent any employer.  Nothing in this newsletter is attributable to or sourced from any specific company.  All case content is generalized to protect employer and client confidentiality.

Plant Floor to Cloud · PO Box 122, Alto, MI 49302