Most of what a small business loses time to isn’t a language-understanding problem, so a language model doesn’t fix it. In a 15-person precoat and finishing shop I worked with, the actual win was getting job data into one structured place — no AI involved at all. AI use belonged in exactly one spot: after the data was clean, answering plain-English questions over it. Get the order right or you’ll spend real money on the wrong thing first.
Why “add AI” is usually the wrong first move
A shop floor running on paper and spreadsheets doesn’t have an understanding-language problem — it has a nothing-talks-to-anything problem. No shared, real-time view of what stage a job was in; someone had to walk the floor and ask. A large language model bolted onto that doesn’t fix the walk-and-ask; it just adds a chat window to it. The fix was mechanical: one system of record, one interface, one source of truth that updates live. That’s what the case study describes, and none of it was AI — it was a database, an API, and an app that works offline on a bad-wifi shop floor.
What actually used AI, and what didn’t
| Piece of the system | Used AI? | Why |
|---|---|---|
| Job and stage tracking, offline sync | No | This is a data-consistency problem — get one write to land once, even offline. A model doesn’t make writes more reliable. |
| Photo attachment, bunk records | No | Structured capture at the point of work; a form, not a conversation. |
| Label cost fix (in-house printing) | No | Once the data existed, this was arithmetic — cost per label, break-even months — not language understanding. |
| “Ask your data a question in plain English” | Yes, once it applies | This is the one place a model adds something a spreadsheet formula can’t: someone who isn’t a database person asking a real question and getting a real answer. |
Where AI is worth paying for
The pattern holds outside that one shop too. AI is worth using when the task is genuinely open-ended — drafting a first-pass reply, summarizing a pile of notes, answering “how many jobs did we lose money on last month and why” over data that’s already correct — and a person is going to look at the output before it goes anywhere. That’s the honest use case: a faster first draft, not an autonomous decision-maker.
Where it isn’t — and where it actively costs you
AI is a bad fit for anything that has to be right every single time with nobody checking: updating a customer record, moving money, deciding whether to reorder stock. A model that’s right 95% of the time and silently wrong the other 5% is worse than no automation at all in those spots, because the failure doesn’t announce itself — it just becomes a wrong record nobody catches until it costs something. The Alpine shop’s editable-never-deletable record design exists for exactly this reason: a correction has to show up in the log, not silently overwrite what happened.
It’s also a bad fit as a first move on messy data. Pointing a model at a spreadsheet full of inconsistent product names and duplicate customers doesn’t clean it up — it produces confident, wrong answers about it. The data has to be structured and correct first. That ordering — data, then AI — is the actual finding here, not “AI is good” or “AI is bad.”
What this means for a $400 first step
The right first question for a 15-person business isn’t “where should we add AI.” It’s “where is data trapped in a place nothing else can read it.” Answering that honestly sometimes means the answer is a plain database and a small app, with zero AI in it, same as it was here. Current pricing covers what that first step costs; book a free call if you want a straight answer on which side of this your situation falls on.
