A new kind of model made the rounds recently: instead of replying in sentences, it returns a category and a confidence score — yes/no, urgent/normal, invoice/quote. It sounds like a technical footnote, but we think it's one of the most practical signals in recent AI news. Because when we look at what small companies actually want from AI, it's rarely conversation. It's a decision.
Most real tasks are classification, not writing
Whenever we map workflows with a client, the same pattern shows up. Does this email go to sales or support? Does this message contain an order? Is this invoice a duplicate or a new cost? Does this application meet the minimum criteria? Is this piece of feedback a complaint that needs an answer today? None of these questions need a well-written paragraph. They need one label and a sense of how sure the system is.
You can ask a chat model to do it, and it works — but it's more expensive, slower and fuzzier. You get back a sentence you then have to parse, worded slightly differently every time. A classification approach gives you structure: a fixed set of options, a numeric confidence, predictable cost. That's exactly what you need when the output feeds another automated step rather than a human's screen.
The confidence score is what makes automation safe
In our experience the real difference is between a system that says "this is a complaint" and one that says "this is a complaint, confidence 0.58". In the first case you either trust everything or nothing. In the second you can draw a line: above a threshold it moves on by itself, below it lands on someone's desk.
That means automation doesn't have to be an all-or-nothing decision. Start by letting the machine handle the clear cases — usually the bulk of the volume — and leave the ambiguous ones to a person. You get the time savings without trading away quality. And because the line is written down as a number, you can move it calmly later, once you've seen how the system behaves on your real data.
How to start: one week, one workflow
We recommend the most boring possible beginning. Pick one repeated decision someone on your team makes by hand every day. Collect a couple of hundred real examples from the past months — emails, forms, invoices, whatever the input is. Write down the categories people actually choose between, and always include an "unclear" option. Then have a human label those examples by hand. This is the tedious step everyone wants to skip, and it's precisely the step that decides whether the project works.
After that it's simple: run the same examples through the model and compare with the human labels. Now you have a number instead of a feeling. If accuracy is good enough, run the first month in "suggestion mode" — the machine proposes a label, a person confirms with one click. Only when confirming has clearly become a formality should you let the confident cases through automatically. Then measure two things: what share of the volume flows through untouched, and how many errors reach a customer.
Where the limits are
This approach isn't universal. If your categories change every month, or there are thirty of them and half overlap, the model isn't the problem — the process isn't settled yet. Fix the process first. We also advise against automating decisions that affect a customer irreversibly: terminating a contract, rejecting a claim, setting a price. There the machine can pre-sort and explain its reasoning, but a person makes the call and the customer needs a clear route to appeal. That's not excessive caution on our part — European rules expect it too.
And if the volume is genuinely small — a couple of dozen cases a week — the honest answer is that automation won't pay for itself. A good checklist is the better tool.