AI is genuinely useful for a narrow set of jobs and genuinely wasteful for most of the rest. The first thing we do is work out which of yours is which — and a fair number of these conversations end with a two-hour fix and no model involved at all.
Where AI actually earns its cost
The jobs where it reliably works have a shape in common: high volume, tolerant of a wrong answer being caught by a human, and currently done by somebody reading text and deciding what to do with it.
- Triage and routing. Reading inbound messages and deciding where they go and how urgent they are.
- First-pass drafting. Producing a reply, summary or description that a person then edits. The edit is the point — this saves the blank page, not the judgement.
- Extraction. Pulling structured fields out of documents nobody wants to read: invoices, statements, applications, leases.
- Answering the same forty questions. The ones your team answers constantly, at hours when nobody is available.
- Classification at volume. Tagging, categorising and flagging where the rules are fuzzy enough that a lookup table has never quite worked.
Where it does not
Anywhere a wrong answer is expensive and nobody will check it. Anywhere the real problem is that data is inconsistent — a model trained or prompted on a mess produces a confident mess. And anywhere the workflow is actually deterministic, which is more often than people expect.
That last one is the most common finding. A surprising share of “we need AI” requests turn out to be a missing form, an absent validation rule, or two systems that were never connected. Those have cheaper, more reliable answers, and we would rather tell you that than sell you a model. Where the answer is plumbing, it belongs in workflow automation.
How we build it
Three principles, all learned the way you would expect.
- Connected to your real data. A model that cannot see your systems can only produce plausible text. Most of the engineering in a useful AI feature is the retrieval and the integration, not the prompt.
- A human in the loop wherever a wrong answer costs something. The design question is never “is it accurate enough” in the abstract — it is what happens on the occasions it is wrong, and who notices.
- Running cost visible from the start. Per-token pricing has a way of surprising people in month three, particularly when a feature turns out to be more popular than expected. We model it before you commit.
What it costs
Two numbers, and people usually only ask about the first. There is the build, and there is the per-use cost that continues forever.
For a small business the second is often the one that decides whether a feature makes sense. A chatbot that handles fifty conversations a month costs almost nothing to run; the same chatbot on a site with real traffic is a different proposition, and the crossover point is worth knowing before you build rather than after. We will show you the arithmetic.
Common questions
Do we need our own model?
Almost certainly not. Training or fine-tuning is rarely the right answer at this scale — the commercial models are good enough, and the work that makes them useful is connecting them to your data and handling their failure cases.
What about our data privacy?
A real consideration and one to settle before building, not after. It shapes which provider is appropriate, what gets sent, and what stays on your side. If your data cannot leave your systems, that narrows the options and sometimes rules the project out — which is a legitimate outcome.
Will this replace someone’s job?
In the work we do, no — it removes a slice of a job, usually the least interesting slice. A five-person business does not have redundant staff; it has people doing four roles each. Taking one of those off their plate is the whole return.
How do we know if it worked?
We agree the measure before starting, and it is a business number — response time, hours recovered, volume handled without escalation. If we cannot name one, that is a strong signal the project is not worth doing.
Further reading
What an AI chatbot actually costs — why the per-conversation running cost matters more than the build price, what drives it, and how to check the idea is not really a form.
Start with the question, not the technology
Tell us what your team spends time reading, sorting or retyping. If AI is the answer we will say so, and if it is a form we will say that instead. More on the division at CAO Technologies, or get in touch.