blog

Paying for AI per person or per job? For automation, the sums look very different

When a small business decides to "do something with AI", the first move is usually to buy everyone a subscription. That's the right call for some things. For the boring, repetitive jobs that most owners actually want rid of, it's often the wrong way to pay.

There are two ways to buy AI, and they suit different work.

Paying per person

This is the chat window. ChatGPT Business, for example, is £15 a seat a month billed annually for a standard seat. The model is simple: so much per person, per month.

That's a good deal when a person is sitting there using it: drafting a tricky email, summarising a contract. They ask, read the answer, and decide what to do with it.

It's a poor fit for a job that runs without anyone sitting there. If the plan is "every invoice that lands in the accounts inbox gets read and the details typed into a spreadsheet", nobody should be copying and pasting each one into a chat window. That's the job you were trying to get rid of, with an extra step.

Paying per use

The other route is the API. That's the same models, but your own software sends them work directly, and you pay for what you use rather than per head.

The unit is the token, which is a chunk of text, roughly part of a word. You pay for what goes in and what comes back. Anthropic's pricing page, checked today, lists its small model, Claude Haiku 5.5, at $0.10 per million tokens in and $0.50 per million out, for normal-sized requests. Its middle model, Sonnet 5.5, is $2 in and $10 out.

A made-up example. Say you get 1,000 supplier invoices a month and want the supplier, date, total and VAT pulled out of each. Call it 3,000 tokens in per invoice and 500 out. That's 3 million tokens in and half a million out a month.

  • On Haiku 5.5: about 30 cents in and 25 cents out. Around 55 cents a month.
  • On Sonnet 5.5: about $6 in and $5 out. Around $11 a month.

If the work doesn't need doing straight away, the same page offers a Batch API with "a 50 percent discount on both input and output tokens". Overnight runs are a natural fit.

Your real numbers will differ. Scanned PDFs and photos of receipts use more tokens than clean text, and longer instructions cost a bit more each time. But the shape holds. For a well-defined job, the AI itself is often the cheapest part.

So where does the money go?

Into everything around it. That's the bit the per-use price hides.

Building the plumbing. Something has to collect the invoices from the inbox, send them off, get the answer back and put it somewhere useful. That might be a small script, a no-code tool like n8n, or a proper link into your accounts software.

Checking the answers. AI makes mistakes, and it makes them confidently. A sensible setup flags anything odd for a person to look at: totals that don't add up, a supplier it hasn't seen before, a blurry scan. Plan for someone spending a few minutes a day on the exceptions, especially at the start.

Keeping it going. Suppliers change their invoice layouts. Models get replaced by newer ones. An automation nobody looks after quietly stops working, usually the week you most need it.

So the honest comparison isn't £15 a seat against 55 cents. It's the cost of setting up and looking after the automation against the hours someone currently spends doing the job by hand.

How to tell which you need

A quick test I'd use:

  • A person reads every answer and decides what to do? Per seat. Buy subscriptions for the people who'll actually use them, not the whole staff list.
  • The same task, on similar documents, over and over? Per use, built into a small automation.
  • Documents that shouldn't leave the building at all? Neither. That's when running a model on your own hardware starts to make sense, and I've written up what that seems to take.

Whichever route, the rules on what data can go where still apply. If you haven't settled that yet, my post on whose account it is is the place to start.

What to do this week

Ask everyone for the one dull task they'd most like a machine to do. Copying details from emails, sorting enquiries, checking delivery notes against orders.

Then for each, write down two numbers: how many times a month it happens, and how long it takes each time. Anything that happens hundreds of times a month and follows the same pattern is a candidate for per-use automation. Anything that needs judgement every time probably belongs with a person and a chat subscription, if anything.

You'll likely find one or two jobs where the sums are obvious. Start with those, and keep a person checking the results until you trust it.

If the job needs joining up with software you already run, such as your accounts package, CRM or shop, that's the kind of work the team at Dev Partners do with API development and integrations. Their page on how they use AI shows where it helps and where a person still checks.