How much does AI automation cost?

Two costs, not one

The build is a one-off: scoping, connecting it to your material, designing how it behaves when it is unsure, and testing it against real inputs.

The running cost is per-use and never stops. This is the one that surprises people, because there is no licence to point at — just an invoice that tracks how much you used it.

Most budgets cover the first and ignore the second.

How the running cost actually works

You are billed per unit of text in and out — both the material you send and the answer that comes back. Two consequences follow, and they are the whole game:

1. Cost scales with usage, not headcount. Unlike per-seat software, adding people changes nothing unless they use it.
2. Long inputs cost money every single time. Sending a 40-page policy document with every question is the most common way to turn a cheap feature into an expensive one.

In practice an individual interaction — a question answered from your own documents, an invoice read, a reply drafted — costs a fraction of a penny to a few pence. That is genuinely cheap.

What makes it expensive

Calling the model per record

The classic mistake. A system processes ten thousand records and calls a model for each one. Each call is trivial; ten thousand are not.

Instead: trigger on events, not rows. Something changed → process that one thing. And cache: the same input should never be sent twice.

Sending more context than needed

Every question that ships your entire knowledge base costs the price of your entire knowledge base. Retrieving the three relevant passages instead of all four hundred is the single biggest lever on running cost.

No ceiling

Without a cap, a bug or a bot can run up a bill overnight. Every AI feature we build has a hard spend cap and cached calls, specifically so cost cannot surprise anyone.

Choosing a bigger model than the job needs

Model families differ several-fold in price. Classification and extraction rarely need the most capable model; judgement-heavy work sometimes does. Match the model to the task rather than defaulting to the top of the range.

A worked example

The Digital Opportunity Scan on this site is a real AI feature we run ourselves. Each scan reads a website, works out what the business does, and produces a set of opportunities.

Measured cost: about 2.3 pence per scan. At a thousand scans a month that is roughly £23 — less than most SaaS subscriptions, for a tool that does real work. It stays there because the inputs are trimmed, the output is bounded, and there is a hard timeout.

What to ask before you commit

• What triggers a model call? If the answer is “every record”, push back.
• What is the monthly ceiling? There should be one.
• What is sent each time? The less context, the cheaper and usually the more accurate.
• What happens if the provider raises prices or retires the model? Both happen. The design should let you swap.
• What does it cost to do nothing? The honest comparison is the hours the manual version takes, not zero.

The maintenance nobody mentions

Models get retired. We hit this on this very site: a model identifier stopped working and the feature failed until it was updated. Budget for occasional maintenance the same way you would for any dependency.

The short version

A narrow AI feature is cheap to build and cheap to run, provided it is designed with caching, event triggers and a spend cap from the start. It becomes expensive through one specific mistake — calling a model once per record — and that is a design decision, not an inevitability.

If you want to know whether AI is the right tool at all before worrying about cost, start with should I use AI in my business?