AI automation pricing: what you actually pay for

"AI automation" covers anything from three steps chained between two tools to a system wired into the ERP that reads thousands of documents. Prices follow. They are also hard to find: of 277 UK AI agencies reviewed by the directory TrustedAIAgencies, 113 showed at least one price on their own website (IT Brief, 4 November 2026).

These are advertised starting prices, so the real invoice is usually higher. They mix one-person firms selling a chatbot set-up with agencies that integrate with business systems. Use them as a floor, not as a budget.

What pushes the price up is not the AI. It is everything around it: the link to the ERP, the checks before anything is written, the approval screen for the sales office, the handling of cases the model cannot deal with. Our pages on sales order entry, quote preparation and supplier invoices describe that work.

If you only want to know where to start, an assessment is sold at £5,000 by SpotDev and at a median of £750 across the agencies in the TrustedAIAgencies study. What such a review covers is described on the AI consultancy page.

Models are billed by the volume of text they process, counted in tokens (roughly a fragment of a word). Providers publish their prices per million tokens, in US dollars. As recorded on 12 November 2026:

Worked example for a document. A PDF page typically uses 1,500 to 3,000 tokens of text, according to Anthropic’s documentation. Take a two-page sales order: 10,000 input tokens, instructions and page images included, and 1,000 output tokens. Those are our assumptions. With Claude Sonnet 5.5: 10,000 × $2 ÷ 1,000,000 = $0.02, plus 1,000 × $10 ÷ 1,000,000 = $0.01, so $0.03 per order. For 2,000 orders a month: $60. With Claude Haiku 5.5 the same sum gives $0.0015 per order, or $3 a month; with Mistral Large, $0.0065, or $13 a month.

Worked example for a message. An assistant that answers from your documents sends the model the question, the passages it found and its instructions: say 6,000 input tokens and 500 output tokens. With Claude Sonnet 5.5: $0.012 + $0.005 = $0.017 per answer, or $51 for 3,000 messages a month.

Model usage is the smallest recurring item. The others are fixed, and they do not go away when volume drops.

Maintenance is not optional. A supplier changes the layout of its orders, a model is retired, the ERP moves to a new version: with nobody watching, the automation degrades without warning. Ask for a monthly report showing how many documents went through with no correction.

Start with a subscription if the need is individual. Move to bespoke when the task is the same a hundred times a week and touches your business data. The two are not exclusive: you can keep a general assistant for writing and commission one automation for order entry.

Nobody can promise you a return before measuring. You can, however, set out the sum yourself with your own figures.

1. Time the task. Ten real cases, start to finish, by the person who normally does it.
2. Count the volume. Documents or requests per month over the last twelve months, not the best month.
3. Multiply. Fictional example: 6 minutes per order × 1,500 orders a month = 150 hours a month.
4. Apply a realistic rate. No automation handles 100% of cases. During the pilot, measure the share processed with no correction, and count the time spent checking the rest.
5. Subtract the running costs. Usage, hosting, maintenance: the table above.
6. Compare with the build price. That gives a payback period, to be checked after three months of real use.

If the sum does not work with cautious assumptions, do not start the project. If the volume is low, a subscription tool or a well-made document template is often enough.