Cost of implementing AI in a business: UK guide (2026)

How much does AI cost for a business: the short answer

AI isn't bought as one product. It comes in three levels, each with a different cost profile: seats on an off-the-shelf assistant (a monthly fee per user), automation that connects AI to the tools you already use (a set-up cost plus usage), and bespoke development (a one-off project plus a monthly running cost).

For bespoke work, UK sources publish these orders of magnitude. Whitehat (March 2026) gives £5,000–£12,000 for an AI audit, £3,000–£20,000 for a small business's first use case and £25,000–£60,000 for a full project over 12–16 weeks. LeverX (October 2026) puts a single-workflow AI agent at £10,000–£30,000 and a multi-step agent with integrations at £30,000–£80,000. Magora publishes AI proofs of concept from £10,000.

A warning on the sources: almost all are companies that sell these services, and none publishes its sample. Use them as an order of magnitude, not as a tariff.

These are third-party market figures, not SocketStack prices. If someone offers ‘AI for the whole company’ without naming a process, ask them to name one before talking about price.

What UK and Irish businesses actually spend on AI

Official data puts the agency figures in context. According to the ONS, about 35% of UK businesses with 10 or more employees used at least one AI technology in June 2026, up from about 12% in September 2023. The figure is 28% for businesses with fewer than ten employees and 49% for those with 250 or more. Surveys differ: DSIT's research, with fieldwork in 2025, found 16% of businesses using AI. In Ireland, the CSO found that 20.2% of enterprises used AI in 2025, and 17.2% of small ones.

On spending, DSIT asked businesses that use or plan to use AI what they spent in 2024. Among the 289 that could give a figure, the median was £2,000 and the mean £19,000: most spend little, a few spend a lot. 31% reported no AI-related spending at all, in many cases because they rely on free or low-cost tools.

Cost isn't the main brake. The ONS puts cost as a barrier for around 7–14% of businesses depending on size, and in DSIT's survey the most cited reason for not adopting AI was a lack of identified need (71%). The first question isn't ‘how much?’ but ‘for which process?’

AI project cost by type

Ranges published by UK sources in 2026. Monthly running costs depend on volume and are covered in the next section.

• The sources disagree on the first project (from £3,000 to £30,000). The difference is usually integration: an AI step that drafts text for a person to paste is cheap; one that reads a supplier invoice, checks it against the order and posts it to Sage or Xero is a software project. See AI invoice and document extraction.
• Mid-market pilots are a different scale. Whitehat quotes £35,000–£120,000 over 8–16 weeks for a pilot or proof of concept in larger organisations.
• If the AI sits inside a larger build, the rest of the software follows the ranges in how much custom software costs.

What you pay every month: tokens, seats and running costs

Language models charge by token, a chunk of text a few characters long. What you send (input) and what the model writes back (output) are priced separately. These are the official list prices as of November 2026, in US dollars per million tokens, as both providers publish them. For sterling, US$1 ≈ £0.74 (ECB reference rates, 4 November 2026), so US$10 is about £7.40:

• Documents example: 1,000 supplier invoices a month, assuming about 4,000 input and 500 output tokens each, is 4 million input and 0.5 million output tokens. On a US$2/US$10 model that's about US$13 a month (≈ £10); on a US$0.10/US$0.50 model, under US$1; on a US$10/US$50 model, about US$65 (≈ £48).
• Chatbot example: conversations cost more because each message resends the history. 3,000 conversations of six turns, at about 3,000 input and 300 output tokens per turn, is 54 million input and 5.4 million output tokens: about US$162 a month (≈ £120) on a US$2/US$10 model.
• Discounts: Anthropic advertises 50% off with batch processing and lower rates for cached, repeated parts of a prompt; OpenAI publishes Batch and Flex prices at half its standard rates. For documents that don't need an instant answer, that's real money.
• Seats: Microsoft 365 Copilot Business is £13.80 per user per month on annual billing in its first-year offer, £16.10 list, excluding VAT. Anthropic lists a Claude Team standard seat at US$20 per month billed annually, or US$25 billed monthly.
• Tool stacks: Expertsure estimates £200–£800 a month for a growing small business with several AI tools and an automation platform, and £50–£200 for a starter set-up.
• Maintenance: 15–20% of the implementation cost a year (ProfileTree), or 15–25% of year-one costs (Whitehat). It covers adjusting prompts, reviewing errors, monitoring and moving to new models when providers retire old ones.

What drives the cost of an AI project

• Integration with your systems. Reading from and writing to Xero, Sage, Business Central or your CRM is usually most of the work, not the model.
• The state of your documents and data. Clean, consistent PDFs are quick; scans, handwriting and twenty supplier layouts take more testing.
• Accuracy and human review. A draft for a person to check is cheap. An automatic posting to your ledger needs validation rules, confidence thresholds and a review queue.
• Volume. Tokens scale with documents and conversations. Low volume rarely justifies a bespoke build; high volume justifies optimising model choice and caching.
• Permissions. An internal assistant that must respect who can see which documents costs more than one over a single shared folder. See private AI assistants for company documents.
• Model choice. At the prices above, the same task costs a hundred times more on the top tier than on the smallest model. The right choice comes from testing on your own data.

Hidden costs of AI

• Data preparation. Gathering example documents, writing down expected results, tidying the knowledge base. Often underestimated.
• Human review time. Someone checks the doubtful cases. Budget for it, especially in the first months.
• Staff training. Whitehat prices it at £3,000–£8,000 over 2–4 weeks.
• Usage growth. If the tool works, people use it more. Set usage alerts and a monthly budget with the API provider.
• Model changes. Providers retire older models; prompts and tests need revisiting when you switch.
• Monitoring and evaluation. LeverX lists monitoring, evaluation and optimisation among the running costs to include in the total cost of ownership.
• Data protection. A processing agreement with each provider and a review under UK GDPR. Whether providers train on API data is covered in our FAQ: does AI train on my data?

How to compare AI quotes: checklist

• A named process and volume: which documents or conversations, and how many a month.
• Set-up and running costs, separately: the one-off price plus an estimate of the monthly cost of model, infrastructure and support.
• A pilot on your own data: tested with your real documents, not a demo dataset, against an agreed accuracy measure.
• Human review: what happens to the cases the AI isn't sure about.
• Integration scope: which systems it reads from and writes to, and who maintains the connection.
• Model and provider: which ones, why, and whether they can be swapped later.
• Accounts in your name: the API account and billing in your company's name, so you see the real usage.
• Ownership: prompts, workflows and code are yours, with copyright assigned in writing.
• Pricing model: fixed price or days. We compare both in fixed price vs time and materials.

Common mistakes when implementing AI

• Starting with the technology. ‘We want an agent’ before ‘we lose 20 hours a week keying in invoices’.
• Buying seats nobody uses. Licences for the whole company before checking who will use them every day.
• Skipping the pilot. Committing to a full build before testing accuracy on your own documents.
• Forgetting the monthly bill. The build is one-off; tokens, hosting and maintenance aren't.
• Automating without review. Letting AI post to the ledger or reply to customers with no human check in the early months.
• Building what you can rent. If Copilot or an off-the-shelf tool does the job, use it. Bespoke pays when the process is yours and the volume is there. If it isn't clear where to start, see AI consulting for SMEs.

How SocketStack works

We start with a free 30-minute call to find one process with real volume; if it's worth building, you get one fixed price per project with the expected monthly running cost set out separately, a working demo every two weeks using your own documents, and code and workflows that are yours. The detail is in how we work.