A managing director of a twelve-person recruitment firm recently told me she’d asked three AI agencies for quotes to build a system that would handle initial CV screening and draft interview notes. The quotes came back at £4,500, £22,000, and £68,000. Same brief, same stated outcome. She couldn’t work out why the gap was so wide, or which number was actually right.
She isn’t unusual. The confusion is structural. “AI system” covers everything from a well-configured SaaS subscription to a bespoke multi-agent platform, and the price reflects that span rather than the quality of the outcome.
This post breaks down what a useful business AI system actually is for a service firm with five to fifty people, where the costs sit, and how to think about the investment before you commission anything.
What does a useful business AI system actually mean?
For a firm your size, a useful AI system is almost always one of three things: a productivity tool layered into software you already use, a custom integration that automates a defined workflow, or a knowledge assistant that answers questions from your own documents. The costs range from tens of pounds a month to tens of thousands upfront, and understanding which category fits your situation is the most useful step before you spend anything.
AI writing assistants and meeting transcription tools (ChatGPT Plus, Claude Pro, Otter.ai, Fireflies) run at £20-£80 per user per month and often sit inside tools you already pay for. Custom integrations, wiring AI into your CRM, automating invoice processing into Xero, building a first-response chatbot, typically cost £5,000-£20,000 to build, with many UK projects landing in the £10,000-£15,000 band. A retrieval-based assistant over your own documents (policies, proposals, service manuals) starts around £15,000 and clusters at £25,000-£35,000 for a well-specified build.
What you are almost certainly not doing at this stage is training your own AI model. Owner-managed service firms use existing foundation models from OpenAI, Anthropic, or Google via API, then add retrieval and a user interface on top. The underlying AI is not your cost driver. Integration, data preparation, testing, and compliance work are.
Why does the total cost catch founders off guard?
The build invoice rarely tells you the full story. Realistic ongoing costs for a custom integration are 10-15% of the original build per year for maintenance, bug-fixing, and model updates, plus a one-off refinement cycle of 15-25% of the original build cost after go-live. A £10,000 integration can therefore cost you an additional £3,000-£5,000 in its first twelve months.
There’s a second omission that catches people out: people costs. UK agency day-rates for AI chatbot and integration work run from £450 to £1,800 depending on seniority and specialism. If you’re working with an agency, that rate is baked into the project quote. If you’re hoping to run a build internally with a junior developer, the comparable all-in cost is £34,000-£40,000 per year in salary before any project overhead. For a well-scoped single-workflow project, a specialist contractor over four to ten weeks is typically the more practical path.
The third and most commonly ignored cost is compliance. If your system processes personal data, such as client records, CVs, or employee information, the ICO’s guidance on AI and data protection requires a Data Protection Impact Assessment where processing is likely to result in high risk. That work is real and billable. It belongs in your project budget from day one, not as an afterthought when a query comes in.
Where does the money actually go on a typical build?
The AI model itself (the API calls to OpenAI, Anthropic, or whichever provider you use) is typically a £50-£500 per month running cost once the system is live. The build cost is almost entirely the surrounding work: integration, data preparation, testing, and compliance. That split often surprises founders who assumed the AI component was the line item to watch.
For a single-workflow agent, say extracting invoice data into Xero or triaging incoming support tickets, a credible UK build range is £10,000-£30,000. For a retrieval system over a document library, expect £15,000-£35,000. In both cases the spend splits across integration and API plumbing (connecting AI to your existing systems), data cleaning and structuring (getting documents or records into a form the AI can use reliably), testing and quality assurance (checking outputs before anything consequential gets actioned), and change management (getting your team to actually use the system).
Security is part of this picture too. The NCSC’s guidance on using large language models safely flags prompt injection attacks, data exfiltration risk, and third-party API dependency as live concerns for organisations deploying AI connected to internal systems. Building in input validation, data segregation, and monitoring adds time and cost to any build. Skipping it creates a liability rather than an asset.
When does the investment make sense, and when should you hold off?
The clearest signal that a custom AI build is commercially justified is a high-volume, repeatable process where quality can be verified by a human before anything consequential is released. Invoice processing, first-line support triage, policy FAQ chatbots, and meeting summarisation are solid candidates because the output is checkable and the volume is predictable enough to model an ROI.
The weaker case is a bespoke, relationship-driven process where every engagement is genuinely different. White-glove advisory, complex negotiations, and matters where a wrong output carries real liability are harder to systematise usefully. The ICO’s guidance on automated decision-making restricts where you can remove human oversight from decisions with legal or similarly significant effects, and in credit, insurance, and employment screening that constraint is active.
You also need the maintenance budget in place before you build. With ongoing costs of 10-15% of build cost per year, a £30,000 system carries a standing annual cost of several thousand pounds. If that figure isn’t already in your operating budget, the system will degrade, become unreliable, and eventually become a management problem rather than a business tool.
If you’re not yet sure where your best candidate process sits, start with off-the-shelf tools at £20-£80 per user per month for three to six months. Measure the actual time saved, the adoption rate, and the bottlenecks that remain. That data is more useful than a gut-feel brief to an agency.
What should you know before you commission anything?
The build-versus-configure question deserves an honest answer before you speak to any supplier. Off-the-shelf AI features inside SaaS platforms you already use, including HubSpot, Xero, Microsoft 365 and Google Workspace, are under-used by the firms that already pay for them. Configuration costs are modest, the vendor handles infrastructure and model updates, and if that covers 80% of your requirement, building something bespoke for the remaining 20% is rarely the right trade.
When a custom build is genuinely justified, regulation is now a practical factor even for straightforward UK service firms. The EU AI Act introduces obligations for providers and deployers of AI systems serving EU customers, and UK SMEs will find many AI vendors already shaping contract terms around it. UK firms in financial services or other regulated sectors face additional requirements: the FCA and Prudential Regulation Authority have jointly consulted on AI and machine learning governance, stressing that firms remain responsible for model risk even when the model sits with a third-party supplier. A vendor telling you compliance is entirely their problem is not giving you an accurate picture of where accountability lands.
The number that matters before you approve any build is the total three-year cost, covering the initial build, post-launch refinement, ongoing maintenance, API fees, and compliance work, compared against the staff hours and error rate you are replacing. Get that figure on paper, compare it to the cost of the problem you’re solving, and the decision usually becomes clear.
If you want help thinking through where AI genuinely earns its cost in your business, Book a conversation.
