What a bespoke AI build actually costs UK mid-market businesses in 2026

The honest answer most consultancies dodge

"How much does it cost?" is the question every consultancy site dodges with a Contact Us button. This guide does not.

Below are honest 2026 ranges for the kinds of bespoke AI builds AIMindShift typically ships for UK mid-market operators. Final cost for your specific scope comes out of the discovery call — but the ranges below will tell you whether you are in the right conversation before you book.

Voice agents

£15k–£40k for a focused build. Single inbound line, brand-voice training, integration with one calendar and one CRM. Qualifying logic built around your real criteria. Hands off to humans on your team's rules.

£40k–£90k for a multi-line, multi-language build. Multiple phone numbers, multiple agent personas, multi-language support, complex routing, deeper CRM integration, after-hours handover protocols.

Ongoing costs: telephony (Twilio, Vonage etc.) at roughly £0.01–£0.05 per minute, plus the AI voice model at £0.05–£0.20 per minute of conversation. For a busy line, ongoing AI + telephony cost is typically £200–£1,500/month.

Chat agents

£12k–£30k for a single-channel build. One channel (website, WhatsApp, or Teams), grounded in your real documentation, capable of taking action across one or two systems (e.g. update an order, raise a ticket).

£30k–£70k for a multi-channel, multi-system build. Same agent across web, WhatsApp, Teams, Slack, SMS. Integrated with CRM, order management, billing, support ticketing. Full handover with conversation context to humans.

Ongoing costs: AI model usage at £0.001–£0.01 per conversation turn. For a moderately busy support deployment, typically £50–£400/month.

Bespoke AI CRM

£25k–£60k for a focused build. A CRM shaped around your sales motion with AI-assisted data entry, scoring against your real win/loss history, and automated reporting. Migration from existing CRM included.

£60k–£150k for a full operational CRM. Multi-team, complex pipelines (B2B + B2C, multi-territory), deep integration with telephony, email, calendar, marketing automation. Custom dashboards and forecasting.

Ongoing costs: hosting and AI usage typically £100–£600/month depending on team size.

Document intelligence (classification, extraction, summarisation)

£20k–£50k for a focused build. One document type (contracts, invoices, claims, policies), one taxonomy, structured extraction into your existing system. Typical throughput: thousands of documents per day.

£50k–£120k for a multi-type, multi-workflow build. Multiple document types, multiple downstream systems, human-in-the-loop review for high-stakes documents, audit-grade logging.

Ongoing costs: AI model usage at £0.005–£0.05 per document, plus hosting at £100–£500/month.

Internal apps and dashboards

£15k–£40k for a focused internal tool. One workflow, one team, replacing a SaaS tool or a spreadsheet/Trello workaround. AI in the background drafting, summarising, classifying, or flagging.

£40k–£100k for a department-wide internal platform. Multiple workflows, multiple roles, integrations with your existing stack, reporting and analytics.

Ongoing costs: hosting £50–£300/month, AI model usage variable.

Self-hosted virtual assistants (Clawdbots)

£30k–£70k for a focused self-hosted build. One assistant, integrated with internal email, calendar, file shares, and one or two business systems. Runs entirely inside your network.

£70k–£180k for a full self-hosted AI platform. Multiple assistants for different roles, deep system integration, hardened security architecture, audit infrastructure, ongoing model maintenance.

Ongoing costs: hardware/hosting on your own infrastructure (often already paid), AI model and infrastructure maintenance typically £200–£1,500/month depending on workload.

What changes the price (in plain English)

• Number of integrations. Every system the build talks to is engineering effort. Two integrations is light; six is substantial; ten is a project.
• Data complexity. Clean data in a well-structured database is fast. Messy historical data across PDF archives and Excel files takes weeks to make usable.
• Security requirements. Standard cloud security is the baseline. SOC 2, ISO 27001, regulated industry compliance, air-gapped deployment — each adds materially to scope.
• Customisation depth. A capability that fits a standard pattern is faster than one that requires bespoke logic at every step.
• Volume and latency. Light, slow workloads run cheap. Real-time, high-volume workloads need engineering for performance — usually more cost, sometimes more infrastructure.

What does not change the price as much as you think

• The underlying AI model. Whether the build uses GPT-4, Claude, Llama, or Mistral matters less than the engineering around the model. Model choice is usually a £-per-interaction question, not a £-of-build-cost question.
• How "advanced" the AI feels. Most of the build cost is integration, data work, and operational engineering — not the AI itself.

Build vs three years of SaaS

A focused bespoke build at £25k delivered in week 6, with maintenance at £4k/year, totals roughly £33k across three years.

A comparable mature SaaS subscription with 40 seats at £40/month, growing 10% per year, totals roughly £65k across three years.

Bespoke wins on three-year TCO at this scale — and that is before the workflow fit argument. Below 15 seats, SaaS typically wins on TCO; above 40 seats, bespoke usually does. The crossover sits somewhere between 15 and 40 depending on price per seat.

The discovery call gives a specific number

Ranges are useful for sizing the conversation. They are not useful for committing budget.

A discovery call gives you a specific costed range for your specific scope: not "voice agents are £15k–£40k" but "for the scope we mapped, your build is £28k ± £6k, with a 4–6 week timeline and £180/month ongoing." That is what you take to the board.