Custom AI Costs in the UK: Scope and Running Budget
Give suppliers the same workflow, data sample and acceptance checks before comparing quotes. The budget tiers here are illustrative references; integration access and ongoing support can change the scope.
A useful quote separates build scope, acceptance checks and operating costs. The table below gives illustrative budget references, not surveyed UK averages or a binding LookADev proposal. Confirm data access, workflow limits, human review and support responsibilities before agreeing a price or delivery date.
1. Illustrative Project Budget References
These planning ranges illustrate differences in scope. They are not evidence of market-wide prices or a payback forecast. Use the current proposal for commercial terms, and resolve unknowns in a small discovery task before committing to a build.
| Engagement Tier | Scope & Architecture | Reference Timeline | Reference Budget (GBP) | Acceptance Check |
|---|---|---|---|---|
| AI Systems Audit | 48h operational diagnostic, workflow leakage mapping, architecture blueprint & ROI model | 48 hours | £800 – £1,400 | Document scope and unknowns |
| Single-Bottleneck Sprint | Production agent (e.g. CV screening, WhatsApp intake, automated extraction) with live CRM sync | 7 business days | £2,500 – £4,800 | Verify one end-to-end workflow |
| Bespoke Semantic RAG | Hybrid vector index (pgvector), multi-source document ingestion, citation guardrails, client portal | 2 – 4 weeks | £8,000 – £18,000 | Test retrieval and source accuracy |
| Multi-Agent Operational Core | Autonomous loops, human sign-off checkpoints, bi-directional ERP sync, observability dashboard | 4 – 8 weeks | £18,000 – £35,000 | Test writes, retries and human approval |
| Enterprise Bespoke AI Platform | Proprietary internal SaaS, fine-tuned SLMs, multi-tenant RLS, complete IP ownership | 8 – 12+ weeks | £35,000 – £60,000+ | Test tenant isolation and handover |
2. Architectural Complexity Breakdown: What Drives the Price?
Scope the smallest workflow that meets your needs. Include representative test cases, failure handling and access controls in each tier; the model alone does not define the work.
Agree which documents the system may read and how updates reach the index. Evaluate retrieval on known questions, verify cited passages and give users an unanswered or human-review path when evidence is missing. Choose chunking and search methods from those tests.
When an AI agent interacts with your database or email server, errors can be catastrophic. Production agent architecture requires sandboxed tool execution, strict JSON schema validation, checkpoint fallbacks, and human approval queues.
List the reads and writes needed in each system. Check API access, sandbox availability, duplicate prevention, retries and reconciliation. Estimate this work from the actual interfaces rather than assigning a fixed percentage of the project fee.
3. Four Running-Cost Checks
Estimate costs per successful task and compare them with billing during a pilot. Include failed attempts and human review rather than pricing only the successful model call.
Measure input and output tokens per task on representative data. Send only the context the task needs; test that trimming it does not remove evidence or safety instructions. Multiply observed usage by forecast task volume and the provider's current rates.
Compare candidate models on the same task set, with latency, error rate and cost per accepted result. Use a cheaper route only when it meets the acceptance checks; retain a human fallback for decisions the system cannot support.
If your provider offers prompt caching, check its minimum length, expiry and billing rules. Budget from the cache-hit rate you observe, not a headline discount. Include uncached calls and any cache-storage charge in the estimate.
Set limits on attempts, task duration and spend, with a human-review path after failure. Test duplicate writes and exhausted budgets before launch. Choose limits for the workflow rather than allowing an agent to retry without a boundary.
4. Agree Scope and Change Control
Compare fixed-price and hourly quotes against the same deliverables. Ask which discovery, testing, rollout and support tasks each includes, and who approves a scope change. Fixed price is useful when acceptance criteria are clear; it does not remove unknown data or integration problems.
An illustrative £3,800, seven-business-day scope might specify structured CV extraction, Bullhorn sandbox sync and human review of each record. Agree a test set and permissions before signing. That example is not a current offer, an automated hiring decision or a promise of time saved.
Confirm delivery dates, defect coverage and post-launch monitoring in the written proposal. Specify ownership of custom code and the licences for third-party services and models; custom code ownership does not mean ownership of a provider's model. Ask for export and handover instructions.
Running-Cost Reference
Supabase's compute documentation explains hourly compute billing independent of database usage and recommends staging load tests. It describes that provider's billing, not an independent UK project-price survey. Add storage, transfer, model usage and support to your own budget. For the build-versus-buy calculation, see the three-year worksheet.
Frequently Asked Questions
How much does it cost to build custom AI software in the UK?
This guide's tiers are illustrative planning references, not a UK market survey or a current proposal. Define the workflow, integrations, evaluation set and deployment requirements, then ask for a scoped quote. Include token usage, hosting and support separately.
When is fixed-price AI development a useful choice?
Fixed price can make a well-defined scope easier to budget. Agree deliverables, acceptance checks, exclusions and the process for changes. Hourly discovery can be useful when data quality or integration access is still unknown; neither contract type guarantees a result.
What ongoing costs should I expect after custom AI deployment?
Estimate model input and output tokens per task, task volume, retries, retrieval and tool fees, plus hosting, storage, backups, monitoring and maintenance. Set application limits and alerts, then verify observed billing. There is no universal monthly ceiling.
How quickly does a custom AI project deliver financial ROI?
Use measured net benefit after rollout, not a promised time-saving percentage. Released staff capacity becomes a cash benefit only when it reduces a cost or supports additional profitable work. Include implementation, adoption and operating costs before estimating payback.
Request a Fixed-Price AI Project Scope
Bring a workflow example, approximate task volume and the systems it must connect to. Discuss the scope with Lucas Martins, then agree a quote and acceptance checks before implementation.