At Uppercut Labs, a custom AI Agent Build starts at £3,000. A scoped first workflow costs £1,950. Before either, an AI Opportunity Sprint costs £1,000, credited against the subsequent build. These are our published starting points, not a claim about an industry average. GBP prices exclude VAT where applicable.
The implementation fee is only one part of the decision. An agent also needs someone to review failures, maintain its information and check changes to models or integrations. Model usage, hosting and any channel fees are another separate line. Ask for all three in a proposal.
Our AI implementation packages
| Package | GBP | Typical delivery |
|---|---|---|
| AI Opportunity Sprint | £1,000 | 1-2 weeks |
| First Workflow Live | £1,950 | 2-3 weeks |
| AI Agent Build | From £3,000 | 3-5 weeks |
| AI Software Build | From £15,000 | 8-12+ weeks |
| AI Operations Retainer | From £600/mo | Monthly |
Delivery windows depend on agreed scope, access to systems, data readiness and review time. The operations retainer covers the agreed live agent or workflow. Broader existing-product and RAG maintenance starts at £750 a month and needs its own scope. Compare the services and pricing.
A worked first-year budget
Take an illustrative small internal agent at the published minimum build price, with 12 months of the minimum operations retainer and a £100 monthly usage/hosting allowance. This is a planning example, not a quote or a measured customer deployment. It assumes that one bounded task fits the minimum scope.
| Item | Calculation | First-year cost |
|---|---|---|
| Implementation | £1,000 sprint + £2,000 remaining build after credit | £3,000 |
| Operations support | 12 × £600 | £7,200 |
| Model usage and hosting allowance | 12 × £100 | £1,200 |
| Total | Build + support + usage allowance | £11,400 |
A £50-£150 monthly usage allowance can be a starting point for a small internal agent. Actual spend depends on request volume, model choice, document retrieval, storage, hosting and retries. Customer-facing messaging may add channel charges. Existing software licences, extra integrations and VAT are not included in the example.
What makes an agent more expensive?
- More systems and actions: every integration has authentication, permissions, field mapping and failure cases.
- Untidy or restricted data: identifying the right source and who may access it can take more work than the chat interface.
- Consequential decisions: a draft summary needs different controls from a system preparing a financial or contractual action.
- Many users or high volume: concurrency, response times and support expectations affect the design.
- A changing process: repeated changes to rules, knowledge and integrations need ongoing engineering.
Ask a supplier to identify the assumptions behind the starting price. A useful scope says which job is included, which tools and actions are connected, what counts as acceptance and how changes are priced. A long feature list without those boundaries is difficult to compare.
Does the time recovered justify the spend?
Measure current handling time on a representative sample. Include checking, corrections and exceptions in the after figure. For example, 200 monthly tasks falling from 10 minutes to four minutes of human effort would recover 20 hours a month. At an assumed £30 per hour that represents £600 of staff capacity. It is not automatically £600 in cash saved.
In that example, a £600 support retainer plus £100 usage would exceed the £600 capacity value before recovering the build cost. On time alone, the case does not work. More volume, a different support scope, fewer errors or a demonstrable revenue effect might change it. Those need evidence, not optimism in a spreadsheet.
When a custom agent is the wrong purchase
If a supported assistant already does the job, test it first. If the process follows fixed rules, use ordinary automation where possible. If nobody owns the process or can identify acceptable outputs, fix that before commissioning software. Our build-versus-buy guide walks through the alternatives.
What to ask before accepting a proposal
- Is the discovery fee credited against the build, and how is that shown?
- Which ongoing costs are mandatory, optional or paid directly to third parties?
- What sample cases must pass before the system goes live?
- Who handles failures, model changes and new requests after handover?
- Who owns the accounts, code and documentation, and what happens if we change supplier?
Bring the task volume, the tools involved and a few representative examples to the first conversation. Those tell us more about a sensible budget than a wish for a general-purpose AI employee. Discuss a scoped AI implementation.