Budgeting for AI Development: What a UK Project Actually Costs

An AI development budget needs to cover getting the feature built, making it dependable and keeping it useful after launch. The model bill is only one line. Data preparation, integrations, testing and the people who handle exceptions can change the total far more than the price of a single generated answer.

Pixelfield’s main AI development page lists feature discovery from £2,000, two-week consultancy from £5,000, PoC or MVP work typically at £8,000–£20,000, and production builds from £50,000. The dedicated AI software page separately lists production integrations inside an existing product from around £10,000. These describe different scopes; they are not a national price average or amounts you should automatically add together.

This guide separates those published prices from a fictional project budget, then shows how to compare proposals without overlooking the work your own team must do.

Start with the scope behind the price

If the system will use tools or change business records, the AI agent development cost breakdown separates integration work, approval controls and ongoing human review from the model bill.

The figures below were checked against Pixelfield’s public service pages on 1 October 2026. They are starting points for a conversation, not a quote for your project.

Sources: AI development services and AI software development. A feature inside an existing product is a different purchase from a new multi-workflow platform. The £10,000 figure applies to the existing-product feature scope described on the dedicated service page; it should not be presented as the starting price for every production AI project. The main page lists production builds from £50,000 and describes advanced enterprise products in that band.

The dedicated AI consultancy page also lists a two-week discovery engagement from £5,000. This is a broader consultancy engagement than the feature discovery starting at £2,000.

For ongoing support, the main AI development page gives an annual planning range of 15–25% of the original build budget for monitoring, retraining and infrastructure. Confirm the actual scope and quote: model API usage, your own staff time and new feature work should each have an explicit allowance rather than being assumed included.

Ask whether discovery or prototype work is credited against a later build, reused within it or priced separately. Do not assume every phase is compulsory. Confirm the VAT treatment, third-party charges and payment milestones in the written proposal; the table is not a tax-inclusive offer.

Break the estimate into work you can inspect

“Build an AI assistant” is too broad to compare. Replace it with a short description of the user, input, output and permitted action. For example: “Staff paste a support request into the existing helpdesk, receive a suggested category and draft reply, and approve the reply before sending.”

That version exposes several distinct pieces of work:

• Data and access: identify approved sources, remove irrelevant material, resolve conflicting instructions and apply access rules.
• The AI task: choose an approach and develop the prompts, retrieval or other logic needed for the specified output.
• Application and integration: connect the helpdesk, handle authentication, show errors and prevent duplicate submissions.
• Evaluation: agree representative test cases, acceptable outcomes and failures that block release.
• Release and operation: add monitoring, usage limits, fallback, documentation and an accountable owner.

Two suppliers may include those tasks under different headings. Compare the deliverables, rather than the number of rows. A cheaper proposal that excludes the helpdesk connection is not a cheaper version of the same product.

The uncertain parts deserve explicit assumptions. “One documented read-only API” is useful. “Integrations included” is not. Record what happens to the quote if access is delayed, source documents conflict or the existing system cannot support the proposed action.

For the technical questions behind that work, see adding AI to existing software. For a prototype, specify the decision it must support: our AI proof of concept example shows how a failed test can justify revising or stopping the project.

A worked AI budget: build plus twelve months of operation

The following is a fictional planning example, not a client project, a Pixelfield quote or a market benchmark. It concerns one internal drafting feature inside an existing application, using an existing model API. It does not include training a foundation model, a new mobile app or an enterprise-wide rollout.

Assume the buyer and supplier have explicitly agreed that these are separate, non-overlapping work packages. Reusable prototype work is already accounted for in the production amount; it is not billed again.

The reserve is spending capacity, not a supplier fee or an automatic extra payment. The internal allowance is an economic cost of staff time; it may not require additional cash if existing employees do the work within their normal hours. Keep those distinctions visible when finance reviews the request.

Now give the operating estimate its own assumptions. In this model, the feature completes 10,000 tasks per month. Each task makes one model call with 1,200 input tokens and 200 output tokens. Assume fictional prices of £2 per million input tokens and £8 per million output tokens. These rates are calculation inputs, not the current tariff of a named provider.

The calculation is:

• Input: 10,000 × 1,200 ÷ 1,000,000 × £2 = £24 per month.
• Output: 10,000 × 200 ÷ 1,000,000 × £8 = £16 per month.
• Base inference: £24 + £16 = £40 per month.
• Allowing 25% additional equivalent model usage for retries and testing gives £50 per month.

That is deliberately a simple workload. Multi-step agents, long documents, image processing, paid tools or repeated retrieval would need additional inputs. Replace the assumptions with measurements from a representative test before approving a production allowance.

In this example, support plus infrastructure is £550 per month, or £6,600 annually: about 24.4% of the £27,000 delivery total, within the published 15–25% range. Model usage and internal human review are additional lines, so the complete operating budget is larger than the support allowance. These figures remain illustrative, not a retainer quote.

Twelve months of operation therefore adds £14,400. The planning envelope is £47,800: £27,000 delivery + £2,400 internal preparation + £4,000 reserve + £14,400 operation. Excluding the unspent reserve, the modelled cost is £43,800.

This is a build-plus-twelve-months comparison, not a claim that all costs fall into the first financial year. It excludes taxes and any unlisted licences, migration, hardware or specialist assurance work. Add applicable items rather than treating a blank cell as zero.

Download the worked budget and assumption worksheet. The CSV contains the model values and calculation notes; it is an editable worksheet, not an automatic calculator.

What changes when the workload grows?

An inference estimate becomes useful when you can explain which input makes it change. If monthly task volume doubles with the same call count and token lengths, the model-use line above doubles from £50 to £100. That does not mean the whole operating bill doubles: hosting and support may remain within their allowances or cross a capacity threshold.

Human review can be more sensitive. The example assumes 400 exceptions among 10,000 tasks, or 4%. If 8% need the same three-minute review, the allowance rises from £600 to £1,200. With the other lines unchanged, the operating total becomes £1,800 per month, adding £7,200 over twelve months.

Those are scenario calculations, not predicted error rates. Test whether the task is actually completed, how often a person intervenes and how long that intervention takes. A draft that needs extensive rewriting has a different cost from a draft that needs a quick approval.

AWS’s guidance on calculating AI costs and value similarly separates direct AI services from related storage, infrastructure and operational costs. Its framework is useful for organising a budget; it does not validate the fictional prices above.

Keep budget approval separate from benefit claims

A cost estimate does not prove that a project will pay for itself. Establish what the current process costs, then measure the new process on comparable work. Include the effort needed to correct mistakes, maintain the source material and manage exceptions.

Time released is not automatically payroll saved. A team may use the capacity to answer customers sooner or handle a backlog without recruiting. Those can be valuable outcomes, but they belong in a different calculation from a reduction in headcount expenditure.

If the baseline is missing, buy evidence before committing to a larger rollout. A useful first phase should narrow a specific uncertainty: whether the source data is usable, whether outputs meet the acceptance criteria, or whether the proposed integration is feasible. It should not merely produce another impressive demonstration.

Compare proposals with the same questions

For a support-specific budget, our guide to customer service chatbot costs separates platform charges from setup, ongoing upkeep and the work left for people.

Before accepting an AI development quote, ask each supplier to answer these questions in writing:

1. What exact workflow, user group and usage volume does the price cover?
2. Who prepares the data, and what happens if it is incomplete or inconsistent?
3. Which systems can the feature read from or write to, and who supplies access?
4. What evidence is required for acceptance, including failure and fallback cases?
5. Which costs recur, and what usage assumptions sit behind them?
6. Who pays for licences, model accounts, infrastructure and human review?
7. What do we receive at handover, and what is included in support?
8. How are scope changes approved, and when can either side stop after a phase?

For an agent, clarify the actions and integrations. For a support chatbot, clarify the knowledge sources, channels and handover. For consultancy, clarify the decision and documents you receive. They can share a budgeting structure without being interchangeable purchases.

If you are still buying advice or discovery, compare published AI consulting fees separately from the implementation budget.

Bring a representative input, the desired output, approximate monthly volume and a list of connected systems to the first conversation. Discuss an AI software project with Pixelfield to establish the scope and projected operating cost before committing to the build.