# Generative AI development company UK for your own documents and data

Canonical: https://appagentix.co.uk/services/generative-ai-development/

Our generative AI development covers assistants, document readers and search over your own files. You see a price in pounds and the monthly running cost before you commit, and a test report on your own cases before launch.

- Proof of concept in 3 to 5 weeks, production feature in 8 to 14 weeks
- A filled example of the test report is on this page, misses included
- The monthly running cost on the quote: a low, an expected and a high month
- Code, prompts, test set and accounts in your name

## How much does it cost to build a generative AI?

One production feature is £15,000 to £45,000 over 8 to 14 weeks, before VAT. A proof of concept takes 3 to 5 weeks on sample data. As of November 2026, one UK agency that prints figures quotes £8,000 to £25,000 for a proof of concept and £25,000 to £80,000 for a production feature. Both exclude VAT, model usage and hosting. A platform with several use cases on one retrieval layer, with user roles and an admin screen, takes 4 to 8 months. Data volume, connected systems and personal data move the price most. Model usage is billed to an account in your name, and the quote shows the monthly estimate. Price your own generative AI development project with the calculator on this page.

- Discovery: 2 weeks, comes off the build price
- Single-prompt assistant: 3 to 4 weeks
- Proof of concept: 3 to 5 weeks
- Production feature: 8 to 14 weeks
- After launch: A monthly support plan, model usage on your account

## Generative AI development services: what we build

Text generation is the AI that UK firms use most often. In September 2026, 21% of those with 10 or more staff used large language models for it (ONS Business Insights, October 2026). Each card names the data the build needs and the promise that goes into the scope.

### An assistant that answers from your documents

Staff or customers ask in plain English and get an answer, with its page reference, from your policies, manuals, contracts or help centre.

Every answer shows its passage, or says that nothing relevant was found.

### Document reading and data extraction

Invoices, claim forms, contracts or CVs become fields in your finance, case or HR system. We check them against about 50 past documents you have marked.

Accuracy is reported field by field, never as one headline percentage.

### Drafting inside a workflow

First drafts of replies, reports or tender answers in your house style, built from ten to twenty examples you have approved.

Drafts are scored against those examples before launch.

### Summaries and triage of long text

Call notes, support threads and case files are cut to the points a person needs, with urgent items flagged. We score them against good past summaries you supply.

The test report lists what each summary left out.

### Search across your own files

Search that matches meaning and follows each person’s access rights, so a question about a notice period finds the clause, whatever its wording.

Nobody gets a result from a file they cannot open.

### Generative AI app development

A feature in your mobile or web product, such as a writing helper that works from your own listings. One team builds the app side and the model side.

The response-time target is in the scope and measured on a mid-range phone.

## From documents to checked output

Generated text is checked before anyone relies on it.

1. Your documents
2. Retrieve
3. Generate
4. Check
5. Person approves

## Four LLM development routes, and what each adds to the price

Each route below costs more than the one above it, so we pay for the next one only after a test on your own cases fails. The calculator on this page prices each route.

What does the model need that it does not already have?

- If the task needs no private facts, such as drafting or summarising text, choose **prompting a hosted model**. It is the smallest build of the four, and its main risk is a confident answer from general knowledge.
- If answers must come from your policies, products, contracts or case files, choose **retrieval over your documents**. It adds a line for loading and searching your files. Each answer shows its passage, and new files need no retraining.
- If a test shows prompting and retrieval still miss a fixed style or format, choose **fine-tuning a model**. It adds about two weeks for data preparation and training, and you repeat it when your data changes.
- If a law or contract stops the data leaving your own network, choose **an open-weight model on your servers**. It adds hosting, security and operations to the price, while open-weight models can trail the best hosted ones.
- If someone quotes you for a new foundation model trained from scratch, choose **the test result first**. That is a research programme, not a service we sell, so ask which test result proves the four routes above fall short.

## The test report you read before launch

A filled example for a policy and contract assistant, where staff at a 120-person UK company ask questions of 600 policies and supplier contracts. Built as one production feature, it comes to a fixed £28,500 excluding VAT, and the calculator on this page prices your own version. The report is not a client record, and it shows one fixed fault and one open item.

**Test report: policy and contract assistant, build 0.9**

| Check | Cases | Pass mark | Result | Notes |
| --- | --- | --- | --- | --- |
| Correct answer with its passage shown | 100 | 95 of 100 | Pass: 96 of 100 | Four misses listed; three trace to poorly scanned pages |
| Asked for legal advice | 15 | 15 refused | Pass: 15 of 15 | Each reply points to the legal team |
| Asked about another employee’s record | 20 | 20 refused | Pass after fix: 20 of 20 | First run answered 2 from a shared HR folder. Access by role fixed, then all 20 re-run |
| Instruction hidden in a PDF | 10 | 10 ignored | Pass: 10 of 10 | Hidden text was read as content, never as an order |
| Out-of-date policy still in the folder | 10 | 10 of 10 | Open: 8 of 10 | Two answers quoted an old travel policy. Build 1.0 archives old files |
| Clause near the end of a long contract | 15 | 13 of 15 | Pass: 14 of 15 | One miss, on page 131 of 140 |

Every pass mark was written down before testing began. A fresh batch of 40 questions, never used for tuning, then scored 39 against a mark of 38. The open row names the build that fixes it, and you decide whether launch waits.

## How a grounded answer is made

One question, followed from the search to the monthly re-test.

1. **Question arrives.** A member of staff asks in plain English
2. **Search your documents.** Only files that person may open are searched
3. **Model drafts.** It writes from the passages found and quotes one
4. **Support check.** No supporting passage means the assistant says it does not know
5. **Person reviews.** Low-confidence answers wait for a named reviewer
6. **Monthly re-run.** Flagged answers join the test set and fixes change the next search

The last step feeds the first: the loop repeats.

## How is generative AI developed?

It starts with a free 30-minute call and a written view on fit, scope and price range within two working days. A project team is usually 4 to 6 people. This plan is one production feature over 12 weeks, while a proof of concept takes up to 5 weeks on sample data, not connected to live systems. The test set comes in week 1, the model comparison in week 2, a prototype in weeks 3 and 4 and the report in week 5.

### Discovery (Weeks 1 to 2)

We interview the people who do the work, read real inputs and agree what a good answer looks like.

You receive: A written scope, a data and risk note and one fixed quote.

### Build the test set (Weeks 2 to 3)

You supply 50 or more real cases, we add awkward ones, and someone on your side marks the correct answer.

You receive: The test set and the pass mark, agreed in writing before a model is chosen.

### Model bake-off (Weeks 3 to 5)

Two or three candidate models run on the test set, and the cheapest one that passes wins.

You receive: A comparison table with each model’s accuracy, speed and cost.

### Build the feature (Weeks 5 to 10)

Each two-week sprint ends with a version you can use. Logging, a flag button and the personal-data safeguards go in.

You receive: A working build every two weeks, with the code in your repository.

### Test run (Week 11)

The feature runs against the full test set and a fresh batch, and we try to break it.

You receive: The test report: pass rate, every miss and the monthly cost estimate.

### Launch (Week 12)

A staged release, first to a small group and then to everyone.

You receive: A live feature, a runbook and a training session for its users.

### Run and improve (Monthly)

The test set runs again each month, because models and your documents both change.

You receive: A monthly report on accuracy, flagged answers, usage and model cost.

## Estimate your generative AI feature in a minute

Pick what the feature does and who uses it, and the range and the build time update as you go.

Scope, a 50-case test set, one fixed-prompt assistant on a web screen, the test report and launch: £8,000 to £12,000, 3 to 4 weeks.

### What it does

- Summarise, sort or triage text: included
- Answer questions with the passage: £1,000 to £2,000
- Draft replies in your house style: £2,000 to £2,500
- Read documents into fields: £3,500 to £6,000

### Where its knowledge comes from

- Only what each user gives it: included
- A fixed library of your documents: £7,000 to £13,000
- Changing documents, synced daily: £10,000 to £17,000

### Who uses it

- One team, up to 50 people: included
- The whole company, access by role: £2,500 to £4,500
- Customers on your website or app: £4,000 to £7,000

### Where it runs

- A hosted model, region on the quote: included
- UK data region, where offered: £1,000 to £2,000
- An open-weight model on your servers: £20,000 to £30,000

### Add-ons

- Link to one system you already run: £2,500 to £4,000
- A second connected system: £2,500 to £4,000
- Review queue: a person approves: £2,500 to £3,500
- Larger test set: 150 cases: £1,500 to £2,500
- Inside your existing app: £3,000 to £5,000
- Scanned pages and handwriting: £2,500 to £4,000
- A second language: £1,000 to £2,000
- Usage and accuracy dashboard: £2,500 to £3,500
- Fine-tuning on your examples: £8,000 to £20,000
- Sign-in with Microsoft or Google: £1,500 to £2,500

An estimate, not a quote, excluding VAT. Model usage and hosting are billed to your own account, so they are not in this figure. The quote shows a low, an expected and a high month. Past the single-prompt assistant, the fixed price follows a two-week discovery, and its fee comes off the build price.

## What it costs to run each month

The bill has four parts: model usage, hosting, monitoring and our support fee. For the example assistant, we assume an expected month of 3,000 questions from 120 staff. Each question reads about 6,000 tokens and writes about 500. Only the support line is paid to us. Usage is counted in tokens, which are small pieces of text, and during the model comparison we count the tokens your real cases use.

The quote then shows a low, an expected and a high month at each provider’s price list. You set a spending cap on the account, and we never mark up model usage.

## How we test that the answers are right

More than 95% of our projects are delivered on or before the agreed date. A demo shows a model at its best, while a test shows how often it is wrong, so we agree these checks before the build.

- **A pass mark written down before testing.** Say the mark is 95 correct answers out of 100. If the result falls short, we report it, fix the cause and leave the mark where it was.
- **Your cases, then a fresh batch.** At least 50 real inputs, marked by someone who does the work, then a second set nobody tuned against.
- **A check that the passage supports the answer.** Any answer drawn from your documents that its quoted passage does not back counts as a failure.
- **Refusal and hidden-instruction tests.** We ask for things the assistant should decline, hide instructions in files it reads, and list every case where it complied.

## What is not included in these prices

Surprises on an AI invoice usually come from these items, so they are listed on the quote as well.

- **Model usage and cloud hosting.** Billed by the provider to your account, with the monthly estimate on the quote.
- **Cleaning messy source documents.** Re-typing or de-duplicating scanned pages is quoted as a separate line.
- **Legal advice and data protection documents.** Your solicitor or data protection lead writes those, and we supply the technical description.
- **Integrations outside the scope.** Each extra system is quoted before it is built.
- **Staff time on your side.** About two hours a week to mark test cases and review builds.

## When generative AI is the wrong tool

Generative AI suits messy text, speech or images where a person can check the output quickly. Finding three clauses in a forty-page contract fits, while a VAT return does not, because ordinary code gets an exact answer every time for less.

Before any build we ask who does the task today and what a wrong answer costs. We also ask whether a search box or a better form would do. A free tool may do: 47% of UK firms able to estimate AI spending spent nothing in 2025 (DSIT AI adoption survey 2026). If one would, or you have no past examples of a good answer to test against, we say so in discovery.

Many pilots stall because nobody wrote down what a correct answer is, which is why the pass mark comes first. Use also stays shallow: only 10% of UK firms with 10 or more staff that use AI use it extensively (ONS, July 2026).

## UK data rules for generative AI development

We are not your lawyers, so we build the technical side and flag the questions your data protection lead or solicitor needs to answer.

- **Personal data in prompts and documents.** If prompts or documents contain personal data, you are normally the controller. The ICO says the provider can be a processor, a joint controller or a controller, depending on who decides why and how data is processed. We list which providers see which data and check each contract, so your data protection lead can settle the roles.
- **A data protection impact assessment.** The ICO states that in the vast majority of cases, using AI on personal data is likely to result in a high risk to people. That triggers the legal duty to carry out a DPIA, and you assess this case by case. We prepare the technical description of the data flow that goes into it.
- **Where the data is processed.** We choose a hosting region for the database and, where the provider offers it, for the model. Where processing happens outside the UK, we flag the transfer mechanism you will need. The region is stated on the quote.
- **Decisions about people.** Since the Data (Use and Access) Act 2025, Articles 22A to 22D of the UK GDPR cover significant decisions taken with no meaningful human involvement. Extra restrictions and safeguards apply to them. We flag these in discovery and keep a person in the loop by default.
- **ICO guidance on AI.** The Information Commissioner’s Office guidance on AI and data protection was last updated in March 2023. The ICO now marks it as under review because of the Data (Use and Access) Act 2025. We check the current version at scoping and use it as the design checklist for fairness, transparency and accuracy.
- **Selling into the EU.** The EU AI Act (Regulation (EU) 2024/1689) sets risk-based rules for AI systems placed on the EU market or used there, depending on what the system does. If EU residents will use the feature, we flag the question for your adviser in discovery.

## Who owns what we build

Most buyers leave ownership until the contract stage, but we write it into the statement of work first.

### Code, prompts and configuration

Committed to a repository in your organisation from the first day, where we work as contributors.

You can remove our access at any time.

### The test set and scoring scripts

The cases, marking guide and scripts are yours, so any other supplier can check our work or replace us.

Delivered at the end of every sprint, not only at project close.

### Cloud and model-provider accounts

The cloud account, model keys and search database are created in your name, and the bills come to you.

No account sits under AppAgentix, so there is no handover day.

### Fine-tuned models and training data

Any weights we train and the cleaned training set are yours, with notes on how to export or recreate them.

The statement of work names every artefact and its holder.

## Which AI is best for generative AI?

We are not tied to one provider, and the usual candidates are hosted models from OpenAI, Anthropic and Google, also sold through the major cloud platforms. Llama, Mistral and other open-weight models are on the list too. Large language models lead the field: 77% of UK businesses using or planning to use AI named LLM-based approaches (DSIT, 2026).

The code treats the model as a setting, so if a provider changes a price or retires a model, we switch, re-run the test set and report the result.

## Questions buyers ask about generative AI development

### How much does it cost to build a generative AI?

It depends on the route. Prompting a hosted model is the smallest build, and retrieval, fine-tuning and private hosting each add to it. Each price is for a build on an existing model, and we do not train new models from scratch. The calculator on this page prices your own version.

### How much does it cost to build an AI app?

An app with one AI feature costs the app price plus the feature price, fixed as one figure after a two-week discovery. With us one generative AI feature adds £15,000 to £45,000 to the app price, excluding VAT.

[See what an app costs to build](/app-development-cost/)

### What are the best RAG solutions?

For you, the best retrieval-augmented generation (RAG) setup is the one that passes a test on your own documents. A build combines a searchable store of your documents, a model and a check on each quoted passage. We compare two or three combinations and keep the cheapest that passes.

### Are there any UK-based AI companies?

Yes. The government’s Artificial Intelligence sector study 2025 identified 10,488 AI companies in the UK, up 79% on 2024. Whoever you shortlist, ask for a price in pounds and a test on your own cases. Ask for a plan for UK data rules and written ownership terms too.

### Is RAG still relevant in 2026?

Yes, for answers that must come from your own documents. Google’s Gemini guide says many models now take 1 million or more tokens, but each request pays for every token sent. Retrieval sends only the passages needed and limits each person to files they may open. For a few short documents, pasting them into the prompt can be enough.

## Related services and prices

- [Agentic AI development](https://appagentix.co.uk/services/agentic-ai-development/)
- [AI chatbot development](https://appagentix.co.uk/services/ai-chatbot-development/)
- [Hire AI developers UK](https://appagentix.co.uk/services/hire-ai-developers-uk/)
- [MVP development](https://appagentix.co.uk/services/mvp-development/)
- [App maintenance and support](https://appagentix.co.uk/services/maintenance-support/)
- [App development cost in the UK](https://appagentix.co.uk/app-development-cost/)
- [AI education system development](https://appagentix.co.uk/industries/ai-education-system-development/)

By [Jainendra Ojha](https://www.linkedin.com/in/jaizendra/), Co-Founder and CEO, AppAgentix. Reviewed November 2026.

[Book a call](https://appagentix.co.uk/requestaquote/)
