AI Development in the UK

About AI Development

Most AI projects fail in the same place: a demo works on ten hand-picked examples, then falls apart on real user input, real data volume, and a real invoice at the end of the month. We build the unglamorous parts that decide whether an AI feature survives contact with production: retrieval that returns the right context, evaluation you can run before every deploy, guardrails on what the model is allowed to do, and cost tracking per request so the bill is a number you chose rather than one you discover.

What the engagement includes

• Retrieval-augmented generation over your own documents and databases
• AI agents with tool access, scoped so they can only do what you allow
• Model Context Protocol (MCP) servers that connect assistants to your systems
• LLM integration into existing products: OpenAI, Anthropic, and open models
• Document extraction and classification for invoices, contracts and forms
• Evaluation suites so a prompt change cannot quietly break output quality
• Cost and token controls with per-request tracking and hard budget limits
• Chatbots and support assistants grounded in your own content, not the open web

Automated decisions under the new UK GDPR articles, training data copyright and the AI security code

In a UK AI build the first design question is whether the system decides anything about a person on its own, because since 5 February 2026 that has been governed by articles 22A to 22D of the UK GDPR, inserted by the Data (Use and Access) Act 2025. The old near prohibition has gone. Solely automated decisions with legal or similarly significant effects are now allowed in wider circumstances, but only with safeguards, and the ICO's draft guidance, updated on 31 March 2026 and put out for consultation, shows how specific those are. A decision is solely automated when there is no meaningful human involvement, and the ICO's examples of significant effects include automatic refusal of an online credit application, e-recruiting without meaningful human involvement and a fraud flag that freezes a bank account. Where the provisions apply, the person must be given information about their own decision rather than a repeat of the privacy notice: which factors contributed, whether profiling was involved and whether another controller's data influenced it, provided at the point of the decision with a way to make representations, obtain human intervention and contest it. The ICO adds that the system should be able to produce an audit trail of the key decision points. Two constraints are easy to miss. The new recognised legitimate interests basis cannot support these decisions, even in part, and a decision that uses special category data needs a condition under article 22B on top of articles 6 and 9. So we build the explanation as an output of the pipeline, storing the features, scores and model version behind each decision, and route contested cases to a reviewer with authority to change the outcome rather than to a queue that can only confirm it. The ICO says this guidance will inform a statutory code of practice on AI and automated decision-making. Whether a particular output counts as a significant decision is for your own counsel.

Training or fine-tuning on content you do not own is an open question in UK law rather than a settled one, and the government's report of 18 March 2026 is its latest word. Under the Copyright, Designs and Patents Act 1988 the data mining exception in section 29A covers non-commercial research only, so, as the report puts it, permission will generally be needed from a right holder before a work is used, which in practice means a commercial fine-tune on scraped articles, images or code needs a licence. The consultation had proposed a broad exception with an opt-out. The report, required by sections 135 and 136 of the Data (Use and Access) Act, records that most respondents rejected it and says the government will not reform copyright law until it is confident reforms meet its objectives. The one court ruling so far, Getty Images v Stability AI, held that a trained model could in principle be an article and so an infringing copy, even if trained abroad, but only if it contains or reproduces a copy of a work, and found no evidence that Stable Diffusion did. The judgment was under appeal when the report was published. Outputs carry their own uncertainty. Section 9(3) makes the author of a computer-generated work the person who made the arrangements for its creation, but the report calls the provision contradictory and proposes removing protection for wholly computer-generated works while keeping it for work made with AI assistance. In practice that shapes three things in a build. The training set is assembled from owned, licensed or openly licensed sources, with the licence terms recorded per dataset, because open licences often carry attribution or other conditions downstream. Data provenance is kept alongside model versions, so you can say what a given model learned from. And where generated output is the product you sell, a human editorial step is worth keeping. How this applies to your own data is for your counsel.

Security questionnaires from UK buyers now reach past the application to the model, and the usual reference is the AI Cyber Security Code of Practice, which DSIT published on 31 January 2025. It is voluntary, the government has put it forward to ETSI as the basis for the global standard TS 104 223, and it sets 13 principles across the life of a system, each assigned to developers, system operators or data custodians. That split matters in an outsourced project, because one supplier can hold more than one role. The Code counts a business that provides a contractual service to embed or deploy a model for an organisation as a system operator, so a team building and running your AI system is a developer and often an operator too, while your organisation remains the data custodian. Several provisions turn straight into engineering tasks. Developers shall keep an audit trail covering the operation and lifecycle management of models, datasets and prompts, so system prompts are versioned and reviewed like code rather than edited in a console. Development and tuning happen in dedicated environments with least privilege, and training data that is not public stays inside them. System operators shall tell end users in an accessible way where and how their data will be used, accessed and stored, for example whether it is used for retraining or reviewed by staff or partners, which means the terms of any model API provider have to answer that question before the interface can. Operators shall log system and user actions and should watch for drift and poisoning over time, and the Code names indirect prompt injection among the risks it is written for. When a model is decommissioned or handed to someone else, the data custodian is involved and the assets are securely disposed of, which settles what happens to fine-tuned weights and embeddings when an engagement ends. Whether the Code appears in your procurement is for your security lead to confirm.

How we work with UK clients

UK engagements tend to start with procurement questions rather than technical ones. Who holds the data, which law governs the contract, whether you can produce a record of processing if the ICO asks. We answer those in the proposal instead of leaving them to a legal review three weeks in, because that review is where UK projects usually stall.

Invoicing is the other thing worth settling early. We are outside the UK, so services we supply generally fall under the reverse charge and your finance team accounts for the VAT rather than paying it to us. Our invoices say so explicitly. We quote in pounds, the figure does not move with the exchange rate mid-project, and payment by bank transfer avoids the card fees that make a five-figure invoice unnecessarily expensive.

Indicative pricing in GBP

Converted from our published Indian rates and rounded. Exchange rates move, so the figure on your quote is calculated on the day it is issued.

Questions UK clients ask

AI Development: common questions

Other markets we work in