DOI : 10.17577/The UK now has 5,800 AI companies, up 85% in two years, turning over £23.9bn between them, according to the Department for Science, Innovation and Technology’s most recent sector study. Headcount followed: 50,040 people in 2022, 86,139 by 2024. Choosing a development partner out of that crowd is harder than it was, because almost every one of those firms now describes itself as an agent company. Gartner is blunter about this than most buyers expect, reckoning that of the thousands of vendors claiming agentic AI capability, only about 130 are real. The distance between a demo that impresses a board and a system that survives its first audit is where most agent budgets disappear. Closing it is a discipline of its own, which is why Crunch-IS is a leader in AI agent development services, with delivery built around the evaluation and control layer that pilots routinely skip. Below is what actually separates the firms that reach production, and six UK companies working in the space.
The failure rate is the story
MIT’s Project NANDA examined 300 public AI deployments, surveyed 153 leaders and interviewed 52 executives for its 2025 report on enterprise AI. Of the pilots it looked at, 95% produced no measurable effect on profit and loss, against $30–40bn of enterprise spending. Gartner expects more than 40% of agentic AI projects to be cancelled outright by the end of 2027, on escalating costs, unclear business value or inadequate risk controls.
Both numbers get quoted constantly and misread just as often. Neither says agents do not work. They say most organisations buy them the wrong way round: a visible pilot, a good demo, no route into the back office. MIT found more than half of enterprise AI spending went to sales and marketing tools, while back-office automation, the category with the strongest measured returns, stayed chronically underfunded.
UK adoption is wide and shallow
Self-reported AI use among UK businesses with ten or more employees climbed from roughly 12% in late 2023 to roughly 35%, on ONS figures. In June 2026, 29% reported using at least one AI technology, eight percentage points up on the year before. Among employers with 250 or more staff it reaches 49%. Information and communication sits at 58%; construction at 13%.
The number nobody quotes is the one that matters for anyone buying agent development. The average adopting business uses 1.6 AI technologies, up from about 1.4 since late 2023. Adoption tripled. Depth barely moved.
Most UK firms bought a single tool, usually a language model for writing text, which is what 17% of businesses report doing, and stopped there. That is the real market: not greenfield, but a lot of organisations holding one shallow deployment and wondering why the second one is so much harder.
Five questions that separate builders from demo-makers
Procurement decks rarely ask these. They should.
- How do you evaluate the agent? If there is no automated evaluation suite and no regression tests against past failures, quality is being judged by whoever last looked at the output.
- What does one task cost at production volume? A model that is cheap at 200 runs a month is a budget problem at 200,000. Ask for the arithmetic, not a range.
- What happens when the agent is confidently wrong? Agents rarely fail loudly. The answer should describe confidence thresholds, human handoff, and a log someone actually reads.
- Who owns the prompts, evaluations and orchestration code? If the answer is the vendor, switching later means rebuilding, and the quoted price is not the real price.
- What is in production right now? A named client, a live date, and what broke in the first month. Firms that have shipped can answer this in a minute.
AI agent development companies in the UK
These six are UK-headquartered and building agents that run in production, whether as client development work or as a platform an organisation deploys. They work in different places — regulated finance, legal, customer operations, public services — so the list is a map of the market rather than a ranking.
Crunch-IS
Builds AI agents as bespoke development work rather than licensing a product, with the agent, its prompts and its evaluation harness left in the client’s codebase. Integration and controls are treated as the deliverable, not the pilot. Suits organisations that have run a proof of concept and need the version that survives audit and volume.
PolyAI
London-based, founded in 2017 by three researchers from Cambridge’s Machine Intelligence Lab. It builds enterprise voice agents for customer operations and has since released an Agent Development Kit so client engineering teams can build and iterate on those agents in their own languages. A fit for contact-centre work where call volume, not document volume, is the problem.
Aveni
Edinburgh-based and deliberately narrow: its platform is trained on financial services data and built for UK FS workflows. That focus is the point, because a generic model applied to advice calls and suitability checks creates compliance exposure rather than removing it. Relevant to firms whose agent use case sits inside FCA-regulated processes.
Luminance
Founded by Cambridge mathematicians and working on legal documents. In October 2025 it launched Agent Lumi, pairing its domain-specific legal model with tooling and memory so the system carries out contract tasks rather than just flagging clauses. Worth looking at for contract review, negotiation and due diligence at volume.
V7
London-based, originally founded as Aipoly in San Francisco in 2015 and rebranded after moving to the UK in 2018. Its V7 Go product runs agentic workflows over documents and visual data, which is the unglamorous middle of most back-office automation. Suits operations teams whose bottleneck is unstructured paperwork.
Kainos
Belfast-headquartered and listed on the London Stock Exchange, reporting revenue of £431.1m for its 2026 financial year, up 17%, with bookings of £505.3m. Healthcare work grew 55% to £74.9m, and the firm has doubled its Responsible AI team. The obvious choice where a public-sector procurement framework and a long compliance review are part of the job.
The boring part is the product
The firms that get agents into production are not the ones with the best models. They are the ones that treated evaluation, cost per task, failure handling and integration as the actual build, and the model as a component. Gartner’s 130 figure is harsh, but it points at something real: most of the market is selling the demo. A buyer who asks the five questions above will find out which half of it they are talking to before signing anything.
