Artificial intelligence in business has stopped being a question of whether and started being a question of where. Around 35% of UK businesses with 10 or more employees now report using at least one AI technology, up from roughly 12% in late 2023 (ONS Business Insights and Conditions Survey, June 2026). Globally the figure is higher still, with 88% of organisations reporting AI use in at least one business function (McKinsey State of AI, 2025).
Adoption is not the same as value. Research from MIT's Project NANDA found that 95% of generative AI pilots produced no measurable impact on profit and loss. RAND puts the broader enterprise AI failure rate above 80%, roughly twice the rate of comparable IT projects. Gartner, reviewing AI in infrastructure and operations in April 2026, found that only 28% of projects delivered the return they were funded to deliver.
Those two sets of numbers describe the same situation from opposite ends. Almost everyone is using AI. Very few are getting paid for it.
This article sets out what the evidence actually shows about the use of artificial intelligence in business, why most initiatives stall, and what the small number of organisations seeing returns are doing differently.
Key points
- UK adoption has almost tripled since 2023, but depth has barely moved. The average adopting business uses 1.6 AI technologies, up from 1.4 (ONS, June 2026).
- Failure is organisational, not technical. Poor data, undefined success measures and weak workflow integration account for most of it.
- The highest returns are showing up in back office and operational functions, not in the sales and marketing pilots that attract most of the budget.
- Regulation in the UK is not a single AI law. It is UK GDPR, sector regulators and an incoming ICO statutory code on AI and automated decision making.
- The businesses seeing returns defined the outcome before they chose the tool.
What artificial intelligence in business actually means
Artificial intelligence in business is the use of systems that perform tasks normally requiring human judgement, applied to a commercial process with a measurable outcome attached.
That definition matters because it excludes a lot of what currently passes for AI usage. A team member pasting text into a chatbot is using AI. A business is not. The difference is whether the output feeds a process that someone owns, measures and improves.
In practice, most artificial intelligence business activity falls into four categories:
- Generation. Producing text, images, code or structured content. The most widely adopted category in the UK, with 18% of businesses using large language models and 16% using visual content creation tools (ONS, June 2026).
- Classification and extraction. Reading documents, invoices, emails or forms and turning unstructured input into structured data.
- Prediction. Forecasting demand, churn, maintenance windows or credit risk from historical patterns.
- Agency. Systems that take actions across multiple steps, such as triaging a support queue or reconciling records between two platforms.
Categories two and three have been in commercial use for over a decade under the name machine learning. Category one arrived at scale in 2023. Category four is where most current investment and most current disappointment sits.
How many businesses are using AI
The honest answer is that it depends what you count.
| Source | Measure | Figure |
|---|---|---|
| ONS BICS, June 2026 | UK businesses with 10 or more employees using any AI | ~35% |
| ONS BICS, June 2026 | UK businesses with 250 or more employees | ~49% |
| McKinsey State of AI, 2025 | Global organisations using AI in at least one function | 88% |
| Eurostat, 2025 | EU enterprises using AI technologies | ~20% |
| ONS BICS, June 2026 | Information and communication sector | 58% |
| ONS BICS, June 2026 | Construction sector | 13% |
The spread between surveys is not a sign that one of them is wrong. They measure different populations with different thresholds. Official statistics that count all firms above a size band produce lower numbers than surveys of self selecting business leaders. When you see a headline AI usage figure, the first question worth asking is who was counted.
What every dataset agrees on is direction and shape. Adoption is rising quickly. Larger firms lead. Digitally mature sectors are pulling away from the rest. And depth is lagging badly behind breadth: the average UK adopter has moved from 1.4 AI technologies to 1.6 in nearly three years, which is close to no movement at all.
That gap between adoption and depth is the real story of artificial intelligence and business right now. Most organisations have arrived. Very few have unpacked.
Why most AI projects fail
The failure research is unusually consistent. MIT, RAND, Gartner and S&P Global all point at the same causes, and none of them are about model quality.
The problem was never defined. Projects start with a tool and search for a use, rather than starting with a cost, a delay or an error rate and asking whether AI addresses it. Gartner's own analysis attributes much of the failure rate to initiatives that were overly ambitious or poorly scoped.
The data was not ready. Gartner forecast that 60% of AI projects unsupported by AI ready data would be abandoned through 2026. If your customer records live in three systems with no shared identifier, no model will fix that. It will inherit it.
Success was defined after launch. The organisations that saw returns set profit and loss metrics before the build started. The ones that did not are now running pilots nobody can either kill or justify.
Nobody owned the workflow. A tool that sits alongside an existing process adds work. A tool that replaces a step in an existing process removes it. Most pilots do the former and are quietly abandoned once the novelty fades. S&P Global found that 42% of companies abandoned most of their AI initiatives in 2025, up sharply from 17% the year before.
The budget went where the returns were not. MIT found that most spend concentrated in sales and marketing, while the strongest measurable returns appeared in back office automation: finance, compliance, document handling and process reconciliation. Less visible work, better economics.
Where the use of artificial intelligence in business pays off first
The pattern across the businesses seeing returns is that they picked narrow, repetitive, high volume tasks with a clear right answer. That is unglamorous and it is where the money is.
Operations and administration
Invoice processing, document classification, data entry between systems that do not talk to each other, contract review, expense checking. These tasks are structured, repetitive and expensive in staff hours. They also have a clear definition of correct, which makes accuracy measurable rather than a matter of opinion.
Customer service
Triage rather than replacement. Routing enquiries, drafting first responses for human approval, summarising long threads before handover, surfacing the relevant knowledge base article. The businesses getting this wrong are the ones that put an unsupervised chatbot in front of customers and measured deflection instead of resolution.
Software development
Among the most mature areas of AI usage anywhere. Code generation, test writing, documentation and code review assistance are now standard in most development teams. The productivity gain is real but concentrated in boilerplate rather than architecture. It speeds up typing far more than it speeds up thinking.
Marketing and content
The most crowded category and the one where measurable returns have been hardest to prove. AI accelerates production. It does not, on its own, improve targeting, positioning or offer. Volume without a strategy simply produces more of whatever was already not working.
Analysis and forecasting
Demand planning, churn prediction, anomaly detection in financial data. These deliver well when the underlying data is clean and the historical pattern is genuinely predictive. They deliver nothing when either condition fails.
A practical approach to artificial intelligence and business strategy
Every framework worth following puts the business problem before the technology. Here is a sequence that reflects what the successful minority actually did.
1. Start with a cost or a constraint, not a tool. Identify a process where you can state the current cost in hours, pounds or error rate. If you cannot measure it today, you will not be able to prove improvement tomorrow.
2. Check the data before you check the vendor. Ask where the data lives, who owns it, how clean it is and whether it can legally be used for this purpose. Most stalled projects failed here and found out late.
3. Define the success measure in week one. A specific number with a date attached. "Reduce average invoice processing time from 11 minutes to 4 minutes by the end of Q2" is a target. "Improve efficiency with AI" is not.
4. Buy before you build. MIT's data showed specialist vendors succeeding at roughly twice the rate of internal builds, largely because internal projects underestimate integration effort. Build when the process is genuinely proprietary. Buy when it is not.
5. Redesign the workflow, do not decorate it. The value comes from removing a step, not from adding a clever one. If the process looks the same afterwards with an extra tool bolted on, you have added cost.
6. Keep a human in the loop where the decision matters. For anything affecting a customer, an employee or an applicant, human review is both good practice and increasingly a regulatory expectation.
7. Review at 90 days and be willing to stop. A pilot that cannot show movement against its stated measure after a quarter should be closed, not extended. The organisations with the worst AI economics are the ones that never cancel anything.
Governance and regulation for UK businesses
The UK has no single AI Act. As of mid 2026 there is no AI Bill before Parliament, and the framework rests on existing law applied by existing regulators. That does not mean there is nothing to comply with. It means the obligations are spread out.
The main threads a UK business should be tracking:
- UK GDPR, as amended by the Data (Use and Access) Act 2025. The automated decision making rules were rewritten, moving from a near prohibition on solely automated decisions with significant effects to a conditions based approach. The practical effect is that the compliance task has shifted from arguing an exception to evidencing safeguards.
- The ICO statutory code of practice on AI and automated decision making. In preparation following regulations that came into force in May 2026. Expect it to carry real weight once finalised.
- Sector regulators. The FCA, Ofcom, CMA and MHRA are each embedding AI expectations into their own binding guidance rather than waiting for cross sector legislation.
- The EU AI Act. Applies extraterritorially. If you serve EU customers, assume it applies to those activities. High risk obligations were pushed back to December 2027 following the May 2026 agreement.
The practical first step for most businesses is smaller than it sounds: list every process where AI or automated logic influences a decision about a customer, an employee or an applicant. That list is your exposure. Most organisations have never written it down.
Alongside that, a short written AI policy covering approved tools, what data may and may not be entered into them, and who signs off new tools will resolve most of the day to day risk. Consumer grade chatbots handling personal data are the most common quiet problem in UK businesses right now.
What good AI usage looks like in a smaller business
Large enterprises dominate the adoption statistics, but the economics often favour smaller organisations. A firm of 30 people can change a process in a week. A firm of 3,000 needs a steering group.
For a small or mid sized business, the sensible shape is:
- One process, chosen because it is expensive and repetitive
- An off the shelf tool rather than a custom build
- A measurement taken before you start
- A named owner who is accountable for the number moving
- A decision point 90 days out
That is a project measured in weeks and hundreds of pounds, not months and hundreds of thousands. It also produces something more useful than a pilot: evidence about whether your data, your processes and your team are ready for the next one.
The businesses that will be ahead in two years are not the ones running the most experiments. They are the ones that finished a small thing, measured it, and used what they learned to choose the next one.
Common questions
What is artificial intelligence in business? It is the application of systems that perform tasks requiring human judgement to a commercial process with a measurable outcome. That covers content generation, document classification, forecasting and multi step automation. The defining feature is that the output feeds a process someone owns and measures.
How many businesses use AI in the UK? Around 35% of UK businesses with 10 or more employees reported using at least one AI technology in June 2026, up from around 12% in late 2023. For businesses with 250 or more employees the figure was approximately 49% (ONS Business Insights and Conditions Survey).
Why do most AI projects fail? The consistent causes across MIT, RAND and Gartner research are undefined problems, data that was not ready for production use, success measures set after launch rather than before, and tools bolted onto workflows rather than integrated into them. The failure is organisational rather than technical.
Where does AI deliver the best return? Back office and operational functions, according to MIT's analysis: finance, compliance, document processing and reconciliation. Most budget goes to sales and marketing, where measurable returns have been weakest.
Does the UK have an AI law? No. The UK regulates AI through existing law and existing regulators rather than a dedicated statute. UK GDPR as amended by the Data (Use and Access) Act 2025, sector regulator guidance, and a forthcoming ICO statutory code on AI and automated decision making are the main obligations. Businesses serving EU customers also need to consider the EU AI Act.
How much should a business spend on AI to start? Less than most expect. A first project should be a single process, an existing tool, a defined measure and a 90 day review. The purpose of the first project is to produce evidence, not transformation.
Is AI usage safe with customer data? It depends entirely on configuration. Consumer versions of common AI tools processing personal data usually create data protection problems around lawful basis, data minimisation and international transfers. Enterprise versions with appropriate contractual terms can be used compliantly, but still require a data protection impact assessment, vendor due diligence and a written internal policy.
Where to start
The gap between businesses using AI and businesses benefiting from it is wide, and it is widening. Closing it does not require more technology. It requires a clearly stated problem, honest data, and the discipline to measure the result.
That is the same sequence that applies to any technology decision. Your goals first, technology second.
If you are working out where artificial intelligence fits in your organisation, or you have a pilot that has stalled and you want an honest read on why, get in touch.
