The number of businesses using artificial intelligence is rising quickly.
That does not necessarily mean the number of businesses achieving meaningful productivity improvements is rising at the same rate.
Buying access to an AI tool and changing how an organisation operates are two different activities.
One introduces a capability.
The other requires an understanding of work, information, responsibility and commercial outcomes.
For businesses evaluating AI investment, the central question should therefore not be which model or platform is most sophisticated.
It should be:
What work can be improved, and what evidence would demonstrate that the improvement is worthwhile?
AI adoption has accelerated across UK businesses
The Office for National Statistics' Business Insights and Conditions Survey provides an indication of how widely AI has entered UK organisations.
In its 8 October 2026 release, the ONS reported that 30% of businesses had used at least one AI technology in September 2026.
That compares with 9% when the relevant question was first introduced in September 2023.
Among businesses employing ten or more people, adoption reached 38%.
The differences by size were notable.
Around 51% of businesses with at least 100 employees reported AI use, compared with 29% of businesses employing fewer than ten people.
These are survey-based figures describing reported adoption. They do not establish the depth of implementation or the resulting financial benefits.
Source: Office for National Statistics, 8 October 2026
The distinction matters.
A company using AI to draft occasional emails is counted as an adopter.
So is a company that has integrated AI into a substantial operational process.
Their underlying capabilities and economic effects can be very different.
Adoption is not the same as integration
A separate ONS analysis published in July 2026 examined AI use among UK businesses.
It found that, among businesses with ten or more employees, reported AI adoption had risen from approximately 12% in late 2023 to around 35% by 2026.
However, the average number of AI technologies used by adopting businesses had increased only modestly, from approximately 1.4 to 1.6.
The ONS characterised adoption as relatively shallow.
This suggests that much of the growth has involved businesses beginning to use AI, rather than necessarily transforming several established operational processes.
The 2026 UK Business Data Survey provides another relevant finding.
Among surveyed businesses using AI, only 21% reported that their AI tools were integrated into existing business systems.
The finding is not a universal measure of technical maturity, but it highlights the distance between accessing AI functionality and embedding it into everyday operations.
Source: ONS, Artificial Intelligence in UK Businesses
Source: UK Business Data Survey 2026
For many organisations, the initial stage of adoption is therefore likely to involve experimentation with accessible tools.
The next challenge is understanding which experiments deserve to become dependable business processes.
Productivity research shows both potential and limitations
Some of the clearest evidence comes from research into AI-assisted customer support.
Erik Brynjolfsson, Danielle Li and Lindsey Raymond studied the introduction of a generative AI assistant using data from 5,179 customer support agents.
Their research, published in The Quarterly Journal of Economics in 2025, found that access to the tool increased productivity, measured by issues resolved per hour, by approximately 14% on average.
However, the benefits were not evenly distributed.
Less experienced and lower-skilled workers experienced substantially larger improvements, approximately 34% in the working-paper analysis, while the effect on highly experienced workers was minimal.
The researchers suggested that AI assistance may help distribute useful practices and knowledge throughout an organisation.
That is commercially significant.
It indicates that AI's value may come from making existing expertise more accessible and operationally consistent, rather than simply replacing human tasks.
The research concerned a particular customer support environment. It should not be interpreted as proof that introducing AI will raise productivity by 14% in any industry.
Source: Brynjolfsson, Li and Raymond, Generative AI at Work
The real unit of analysis should be the workflow
Businesses often evaluate software by examining individual features.
Can the tool write emails?
Can it summarise documents?
Can it answer customer questions?
Can it connect to a database?
These capabilities may be useful, but they do not independently establish a business case.
A more informative approach is to examine the complete workflow.
Consider a company that receives product enquiries from distributors.
An employee may need to read the message, identify the product, check the relevant documentation, examine availability, establish appropriate terms, prepare a response and update internal records.
An AI tool capable of drafting the response solves only one part of that sequence.
If the employee still needs to gather information manually from five systems, the overall time saving may be limited.
A properly designed system might instead retrieve approved product information, prepare a draft response, flag missing details and create an internal record.
A human can then review and approve the result.
The value comes from coordinating the workflow, not simply adding a language model to it.
A simple financial model for evaluating automation
Suppose a business processes 2,000 routine administrative cases per month.
The average case requires eight minutes of staff time.
That represents approximately 267 hours of work each month.
Assume a proposed system reduces the average handling time to five minutes, including necessary human review.
The difference is three minutes per case.
Across 2,000 cases, the business saves 6,000 minutes, or 100 hours per month.
At an illustrative fully loaded labour cost of £25 per hour, that represents £2,500 of released staff capacity each month.
This is not automatically a £2,500 cash saving.
If employees remain on the payroll, the benefit may arise through additional output, shorter backlogs or improved service rather than reduced expenditure.
The system also has costs.
It may require development, integration, maintenance, model usage, information security controls and ongoing quality monitoring.
Suppose those costs total £900 per month after implementation.
The illustrative gross value of released capacity exceeds ongoing system costs by £1,600 per month, before allowing for initial investment, residual errors and changes in workload.
The economic case becomes stronger if the released time can be redirected towards genuinely valuable work.
It becomes weaker if the automation frequently fails or merely transfers administrative effort elsewhere.
Why automating a poor process can make it worse
There is a common implementation mistake.
A business identifies a repetitive task and immediately begins automating it.
But repetition does not establish that the task is necessary.
Some processes exist because departments use incompatible systems.
Others survive because nobody has reviewed the original requirement.
A team might manually transfer information from one spreadsheet to another each week.
An automation could perform that transfer in seconds.
But if the same information can be accessed directly through one reliable system, eliminating the transfer may be more effective than automating it.
The distinction is between process improvement and process acceleration.
A faster inefficient process may remain inefficient.
This is why AI implementation should begin with an operational assessment.
What is the task intended to accomplish?
Which information is required?
Where does it originate?
Who is responsible for decisions?
Which actions genuinely need human judgement?
What happens when the system produces an incorrect result?
These questions should determine the solution.
The choice of technology comes afterwards.
Reliability matters more than the demonstration
AI prototypes are often impressive because they demonstrate a successful outcome under controlled conditions.
A production system must operate across a wider range of circumstances.
Inputs may be incomplete.
Documents may contain conflicting information.
External services may be unavailable.
A customer may submit an unusual request.
An AI model may produce an answer that appears convincing but is incorrect.
The business must decide which errors are tolerable and which require human intervention.
For a low-risk internal summary, occasional minor mistakes may be manageable through review.
For contractual commitments, financial calculations or sensitive customer information, the required controls are considerably stronger.
Good systems therefore need more than a capable model.
They need appropriate data access, validation rules, monitoring, escalation procedures and clear responsibility.
A useful automation is one the business can depend on.
What an AI implementation review should measure
Before building a system, an organisation should establish a baseline.
That might include the time required for each task, number of cases processed, error rate, cost and staff involvement.
Following implementation, the same measures should be collected.
A sound evaluation should consider:
| Dimension | Question |
|---|---|
| Productivity | Has the time required per completed task fallen? |
| Quality | Are outputs accurate and suitable for their purpose? |
| Reliability | Can the workflow handle exceptions and failures? |
| Capacity | Can the team process more work without disproportionate additional resources? |
| Financial value | Do the benefits justify development and operating costs? |
| Risk | Are information security and human oversight appropriate? |
Where possible, compare the new workflow with a control group or a comparable process that has not changed.
This helps distinguish the effects of the technology from seasonal demand, staffing changes or unrelated operational improvements.
The strongest AI strategy may involve building less
Businesses do not necessarily need bespoke AI systems for every requirement.
An existing product may already solve the problem effectively.
A workflow may be improved through better system configuration.
Some work may be more suitable for a specialist employee.
Other tasks may justify a custom application because the business has particular information, processes or integration requirements.
The commercially sensible decision is the one that produces the most appropriate outcome for the resources involved.
The rapid spread of AI makes this distinction increasingly important.
Technology is becoming more accessible.
The scarce resource is not necessarily access to an advanced model.
It is the ability to identify useful applications, implement them reliably and measure their contribution to the business.
The starting point should be the work that needs improving, not the availability of a new tool.