AI use differs by sector, business size and survey definition. The useful question is not whether your industry looks ahead or behind, but which process is worth improving safely.
01
An industry league table can make a sensible business panic
Imagine a regional manufacturer with 70 employees. Its office team still rekeys supplier information, chases missing quality records and assembles a weekly production report from several spreadsheets.
The directors read an AI adoption table. Technology businesses appear near the top. Manufacturing sits much lower. The immediate reaction is that the company has fallen behind and needs an AI strategy quickly.
That reaction is understandable. It is also how a useful market statistic turns into an expensive shopping list.
An adoption percentage can show where activity is concentrated. It cannot tell this manufacturer whether a chatbot, forecasting model, document reader or machine vision system fits the work. It cannot show whether the source data is reliable, whether staff will use the result or what a wrong answer could cost.
The percentage is context. The strategy begins inside the business, with one process and one result worth improving.
02
First check what the adoption number actually measures
AI adoption figures often look more precise than the underlying question. Current UK sources illustrate the problem well.
The Department for Science, Innovation and Technology reported in January 2026 that 16 per cent of UK businesses used at least one AI technology. Its representative survey included micro businesses and found higher use among larger firms and knowledge-intensive sectors.
Office for National Statistics analysis published in July 2026 reported that around 35 per cent of UK businesses with 10 or more employees used at least one AI technology. The UK Business Data Survey reported 41 per cent among businesses that handled digitised data. Different populations, definitions and survey methods produce different answers.
| Check | Why it changes the result | Example |
|---|---|---|
| Who was included? | Large firms tend to report more AI use than small firms | All businesses, firms with 10 or more employees, or only businesses handling digitised data |
| What counts as adoption? | Occasional use is different from a production system | One employee using a writing tool, a formal pilot, or an integrated live process |
| Which technologies count? | A broad definition captures more activity | Text generation, machine learning, computer vision, robotics or automated decisions |
| Who answered? | Leaders and staff may see different levels of use | A director may report company policy while employees report informal daily use |
| When was it measured? | AI products and workplace habits are moving quickly | Fieldwork date matters more than the date printed on a later article |
The supplied Presenc AI research uses another definition: at least one AI tool deployed in production. Its sector table is useful as a directional comparison, but its estimates should not be mixed with official UK figures as though every source measured the same thing.
Before repeating any adoption statistic, read the methodology. A smaller number with a clear definition is more useful than a dramatic number nobody can explain.
03
Industry differences describe the work, not the winner
Sector differences are still useful once the figures are treated carefully. They show where suitable data, repeatable decisions and digital systems make adoption easier.
The ONS found the information and communication sector had some of the highest reported AI use in June 2026, while construction was much lower. OECD data also shows higher adoption among information and communication businesses and professional services than across firms overall.
| Business setting | Useful starting point | Important boundary |
|---|---|---|
| Professional services | Search approved knowledge, prepare case summaries or classify enquiries | Advice and final decisions still need accountable professional review |
| Manufacturing | Read supplier documents, flag missing quality records or support maintenance planning | Poor sensor or document data will create unreliable output |
| Finance and insurance | Prepare evidence, detect unusual patterns or support document review | Access, explainability, regulation and approval controls are central |
| Retail and ecommerce | Classify support requests, improve product information or support demand planning | Customer data and automated pricing need clear limits |
| Construction and field services | Organise job evidence, extract form data or draft site updates | Connectivity, inconsistent records and safety decisions limit full automation |
That does not make a construction company less ambitious or a software company automatically good at AI. The work is different. A digital business already has searchable information, software teams and frequent data-led tasks. Work on a building site includes physical conditions, fragmented subcontractors, safety responsibilities and information that may arrive through photographs, calls and paper.
The right lesson is to look for the conditions behind the rate. Which information is already digital? Which decisions repeat? Where is there enough volume for automation to matter? Where must a qualified person remain responsible?
04
Measure depth, not just presence
A business can truthfully say it uses AI because somebody has a Copilot licence. That tells us almost nothing about operational change.
The ONS found that adoption among UK businesses with 10 or more employees had risen from around 12 per cent in late 2023 to around 35 per cent by June 2026. Yet the average number of AI technologies used by adopting businesses moved only from roughly 1.4 to 1.6. Only 10 per cent of adopting businesses reported extensive use.
The UK Business Data Survey found another gap. Among businesses using AI, 21 per cent said their tools were integrated into existing business systems. Larger firms were much more likely to report integration than smaller organisations.
This is the difference between access and adoption. Access means the tool is available. Useful adoption means it has a defined job, appropriate data, an owner, a review route and a measurable effect on the process.
For the manufacturer, an employee using AI to rewrite an email is use. A controlled service that reads supplier certificates, checks required fields and routes exceptions to the right person is deeper adoption. The second option may create more value, but it also needs better data, testing and ownership.
05
Choose the process before choosing the AI tool
Industry examples can help a team notice possibilities. They should not decide the implementation.
Start with the work people already perform. Map the steps from request to completion. Note where information is retyped, checked, summarised, matched, classified or chased. Then separate the stable rules from the judgement that depends on experience or context.
The manufacturer might discover that missing supplier records cause the most delay. The team receives documents by email, checks six required fields, updates the same system and asks the supplier for anything missing. That is a stronger starting point than a vague target to use predictive AI because other manufacturers do.
A simpler automation may handle some of the job without AI. File naming, required-field validation and routing can often use ordinary software rules. AI earns its place where the input varies enough that fixed rules struggle, such as extracting information from differently formatted documents or preparing a short summary for review.
06
Score one candidate use case honestly
A useful first AI project is valuable enough to matter and contained enough to learn from. It should not begin with a decision that can cause serious harm when the system is wrong.
| Factor | Question | Healthy signal |
|---|---|---|
| Process stability | Do people broadly agree how the work should happen? | The main steps and exceptions can be described without three conflicting versions |
| Data readiness | Is the necessary information available and lawful to use? | Inputs have known owners, formats, access rules and retention needs |
| Volume and friction | Does the task happen often enough to justify change? | The baseline shows repeated time, delay, error or missed follow-up |
| Consequence of error | What happens when the AI is wrong? | A mistake is visible, reversible and held for human review |
| System route | Can the result reach the existing record or next step? | The workflow does not end with somebody copying text from another window |
| Measurement | Can the test prove whether the process improved? | Time, completion, error, quality or response measures exist before the pilot |
Score the candidate against the process, data, consequence, review route and baseline. A fashionable use case with weak data should lose to a quieter one the business can measure and control.
07
Put ownership and data rules in place early
Governance should grow with the risk, but it should exist from the first useful test.
The UK Business Data Survey found that only 17 per cent of businesses using AI reported having formal or informal AI policies or guidance. It also found that 73 per cent of businesses handling digitised data felt uncomfortable with their data being used to train external AI models.
Those concerns are not a reason to stop. They are a reason to define which tools are approved, which data may be entered, where results are stored and who reviews them. Staff should know when they are using AI and what to do when the output is uncertain.
The manufacturer does not need a fifty-page policy before testing supplier-document extraction. It does need a named owner, approved data boundary, sample set, error log, human check and a way to stop the test without losing the original process.
08
Run a small test that can earn the next investment
A pilot should answer a business question, not merely demonstrate that a model can produce an impressive output.
For the supplier-record process, the test might use a representative set of historic documents with sensitive data handled appropriately. The team can compare extraction accuracy, review time, missed fields and exception handling against the current method.
Set the stopping rules before the test starts. If the output cannot be checked quickly, if important document types fail repeatedly or if the integration creates another disconnected queue, the design needs changing. A weak pilot is useful when it prevents a much larger commitment.
If the result is reliable and the team accepts the new route, connect it carefully to the live process. Monitor exceptions and keep the person responsible for the business decision visible. Scale the pattern only after the first workflow has earned trust.
09
Benchmark your progress against your own operation
AI adoption by industry is worth watching. It helps leaders understand where investment is concentrated, which uses are becoming ordinary and where their sector may face practical barriers.
It is a poor target on its own. A business does not become better because its tool count matches a sector average. It becomes better when a real process gets faster, clearer or more reliable without creating a larger risk somewhere else.
Return to the manufacturer. The useful decision was not how to catch the technology sector. It was whether one document-heavy process could be improved with controlled extraction and human review. The result can be compared with the company's own baseline, not somebody else's adoption headline.
Use sector statistics to ask better questions. Build the roadmap from the work your business is actually responsible for.
Useful questions
Before following an AI adoption trend, ask:
- Who was included in the survey and what counted as adoption?
- Does the sector comparison include businesses like ours in size, geography and digital maturity?
- Which real process are we trying to improve?
- Could ordinary automation solve the stable parts more simply?
- Is the necessary data reliable, accessible and lawful to use?
- Can a person review mistakes before they cause harm?
- What baseline will show whether the pilot improved anything?
- Who owns the workflow after the demonstration ends?
