At a glance
AI adoption is rising, but many employees have not been trained to use it well. The practical response for a small business is not to buy several tools at once. Start by mapping where work is delayed, identify the teams and tasks involved, and train the people who will check the results. AI can reduce repetitive administration, but it still needs clear rules, human review and a modest first project. There is no single published cost, so budget for software, setup and staff time.
What has happened
The gap is not simply between businesses that use AI and businesses that do not. It is also between introducing a tool and having people who understand what it can do, where it can go wrong and how it fits into an existing process.
The Office for National Statistics reported in 2026 that its Business Insights and Conditions Survey analysis is tracking how UK businesses are applying AI and the factors shaping its uptake. Its more detailed 2026 article says that self-reported use among businesses with 10 or more employees rose from around 12% to around 35% between late 2023 and 2026. However, the average number of AI technologies used by adopting businesses rose only modestly, from around 1.4 to around 1.6.
That suggests adoption is often shallow. A firm may be using one tool for drafting, searching or image processing without changing how work moves through the business. The risk is that staff are expected to work out the new process for themselves.
Research published by the Department for Science, Innovation and Technology in 2026 found that adoption remains modest overall and varies by business size and sector. The decision for a small firm is therefore not whether it is behind a national race. It is whether one well-chosen task can be improved without creating more checking, confusion or risk than it removes.
Who this applies to
- Small businesses with around 5 to 50 staff that are considering AI, have already introduced a tool or suspect staff are using tools without a shared approach.
- Owners and managers who are losing time to enquiries, quotations, invoices, reminders, document handling or repeated data entry.
- Firms where one team is experimenting with AI but other teams do not know what is changing or who is responsible for checking the output.
- Businesses that need a practical starting point rather than a technical AI strategy.
- Sole traders are not directly represented in the 2024 employer survey cited below, so its findings should not be treated as evidence about them.
Map the skills gap in each team
Start with the work, not the software. Ask each team to describe the tasks that take time, are repeated frequently or regularly wait for information from another team. Record what the task involves, what information it uses, who checks it and what happens when it is wrong.
A useful skills map has three layers. The first is basic use: can a person give clear instructions to a tool, provide the right context and recognise an answer that needs checking? The second is job knowledge: does the person understand the customer, product, contract or process well enough to judge the result? The third is responsible use: does the team know what information must not be entered, when a human must approve an output and how an error should be reported?
The Department for Science, Innovation and Technology employer survey, carried out between 19 March and 7 June 2024 and published in 2026, found that 31% of employers used AI. It also found that 61% had no current staff working with AI, while only 11% had staff undertake AI training in the previous 12 months. These figures are more than a year old by August 2026, but they show why buying access is not the same as building capability.
Look at teams separately because their risks and useful skills differ.
| Team | Likely gap to check | Sensible first capability |
|---|---|---|
| Sales and enquiries | Responses are drafted inconsistently or leads are not followed up | Turning enquiry information into a clear next action, with a person approving replies |
| Estimating and quotations | Staff retype details or spend time finding previous information | Structuring customer requirements and checking that a quote contains the required details |
| Operations and engineering | Information is spread across job notes, emails and schedules | Finding relevant information and flagging missing or conflicting details |
| Finance and administration | Repeated invoice, reminder and data-entry work | Confirming source data, exceptions and approval points before anything is sent |
| Management | Little visibility of where work is delayed | Reading a simple dashboard and deciding what action follows |
Do not assume the most enthusiastic employee is automatically the right owner. Choose someone who understands the process and can explain its exceptions. A person who knows when a job, quote or invoice is unusual is often more valuable than someone who has tried the most AI tools.
Choose tasks that are ready for automation
The best early candidates are tasks with a clear beginning and end. They usually involve information that already exists in a consistent place, follow repeatable rules and have an obvious person who can check the result.
Examples include collecting website enquiries into one queue, sending a reminder when a quote has not been answered, preparing an invoice from approved job information, or showing managers which jobs are waiting for a next step. These uses do not remove the need for judgement. They reduce the chance that a routine step is forgotten.
Be more cautious when the task affects safety, employment, legal commitments, financial approval or a customer’s eligibility. A system that drafts a response may be useful. A system that sends an unreviewed commitment to a customer may not be. The distinction is whether the tool is assisting a decision or making it without a suitable control.
The ONS’s 2026 analysis found that improving business operations was the most common use of AI among larger businesses, reported by over 60% of them. It also reported effects on creative or design roles for over half of businesses using visual content AI, and similar effects on administrative or clerical roles when using image-processing AI. Those findings relate to larger businesses and specific uses, so they are not a direct forecast for a firm with 5 to 50 employees. They do indicate where process changes are already being observed.
Use this test before selecting a task:
- Is the work repeated often enough that saving time would matter?
- Does it use information the business already holds?
- Can the desired result be described in plain language?
- Can someone check the result before it creates a customer, financial or operational consequence?
- Is there a simple measure, such as time taken, missed follow-ups or late quotations?
- Can the business stop the process easily if the result is poor?
Tasks that fail several of these tests are not necessarily impossible. They are simply poor first projects. A complex quoting process, for example, may need data cleaning and agreement about pricing rules before any automation is useful.
The choice is usually between assistance, workflow automation and a dashboard. They solve different problems.
| Approach | What it does | Main trade-off |
|---|---|---|
| AI assistance | Helps a person draft, summarise or find information | Quick to try, but results and checking can vary |
| Workflow automation | Moves information or starts a routine action when a condition is met | More consistent, but depends on accurate source data and agreed rules |
| Dashboard | Brings pipelines, jobs, quotes and enquiries into one view | Improves visibility, but does not fix a process that nobody owns |
Do not measure success by how impressive the output looks. Measure whether a real bottleneck has changed. If the problem is lost enquiries, count enquiries received, assigned and followed up. If the problem is late quotes, measure the time from a complete enquiry to an approved quotation. If the problem is invoice administration, measure exceptions and rework as well as time saved.
Build a safe, manageable AI starting point
A manageable starting point has one business problem, one accountable owner and a defined review period. It does not require every employee to learn every tool. The owner should first agree what the project is meant to improve and what it must not do.
Write the process down before changing it. This can be a simple description of the current route from enquiry to quote, job to invoice, or quote to follow-up. Include the points where information is copied, where work waits and where mistakes are found. That baseline gives staff something concrete to compare with the proposed change.
Next, decide what information the system may use. Separate ordinary operational information from confidential customer, employee, financial or commercially sensitive information. Check the terms and settings of any chosen service, restrict access to people who need it and avoid putting sensitive material into a tool until the business understands how it is handled. A policy should be short enough to use. It should say which tools are approved, what information is prohibited or restricted, when checking is required and who to ask when something is unclear.
The Department for Science, Innovation and Technology's AI Adoption Research was published in January 2026 after research conducted with 3,500 businesses between 12 February and 2 May 2025, supported by 100 qualitative interviews. It says the survey asked the person responsible for technology oversight, and notes that this approach did not provide insight into shadow AI adoption. For a small firm, that is a reason to ask staff what they are already using, not a reason to assume that an approved tool is the only one in use.
Give the first project a narrow boundary. A sensible sequence is to test privately with old or non-sensitive examples, compare the result with the existing process, then run it alongside the current method for a short period. Only after the owner is satisfied should the new method handle live work. Keep a manual route available.
For each step, record four decisions in plain English:
- What starts the process?
- What should the system prepare, suggest or move?
- What must a person check?
- What happens when information is missing or the result looks wrong?
The last question matters most. An automation that works only when every record is complete will fail at the first unusual job. Build an exception route, and make it obvious to staff how to pause or correct the process.
The ONS reported in January 2026, using data collected in late December 2025, that 25% of businesses were using some form of AI and 15% planned to adopt one within the next three months. It also reported that 4% of businesses already using AI said their overall headcount had decreased as a result. These are reported survey results, not a promise of job reductions or savings for a particular firm. They support a measured approach: decide whether the aim is to release staff time, improve service, handle more work or reduce errors, rather than assuming automation should mean fewer people.
Keep the first implementation small enough that the owner can see what happened. A single enquiry route, quotation reminder or invoice check is easier to govern than a company-wide collection of disconnected experiments. If the process works, extend it one adjacent step at a time.
Train staff and measure whether it helps
Training should be attached to the task people actually perform. A short session on writing useful instructions is less valuable than practising how to prepare an enquiry summary, identify missing information and check a draft reply.
Everyone involved needs the same basic understanding, but not everyone needs the same depth. General users need to know the approved tool, the information rules and the checking process. Process owners need to understand triggers, exceptions and how to correct records. Managers need to know what the measures mean and when to stop a workflow. A small number of people may need deeper technical support, but that is not a prerequisite for a useful first project.
The 2024 employer survey published by DSIT in 2026 found that 21% of employers thought demand for AI skills was likely to increase over the following 12 months. It also found that 4% had tried to recruit someone to work with AI in the previous three years. Those findings are not a forecast of what every small firm should hire for, and the fieldwork is now more than two years old. They do suggest that training existing staff deserves consideration before creating a specialist role.
Measure three things: the outcome, the effort and the quality. Outcome might be response time or the number of overdue follow-ups. Effort includes setup, corrections and time spent checking. Quality includes errors, complaints and work that had to be redone. Take a baseline before the trial and review the results with the people doing the work.
Do not claim a saving simply because a draft was produced faster. If staff spend the saved time checking unreliable outputs, the process may not have improved. Ask employees what became easier, what became harder and which exceptions were missed. Their feedback is part of the measurement, not an optional opinion.
Where this falls short
- The published figures measure different groups and use different methods. The ONS 2026 article focuses on businesses with 10 or more employees, while the DSIT employer survey covered employers excluding sole traders.
- Adoption figures are self-reported and may not capture unapproved or informal use. DSIT's 2026 research specifically notes that its technology-oversight interviews did not reveal shadow AI adoption.
- A national adoption rate does not show what a particular trade, manufacturer, engineering firm or professional practice should automate.
- AI output can be inaccurate, incomplete or unsuitable even when it sounds confident.
- Time saved in one task may be absorbed by checking, data cleaning, setup or staff training.
- The sources report adoption and skills conditions. They do not establish a guaranteed return on investment for a small business.
Worked example
Illustrative example: a small engineering firm receives enquiries by email and through its website, then records them manually before preparing quotations.
The firm chooses one project for six weeks. In the first week, it maps the existing process and agrees that AI may summarise an enquiry and flag missing information, but a member of staff must approve every customer reply and quotation. In week two, it tests the process on old, non-sensitive enquiries. In weeks three and four, it runs the new method alongside the existing one. The final two weeks are used to review results and correct the rules.
For illustration, suppose the firm receives around 20 enquiries a week. Before the trial, staff record them in different places and some follow-ups are late. The target is not an invented claim about revenue. It is to reduce the time spent copying information and make every enquiry show an owner and next action.
The manager compares the baseline with the trial by recording response time, overdue follow-ups, corrections and staff checking time. If the process saves a few hours but creates repeated errors, it is stopped or redesigned. If it reduces missed next steps without weakening review, the firm can consider connecting the approved enquiry record to a quotation reminder.
The important result is a controlled decision, not the use of AI for its own sake.
Implevo's View
For most small firms, this is worth acting on when a clear administrative bottleneck already exists. Lost enquiries, late quotes, forgotten reminders and repeated invoice work are easier places to start than an open-ended attempt to “use AI”. The first investment should be understanding the process and involving the people who do it.
It is not worth adopting a tool merely because other businesses are discussing AI. The ONS reported in 2026 that adoption among businesses with 10 or more employees had reached around 35%, but also described adoption as relatively shallow. That is evidence of use, not evidence that every business needs another system.
We would watch whether staff training keeps pace with new workflows, whether shadow use becomes visible, and whether claimed time savings survive quality checks. If you want to examine one process without committing to a large programme, Implevo's Discovery Day is a low-pressure place to start.
References
- UK sectors split as AI adoption races ahead of workforce readiness, Small Business UK, 11 August 2026
- UK sectors split as AI adoption races ahead of workforce readiness - Small Business UK, "UK "professional services" (SME OR small firm) (AI OR automation OR productivit, 11 August 2026
- AI in UK businesses - GOV.UK, gov.uk, 20 July 2026
- AI Skills for Life and Work: Employer survey findings - GOV.UK, gov.uk, 28 January 2026
- AI Adoption Research - GOV.UK, gov.uk, 28 January 2026
- AI Adoption Research - GOV.UK, gov.uk, 28 January 2026
- Artificial intelligence in UK businesses: 2023 to 2026, ons.gov.uk, 20 July 2026
- Executive summary with introduction and next steps - GOV.UK, gov.uk, 29 October 2025
- Business insights and impact on the UK economy - Office for National Statistics, ons.gov.uk, 8 January 2026
- Release home - Employer Skills Survey - Explore education statistics - GOV.UK, explore-education-statistics.service.gov.uk, 24 July 2025