New OECD figures show that large enterprises are more likely to use AI. For smaller firms, the practical challenge is to find a recurring problem, fit the tool into everyday work and measure what changes for staff and customers
In a small firm, a single customer enquiry can pass through the person who answers the phone, the person who knows how long a job will take and the owner who decides whether to accept it. Sometimes all three are the same person. A generative AI assistant can draft a reply in moments. Its contribution to the business depends on whether it helps that person assemble the right information, make a decision and keep a promise to the customer. That distinction between access to a tool and a change in the way work gets done is easy to miss when adoption is expressed as a percentage.
On 5 November 2026, the OECD reported that in 2025 an estimated 20% of enterprises with at least ten employees across the OECD used AI, compared with 52% of enterprises with at least 250 employees. The comparison is between larger firms and the full population covered by the estimate, which includes those larger firms. It cannot be read as a comparison with the self-employed or as evidence that large firms get more value from each implementation. It establishes a gap in reported use and leaves open what firms are doing with that access. OECD release and Digital Economy and Society Outlook 2026.
An earlier OECD survey helps explain why the distinction matters. Its non-representative sample covered more than 2,000 SMEs across 12 OECD countries. The authors describe continuing uptake of ready-made products alongside uneven integration into day-to-day operations. Respondents’ experience points to constraints such as time, skills and maintenance costs. The survey can illuminate those barriers without establishing how common each is across the whole SME population. For an owner weighing a purchase, the practical question is whether the firm has the capacity to build and maintain a reliable process around the tool. 2026 OECD D4SME Survey.
Start with the work that gets stuck
Imagine a small tailoring workshop that receives alterations and custom orders by email. This is an illustrative example. A customer may include a desired date and a description of a garment while leaving out measurements or the occasion that makes the deadline firm. The workshop spends time reading messages, identifying missing details and asking follow-up questions. A plausible first use of AI would be to turn each enquiry into a short intake record and draft a question about the missing information. A member of staff would compare both with the original message before responding. Prices, capacity and delivery commitments would remain with the people who know the work.
The strongest argument for such a pilot would come from a review of actual enquiries. If most cases already arrive complete, a templated response or a form could do the work with less upkeep. If descriptions vary widely and staff repeatedly have to turn them into the same categories, assisted extraction may be worth testing. The distinction also matters when the problem sits downstream. A faster reply makes little difference if every order waits for the same overbooked specialist. The owner needs to locate the delay that affects the customer and trace how the proposed change would alter it.
A narrow pilot can also surface hidden judgement. The workshop might discover that two apparently similar requests differ because one fabric is difficult to match or because a customer expects the original finish to be preserved. Those distinctions may never have appeared in the job record because an experienced employee could recognise them and ask the right question. In that case, adding a field or changing the intake conversation could be valuable in itself. The exercise gives the firm a clearer view of what it knows, where that knowledge lives and which parts it can safely ask software to organise.
Build a dependable handover
The minimum integration can be quite modest. A reviewer needs to see what the customer supplied, what the assistant inferred and which items still require confirmation. A field that says “deadline unknown” should remain visibly different from a proposed date. In the workshop example, an assistant could draft a record in a shared queue with a link to the original enquiry. A person would confirm it before it reaches the scheduling system. That boundary creates a clear place to correct an error and prevents an attractive draft from acquiring the status of a business commitment simply because it arrived in a familiar format.
Responsibility for the underlying information needs equal attention. A reference sheet may contain approved materials, usual turnaround times and the conditions that require a specialist review. Someone has to decide when that sheet is out of date and who can update it. If the assistant has access to customer messages, the firm also needs to understand how those messages are stored, who can see them and whether the supplier can reuse them. A small initial test can use appropriately cleared examples and give the tool only the information necessary for the task. Expanding access should follow a specific operational need.
Large companies can sometimes assign separate teams to integration, procurement and security. A microbusiness often has to fit those responsibilities into an owner’s working week. That can favour a small, understandable arrangement with a clear manual fallback. For a sole trader, the fallback may be a familiar inbox and an order spreadsheet. For a larger operation, it may involve rolling back an integration while preserving its audit trail. In both cases, the ability to resume work after an outage or a poor software update has a cost, and that cost belongs in the decision to adopt.
Measure the completed job
Published workplace research shows how much the outcome depends on the task and the people using the tool. Erik Brynjolfsson, Danielle Li and Lindsey Raymond studied a staggered rollout of a conversational assistant involving 5,172 customer support agents. Their published study found an average increase of 15% in issues resolved per hour, with larger gains among less experienced and lower-skilled agents. Among the most experienced and highest-skilled workers, small gains in speed were accompanied by small declines in quality. These are observations from a particular support operation. A workshop could borrow the discipline of measuring resolved work while making its own estimate of the effect. Generative AI at Work.
For the illustrative pilot, the owner might compare assisted and ordinary intake over a period that includes typical work and awkward exceptions. Time from initial enquiry to a usable record is one measure. The number of follow-up exchanges, the accuracy of the record and the frequency of missed requirements give it meaning. A job accepted quickly with the wrong deadline is a poor result. So is a workflow that saves the person reading enquiries a few minutes and adds a longer check for the person preparing quotes. The unit of comparison needs to include both people and the final outcome.
The economics then depend on what the firm can do with any time it gains. Short, scattered savings may make the day easier without changing how many orders can be completed. That can be valuable, but it has a different financial meaning from gaining enough contiguous time to take on work or reduce overtime. Subscription charges are easy to see. Time spent reviewing output, updating references and investigating a bad record is easier to lose track of. Recording these costs alongside service quality can help an owner distinguish a better working day from a measurable improvement in business capacity.
Earn the case for wider use
Evidence from a pilot should help a firm choose where to extend assistance. Suppose the workshop finds that standard alterations are easy to capture, while bespoke orders regularly require a discussion with the customer. It might keep the assistant for common enquiries and direct bespoke work straight to an experienced colleague. The same distinction can guide training. Staff need to know how to spot ambiguous measurements, read an uncertain suggestion and flag a new kind of mistake. They also need a way to record that mistake so the next reviewer can learn from it.
Scale introduces a different set of decisions. A successful intake draft can tempt a firm to let the tool write directly to a calendar, issue a quote or change a supplier order. Each move grants new authority and changes the consequences of an error. The firm can assess those moves separately, asking who can approve them, what evidence is available at that point and how an erroneous action would be reversed. A useful drafting tool may have an excellent business case with review in place. Its results say little about the safety or economics of delegating a commitment that the business must honour.
The OECD figures describe a gap in the uptake of a broad class of technologies. For the owner of a small firm, the decision comes down to a more specific set of questions. Can the business identify a recurring difficulty, supply dependable context and check whether the proposed change makes the completed job better? A modest implementation that answers those questions gives the firm something it can build on. It also gives an owner grounds to stop when a simpler change serves the customer just as well, before recurring software and maintenance costs become part of the way the business operates.
Image credit: Tailor sewing machine by Sajid213, CC BY-SA 3.0. Illustrative photograph. Resized for display.