A member of your support team can draft a reply with AI in seconds. To decide whether that belongs in your support workflow, you need to know how often the draft is correct, how long it takes to review, and who handles questions it cannot answer. Recent adoption surveys put that distinction in perspective.
What the October surveys found
The UK government’s October 29 report found that 32% of UK businesses with at least five employees reported using one or more AI technologies in its 2026 survey, up from 16% in the 2025 wave. The 2026 survey interviewed 3,503 businesses. Read the UK AI adoption survey.
Goldman Sachs reported that 86% of respondents to its 2026 survey of 10,000 Small Businesses program alumni used AI and 15% had fully integrated it into core operations. Those figures describe program alumni who responded to the survey. The UK and Goldman Sachs surveys cover different groups, so each figure is best read in the context of its own sample. See the Goldman Sachs survey findings.
Pick one task with a clear review step
Take a support inbox that gets the same questions about delivery and returns every week. A first pilot could draft replies from approved help articles and show the source beneath each draft. A support agent checks the policy, edits the reply, and sends it. Tickets involving exceptions, account changes, or missing information stay with the team.
Give the pilot owner a narrow set of approved articles and decide which ticket data the tool can use. Start with redacted examples. Include a standard delivery question, a request that depends on the customer’s region, and a case where the policy provides no answer. Check that the tool sends uncertain cases to a person and keeps ticket text from authorizing actions such as issuing refunds.
Measure review time and mistakes
Before the pilot, time how long agents spend finding the policy, drafting a reply, and checking it. During the pilot, time the same work on comparable tickets, including reading sources, editing drafts, and writing replacements when needed. Track wrong policy claims and reopened tickets as well as minutes saved.
Decide whether to expand the pilot
Set a review date and agree on acceptable handling time and answer quality before starting. If drafts save time and agents can consistently verify the answers, add one nearby ticket category and repeat the checks. If corrections take up the time you saved, look at the cases that failed. Stale policies, missing source material, or too broad a ticket category each point to a different fix.