You run a firm of somewhere between five and fifty people. The vendor emails keep arriving, a competitor has started mentioning AI in its pitch, and at least one of your team is already pasting work into ChatGPT, sanctioned or not. The question underneath it all is blunt. Is this worth the money and the attention, or is it another software cycle that will cost a few thousand pounds and deliver a login nobody uses?
The evidence now supports a clearer answer than it did two years ago, and the answer depends less on your sector than on the shape of the work you would point AI at, and on how you would buy it.
What choice are you actually facing?
Strip away the vendor noise and you are choosing between three positions. Do nothing yet, run a narrow pilot of an off-the-shelf tool against one repetitive process, or commit to a bespoke AI build. For a firm of five to fifty people, the evidence now points firmly to the middle option, and almost never to the third.
You are also deciding alongside plenty of undecided peers. A 2024 YouGov survey of 1,000 decision-makers in UK businesses with up to 250 employees found 31 per cent already using AI tools and a further 15 per cent planning to. The UK government’s own adoption research puts overall business usage lower, at around one in six firms. Across OECD countries, the share of firms using AI grew from 5.6 per cent in 2020 to 14 per cent in 2024, still well behind cloud accounting or CRM. If you have not moved yet, you are not late.
That context takes the panic out of the decision. The productivity evidence says AI works well for certain kinds of work and poorly for others. So the useful question is whether your firm has enough of the work it is demonstrably good at, and whether you can buy it in a way that fails cheaply if it fails at all.
When is a focused AI pilot worth it?
A pilot earns its keep when the target process is repetitive, high-volume, text-heavy and measurable. Answering routine enquiries, triaging inbound email, processing invoices and drafting standard documents all qualify. Where those conditions hold, independent evidence puts time savings at around five per cent of work hours per user, with customer-facing throughput gains of ten per cent or more.
The numbers are consistent across studies. Analysis by the Federal Reserve Bank of St Louis, using representative survey data from late 2024, found workers who used generative AI saved an average of 5.4 per cent of their work hours, about 2.2 hours in a 40-hour week. During the hours AI actually assisted, they were roughly a third more productive. A separate NBER field study of 5,179 customer support agents found an AI assistant lifted issues resolved per hour by 14 per cent on average, and by 34 per cent for the newest staff, because the tool spread the habits of the best performers to everyone else. In the UK government’s adoption research, 56 per cent of firms using AI reported a rise in overall productivity.
The named small-firm cases point the same way. Suitor, a five-person Australian suit rental company, added a chatbot and watched response times drop from three minutes to six seconds, with 85 per cent of queries handled without a human. An invoice automation pipeline built for a property management firm covered its own cost within eight weeks. Bella Santé, a Boston medical spa, layered a chatbot on top of its call centre rather than replacing it, and attributed over $66,000 of revenue to it within six months. None of these were grand change programmes. Each was one tool aimed at one process.
When is holding off the better call?
Holding off makes sense when the foundations are missing. If your invoicing, scheduling and client records live in spreadsheets and inboxes, if your data is patchy, or if the work itself is bespoke and low-volume, AI tools have nothing reliable to grip. The money is better spent on basic digital plumbing first, and the pilot will still be there next year.
The readiness data backs that caution. In the UK government research, only about a third of firms planning to adopt AI felt ready to implement it, and limited skills sat alongside lack of identified need as the top barriers. The British Business Bank makes the same point from the other direction, warning that AI can be expensive, complicated and time-consuming to implement where a firm lacks in-house skills and has to buy them in.
There is also a category of work AI handles badly at any price. If your value sits in bespoke, judgement-heavy projects with little repetition, a generic tool will produce plausible but off-target output that costs more to fix than it saves. And if your client relationships trade on a personal voice, be careful about delegating customer-facing words to a machine. In the YouGov survey, 57 per cent of decision-makers worried AI could flatten their firm’s creativity. That worry is legitimate, and the fix is scope. Keep AI on the repetitive internal work and keep humans on the words clients actually read.
What does getting the call wrong cost?
The wrong call cuts both ways. Overcommit and you can burn five figures and months of attention on a project that never reaches production. Wait too long and you concede a growing productivity edge to competitors who moved first. Careless adoption adds a third cost, because UK data protection duties apply in full the moment AI touches personal data.
Start with overcommitting. S&P Global survey data, reported by CIO Dive, shows 42 per cent of companies scrapped the majority of their AI initiatives in 2025, up from 17 per cent a year earlier, and the average organisation abandoned 46 per cent of AI proofs-of-concept before production. Those are enterprise figures, and enterprises can absorb failed experiments. A twenty-person firm cannot. Cost analyses put even small bespoke automation projects at $10,000 to $50,000 all-in, while a 2022 UK government study found smaller firms that adopted AI spent an average of £9,500. That is real money for a tool that may never embed.
Waiting has a gentler cost curve, but it is not free. The same government research found 77 per cent of AI-using firms had seen no revenue change yet, so the competitive gap is still narrow. The productivity data says it compounds, though. Two hours a week per person, across a team of fifteen, is roughly a full working day of capacity every week that faster competitors are banking and you are not.
The third failure is carelessness. The Information Commissioner’s Office is explicit that UK GDPR applies in full when AI processes personal data. Feed client records into an unvetted tool and you have created a data protection incident that no amount of saved admin time will pay for.
What should you ask before you decide?
Six questions separate a disciplined purchase from an expensive experiment. They force you to name the process, price the problem, and check the tool can fail safely. If you cannot answer them in an afternoon with your operations lead, treat that as the first finding, because a firm that cannot measure the problem cannot measure the payback either.
Before signing anything, work through these:
- What exact process are we improving, and what does it cost us today in time, errors and delay?
- Is the data this tool will touch clean, lawful and secure enough to automate?
- Will it need integration with our CRM, accounts or ticketing systems, and who does that work?
- Who owns the workflow after go-live, including prompts, monitoring and upkeep?
- What is the payback period if the results are only half as good as the vendor promises?
- Could better process design, or software we already own, solve this more cheaply?
Run them against a single process, give the pilot twelve weeks and three metrics, and make the renewal decision on numbers rather than the demo. On that basis, for a service-led owner-managed business, the answer to the headline question is a qualified yes. Worth it when it is small, specific and measured. Rarely worth it as a leap of faith, and never worth it as a bespoke build bought to keep up appearances.
