In the previous article, I argued that buying AI licences is unlikely to resolve an inefficient business process on its own. That raises a reasonable question: what might a sensible approach actually look like for a smaller company?
Consider a hypothetical 20-person business that services and maintains equipment for commercial customers. There are field engineers, a small office team, an operations manager and a founder who is still closely involved in day-to-day decisions. The company is busy, but quotations take longer than they should, service reports arrive late and the same customer details are entered into several systems. This is an illustration, not a client case study or a claim about achieved results.
First, decide what needs to improve
The OECD’s December 2025 discussion paper on AI adoption by small and medium-sized businesses argues for different adoption approaches according to digital maturity, the complexity of the intended use and its scope. That seems a sensible starting point for a smaller company: understand the work and its constraints before selecting a product. The OECD analysis covers G7 economies, not a measured outcome for this hypothetical UK business. [1]
I would ask the founder and the people doing the work to identify a handful of recurring problems. Which activities consume disproportionate time? Where do customers experience delays? Which errors create rework? And where is the business dependent on information held by one person?
Suppose the initial discussion identifies five possible opportunities:
- Service reports — use AI to turn approved engineer notes into a draft report, with an engineer checking the result before issue. A plausible first pilot.
- Quotations — prepare drafts from completed specifications, but retain authorised pricing and margin approval. Worth exploring next.
- Customer enquiries — categorise incoming requests for staff to review and route. Potentially useful, with data-handling checks.
- Invoice matching — investigate repeated discrepancies first; inaccurate records may be the main problem, not an absence of AI.
- Autonomous scheduling — defer decisions that could commit engineers, affect customers or create safety and service risks without adequate controls.
This is not a scientific ranking. It is an initial judgement about likely value, readiness and risk. The purpose is to choose one manageable starting point, while recognising that some opportunities may be better served by ordinary process improvements or existing software.
Days 1–30: understand and measure the work
I would choose service-report preparation for the initial assessment. The operations manager becomes the business owner; an experienced engineer is responsible for technical approval. Before introducing a tool, the company records how many reports are produced, how long preparation and checking take, how often corrections are required and how quickly customers receive the final document.
The team also traces the process from the engineer’s notes to the customer-facing report. Perhaps a missing serial number causes repeated calls, or the template asks for information already recorded elsewhere. Those issues should be corrected regardless of whether AI is introduced.
By day 30, the company should have a baseline, a named owner, a documented workflow and agreement on what a successful pilot would need to demonstrate. If those basics are missing, I would not rush into implementation.
Days 31–60: design a controlled pilot
The next stage is to test an approved tool against a limited selection of representative reports. I would begin with appropriately anonymised or otherwise authorised information, and check the supplier’s data-handling arrangements before using confidential customer records. The Information Commissioner’s Office provides guidance on assessing AI-related data protection risks, although its existing guidance is under review following legislative changes. [2]
AI produces a draft; the engineer remains responsible for confirming technical facts and approving the final report. The test set should include incomplete notes, unusual jobs and conflicting information rather than only straightforward examples. Every correction is recorded, together with the time spent checking the output.
For a small business, governance need not mean a large policy manual. It does mean deciding who may use the tool, what information may be entered, who checks outputs and what happens when the system gives an unreliable answer. The NCSC’s secure AI guidance highlights the importance of secure deployment, monitoring and operational responsibility, particularly where AI is integrated into business systems. [3]
Days 61–90: measure the result and decide
Assume, purely for illustration, that the company produces 40 service reports a week. The baseline is 18 minutes per report, including checking. If a pilot demonstrates a reduction to 12 minutes, again including checking and corrections, the capacity released would be four hours a week — about 184 hours over 46 working weeks.
Four hours released each week is not the same as four hours of cash savings. The benefit depends on quality, licence and support costs, customer outcomes and whether the business can use that capacity for worthwhile work.
I would compare the pilot with the original baseline: total staff effort, factual accuracy, rework, turnaround time, customer complaints, additional costs and security or privacy issues. A quick draft that needs extensive checking is not necessarily an improvement. Nor is a faster report useful if its technical accuracy deteriorates.
At the end of 90 days, the management decision should be explicit: extend the pilot, adopt it with agreed controls, fix the process and try again, or stop. Expansion to quotations or customer enquiries should depend on what the first exercise actually established, not on a general enthusiasm for AI.
A small start can still be strategic
I do not think a 20-person business needs an elaborate AI transformation programme to make useful progress. The starting point can be one process, one accountable owner and a measured pilot. But there should be a clear link to the business objective and a plan for managing the change once the initial trial has ended.
The most useful outcome of the first 90 days may be a working AI-assisted process. Equally, it may be a clearer understanding of why an existing process is inefficient, and a decision not to use AI yet. Either would be more valuable than an expensive roll-out with no convincing evidence of improvement.
In the next article, I will look at one of the more difficult questions: when AI releases hundreds of hours of staff time, how much of that represents genuine financial value?
Sources and methodological notes
[1] OECD, AI adoption by small and medium-sized enterprises: OECD discussion paper for the G7 (9 December 2025). Comparative adoption pathways and enabling conditions; not UK intervention results. Source
[2] ICO, Guidance on AI and data protection (accessed 22 October 2026). Source — Current ICO guidance notes that it is under review following the Data (Use and Access) Act.
[3] NCSC, Guidelines for secure AI system development (27 November 2023). Source — Guidance primarily for AI providers; relevant principles include secure deployment and operation.
