Consider a hypothetical 20-person business that decides to invest in AI licences for everyone. Staff begin drafting correspondence more quickly, producing summaries and getting help with routine research. The early feedback is encouraging. Yet a few months later, quotations still take too long, customer information is difficult to find and managers continue to chase the same approvals.
The AI tools may be working perfectly well. The problem is that faster individual tasks do not necessarily add up to a more productive business. I think this distinction deserves more attention before organisations commit to widespread deployment.
The business problem comes before the product
An Office for National Statistics study of UK firms identified difficulty finding suitable activities or use cases as the most commonly reported barrier to AI adoption in 2023 (39%). Cost followed at 21%, with AI expertise and skills at 16%. [1] More recent research from The Productivity Institute, using a panel of UK firms, found that existing management capabilities, particularly performance measurement, were associated with subsequent AI adoption. This does not establish that those practices cause a successful deployment, but reinforces the value of understanding how work is managed. [5]
The figures suggest a rather practical starting point: identify work that genuinely needs improving. A licence, however capable, cannot establish whether slow delivery is caused by poor information, an unnecessary approval, a lack of capacity or an unclear handover between teams.
Nor is AI necessarily the answer. Sometimes the best improvement is a better form, a clear decision rule, a single customer record or an existing software feature that nobody has bothered to configure properly.
Follow the work from beginning to end
Take the quotation process in that hypothetical business. An enquiry arrives by email. Someone collects the specification, looks up the customer history and prepares a price using a spreadsheet. A colleague checks availability, a manager approves the margin and the final quotation is sent to the customer.
AI might prepare a first draft of the response in seconds. That could be helpful. But if the customer records are incomplete, the price list is inconsistent or approval is sitting in an unattended inbox, the overall process may not become much quicker. In some cases, a faster first step simply creates a larger queue at the next one.
I would map the process before selecting a tool: where does work wait, where is information re-entered, which decisions need an experienced person, and which exceptions cause rework? The purpose is not to produce an elaborate process diagram. It is to find out what actually limits performance.
Information and accountability still matter
The UK Business Data Survey 2026 found that 21% of businesses already using AI had integrated it into existing business systems. Just 5% reported a formal written AI policy. The survey covered businesses handling digitised data, so these percentages should not be taken as estimates for every UK business. [2]
Integration alone is not a measure of success, and a small company may quite reasonably begin with a standalone assistant. However, unreliable source data will still produce unreliable outputs. It is worth agreeing which records are authoritative, what information an AI tool may access, and who is accountable when the result is wrong.
The appropriate controls depend on the task. Drafting a non-sensitive internal summary is not the same as recommending a customer price or making a decision affecting an employee. The National Cyber Security Centre advises that security needs to be considered across the design, development, deployment and operation of AI systems. That is particularly relevant as businesses move from individual assistants towards integrated workflows and agents. [3]
Five questions before the next AI licence
For a smaller company, I would start with a short conversation involving the people who perform the work and the person accountable for its outcome. Five questions are usually enough to establish whether an idea deserves further investigation:
- What is the desired business outcome? Faster responses, fewer errors, more completed orders or lower cost need different measures.
- Where does the current process actually slow down? Check handovers, approvals and rework, not just time spent writing or searching.
- Is the information accurate and accessible? An attractive demonstration using tidy sample data may conceal a much messier operational reality.
- Who owns the decision and checks the result? Define the human review, failure response and limits of automated action before deployment.
- How will the benefit be measured? Record the starting position, the full cost of change and the result after the initial enthusiasm has settled.
If there is no clear business outcome, no owner and no reliable way to measure improvement, I would hesitate before committing to more technology.
Start small, but properly
A sensible pilot need not be expensive or complicated. Select a repetitive, relatively low-risk process with an identifiable owner. Establish a baseline, test with real users and approved information, and deliberately include the awkward cases that rarely appear in a product demonstration. Record not only time saved, but the checking, correction and support effort introduced by the change.
Government-commissioned AI adoption research also found that many businesses interviewed had no formal metrics for attributing productivity or revenue changes to AI. That is a useful reason to agree the measures before a pilot rather than after an impressive demonstration. [4]
The result of a pilot may be to scale the AI solution, improve the underlying process first, choose conventional automation or decide that the business case is not strong enough. Any of those can be a sound decision if supported by evidence.
The difficult part is the change, not the purchase
Buying an AI licence is straightforward. Changing how work is organised, who makes decisions, what information can be trusted and how performance is measured takes more thought. For founders and leadership teams, that work may be less exciting than an impressive demonstration, but it is where much of the lasting value is likely to be found.
The next article will take a hypothetical 20-person company and work through what a proportionate first 90 days of AI adoption could look like.
Sources and methodological notes
[1] ONS, Management practices and the adoption of technology and artificial intelligence in UK firms: 2023 (24 March 2025). Source — Management and Expectations Survey; barriers are based on 2023 responses; management findings are associational.
[2] DSIT, UK Business Data Survey 2026 (18 June 2026). Source — Survey conducted October 2025–January 2026; results relate to businesses handling digitised data.
[3] NCSC, Guidelines for secure AI system development (27 November 2023). Source — Guidance primarily intended for AI system providers, relevant to adoption and integration decisions.
[4] DSIT, AI Adoption Research (28 January 2026; updated 13 February). Source — Business interviews February–May 2025; perceived benefits and revenue outcomes are self-reported.
[5] The Productivity Institute, More than just Plug and Play: Early Evidence on Organisational Capital and AI Adoption (28 September 2026). UK management-panel working paper; observational association, not proof of causality. Source
