The AI Investment Boom: Are Businesses Buying Value or Buying Into the Hype?

Debt and adoption are different stories

The Bank of England’s July 2026 Financial Stability Report discussed the rapid expansion of investment in AI infrastructure and the possibility that expectations about future returns could be tested. It also described the immediate system-wide risk as moderated by the relatively modest stock of AI-related debt at the start of 2026. The report says AI firms’ use of credit markets was accelerating, while noting little evidence so far that this had crowded out other borrowers. That is a warning about financing and market exposure, not a prediction that every AI project will fail.

Infrastructure investment means financing data centres, chips, electricity and the networks that support large-scale computing. It involves long-lived assets, large capital commitments and assumptions about future demand. A small business adopting an existing AI service is making a different decision: usually a subscription, some implementation effort and a choice about whether a defined process improves.

Supplier exposure still matters: infrastructure capacity, service terms or prices can affect a customer over time. The Bank’s discussion is context, not a substitute for a company-specific business case.

Start with a process, not an AI budget

A weak proposal begins with a tool and searches for work to justify it. A stronger one starts with a recurring problem: for example, staff spend hours finding information across approved documents, preparing routine first drafts or classifying incoming requests. Describe the current workflow before changing it. Who does each step? How many cases arrive? Where do delays, errors and rework occur?

Then ask whether AI is needed at all. A clearer form, a rules-based automation or a change to an approval queue may solve the problem with less cost and uncertainty. If AI remains a reasonable option, define the desired improvement in terms a manager can observe: shorter turnaround, fewer corrections, more cases handled without reducing quality, or more time available for work that needs judgement.

This is distinct from the broader questions in The real cost of business AI and AI agents change the economics of automation. Here the focus is whether an SME is buying measurable operational value while the wider AI infrastructure market makes large financing bets.

Put the whole cost into the calculation

A licence price is only one line in the investment. Add configuration, data preparation, connecting systems, staff training, checking outputs, support, security review and the time spent handling exceptions. Consider whether the work saved becomes a cash saving, additional capacity or simply a time benefit. Those outcomes are useful, but they are not interchangeable.

Consider a hypothetical small firm where a team spends 12 hours each week preparing routine client summaries. At an assumed loaded labour cost of £30 an hour and 48 working weeks, the current effort represents £17,280 a year. If a pilot safely removes 30 per cent of that work, the theoretical annual capacity released is worth £5,184. This is an estimate, not a promised saving: the team must verify quality and establish what happens to the time released.

Suppose the first-year costs are £1,800 for subscriptions, £1,200 for setup, £1,500 for staff training and £1,500 of human review and exception handling. Total cost is £6,000, so the first-year net benefit is negative £816 against the assumed time value. If setup and training do not recur, later annual costs might be £3,300 and the same verified capacity benefit could yield £1,884 a year. If the hours cannot be redeployed or output quality falls, the financial case is weaker. The figures are deliberately hypothetical; a real business should replace every assumption with measured data.

Measure outcomes, dependencies and the downside

Record a baseline before the pilot. Measure completed work, elapsed time, correction rates, customer impact and staff review effort. Compare like with like, including busy and quiet periods where relevant. A dashboard showing prompts, log-ins or generated text does not establish commercial value.

Also write down the dependencies: what data the system receives, which supplier operates it, whether a person must check each output, how the service can change, and whether the workflow can continue if it is unavailable. Include recurring fees and likely integration work in a longer-term view. This complements, rather than replaces, the infrastructure questions in The hidden infrastructure risk behind business AI.

Decide in advance what would cause the pilot to stop. If error rates rise, staff spend longer checking than they save, the provider changes its terms or the process owner cannot explain an output, pause and review. Reversibility has commercial value: a limited test gives the organisation evidence without locking it into a large commitment.

A sound investment is an evidence-led one

A business does not need to predict the future value of the entire AI market. It needs a proportionate decision about its own objectives, current process, total cost, risks and alternatives. Start small where the task is suitable, keep a person accountable for quality and measure what changes after adoption.

AI can be commercially valuable when it removes low-value effort, improves access to information or helps a team serve work more consistently. The point is not to avoid investment. It is to avoid confusing industry excitement, vendor claims or borrowed market forecasts with proof that a particular workflow works.

The best AI investment is not necessarily the most advanced technology. It is the one that produces a measurable improvement at an acceptable cost, with a clear owner and a sensible plan if the evidence changes.

Sources and further reading

• Bank of England, Financial Stability Report, July 2026 (PDF)