Small-Business AI Adoption: Reading JPMorganChase Institute Research

How many small businesses use AI? The answer changes depending on whether a researcher counts a free experiment, a paid subscription, an integrated workflow or a measurable business benefit. JPMorganChase Institute's April 2026 research offers one particularly useful perspective because it examines actual payments to AI services by a defined sample of US business-bank customers. That is a narrower but clearer question than asking owners whether they have ever tried an AI chatbot. It helps show how paid adoption has changed, while leaving important questions about effectiveness unanswered.

The report, Understanding the use of AI among small businesses, was published on 14 April 2026 and examines de-identified Chase Business Banking transactions from 2019 through 2025. The findings concern the studied American banking population; they must not be presented as UK-wide small-business adoption rates. A payment is a real economic signal, but it cannot tell us on its own whether an employee used the service daily, whether an AI answer was correct or whether the business earned more profit. Those distinctions matter when applying the research to purchasing decisions.

What the transaction-based method can and cannot see

The researchers identify payments to AI-related services and use them to describe adoption behaviour. They distinguish cumulative adoption from use indicated by payments in a given period. A business that paid once might appear in a cumulative adoption measure even if it later cancelled the service. Someone using a free AI assistant might not appear in the transactions at all. The definition is valuable precisely because it describes an observable event rather than an informal impression, but it remains one lens on the market.

It is not a census of every American company or a survey of each employee's software practices. The sample comes from Chase Business Banking customers who meet the report's analytic criteria. Firms with different banking arrangements may differ. The report does not permit a simple inference that one country is more advanced than another without a comparable method. When repeating any of its figures, retain the sampling context, period, definition of AI-related spending and original publication link.

Understand why newer business cohorts appear to adopt faster

One of the report's striking findings concerns the time from firm formation to paid adoption. The 2025 business cohort reached a 10% cumulative AI adoption threshold in about six months, whereas the 2019 cohort required roughly 77 months. These figures are from the report's measured sample and should be interpreted as the timing of identified payments. They do not prove that newly formed businesses use AI more effectively or that all established companies are slow to innovate.

Businesses founded later had access to a different software market from firms operating in 2019. New products and pricing models became available, and AI functionality became more prominent in everyday digital services. There may also be differences in the types of firms formed during each period. A responsible reading describes the faster observed uptake without asserting that any one factor caused it. For a UK business owner, the practical question is whether suitable tools now exist for their particular workflow, not whether they match another firm's adoption age.

Distinguish employer and non-employer firms

The Institute reported cumulative AI adoption by December 2025 of around 26.1% among employer firms and 15.3% among non-employer firms within the study's sample. The difference is interesting but not proof that hiring employees causes an enterprise to buy AI. Employer firms may have different revenues, tasks, administrative demands and opportunities for internal specialisation. The research offers associations within the data; interpreting causes requires more evidence.

For a sole trader, an affordable assistant that reduces repetitive document work may be suitable despite an apparently lower adoption rate among comparable non-employers. For a ten-person business, a shared knowledge and response workflow may be more important than personal productivity tools. The report does not decide between those approaches. A small firm should assess its existing software, who will supervise outputs and what business process is actually constrained.

Separate occasional purchase from sustained, useful use

A subscription proves neither regular use nor successful integration. A business may trial an AI product, find it difficult to connect to approved information and abandon it after a month. Another may use a small, inexpensive feature reliably for a well-defined task. The transaction study does not directly measure those operational outcomes. A responsible buyer should therefore look beyond the market adoption graph to the staff work that would change after purchase.

Define the task in ordinary language. Examples might include locating an approved policy paragraph, triaging incoming enquiries for human review or drafting an internal action list after a meeting. Test whether the tool provides correct source links and handles missing information appropriately. Record the time employees spend verifying output and correcting errors. A nominal saving in drafting effort can disappear when a confident but inaccurate answer takes longer to investigate than writing the message manually.

Interpret AI spending without overlooking implementation cost

The Institute considers how spending changes by business cohort, but the price of a subscription is only one component of the cost of adoption. Staff training, administrator attention, security settings, workflow redesign, evaluation and exit arrangements can be material for a small company. A cheap tool can be costly if employees must paste the same information between disconnected applications; a more expensive integrated product is not necessarily better if the underlying process is weak.

Prepare a simple total-cost model: subscription, relevant usage fees, configuration, staff hours, verification effort and support. Keep the assumptions visible. Do not claim a return on investment before the firm has measured an outcome or established a credible forecast with uncertainty. If a provider advertises a percentage productivity improvement, ask which tasks were measured and whether the test resembles the firm's own operation. Vendor claims and transaction studies answer different questions and should not be merged into a single benefit statement.

Use an operational pilot rather than copying a national statistic

Choose a bounded workflow with clear ownership and a reasonable sample of representative tasks. Establish how long the work takes today, how often corrections are required and which customer or compliance risks matter. Test the proposed AI tool using non-sensitive examples initially. Include ambiguous and difficult cases, not just the easiest prompts. Assess whether results are supported by authoritative information and whether staff can override an unsuitable recommendation.

For customer service, measure genuinely resolved requests, escalations and repeat contact, not merely messages generated. For administrative writing, measure the end-to-end time including human review. For document retrieval, measure how often the employee can identify the correct current source. The pilot should have a practical exit if the product cannot satisfy basic requirements. A paid adoption statistic can explain why the business might investigate AI; the pilot determines whether this particular use is worthwhile.

Protect customer information and preserve human responsibility

Introducing an AI service may change where business information is processed. Map which data enters the tool, whether the vendor retains it, which staff can access it and how the organisation can correct or remove information where appropriate. The UK Information Commissioner's Office explains principles including data minimisation and accuracy. Apply the relevant legal and contractual requirements to the real processing arrangement, especially when personal or confidential data are involved.

Keep consequential decisions with authorised employees. AI may summarise a support history, but should not invent a promise about a refund, delivery or contract change. Where its source is missing or contradictory, the tool should direct the employee to review the original record. Establish how incidents, false classifications and unwanted disclosures are investigated. These safeguards are part of the implementation cost and success criteria, not optional additions after the subscription is purchased.

Read the research as context, not as a universal recommendation

The report provides evidence that paid AI adoption rose and that observed patterns varied across the studied businesses and cohorts. It does not establish a direct productivity or revenue gain from those purchases, nor does it assess every type of AI use. An owner should therefore distinguish three questions: are other businesses paying for AI, is a tool technically suitable for this workflow, and did it produce a worthwhile result in our own operation? Only the first question is directly answered by the transaction study.

Keep the original paper beside any presentation citing its data. State that the research is US-based, payment-based and covers 2019–2025 activity. If another report gives a different adoption rate, examine the methods before deciding that the two disagree. A UK self-report survey and a US transactions dataset may both be correct within their own definitions. A careful comparison is more useful than a claim that adoption has reached one universal percentage everywhere.

Paid adoption is a meaningful signal that businesses are allocating money to AI services. The commercial question is what happens after the purchase. Begin with a limited, supervised task, preserve original evidence, calculate the real operating cost and review difficult cases. That is a more dependable way to use the research than treating a percentage as an instruction to buy software.