Small business AI adoption: who is ahead, who is behind, and why

Small business AI use is rising but uneven. The US Census Bureau reported in May 2026 that 19.8% of businesses used AI, against 32% of firms with 100 to 249 employees and under 20% of firms with four or fewer. The OECD (2021) names awareness, skills, investment and technology lock-in as digital gaps for small firms.

EasWrk is a hosting platform, not a research institute. This essay summarises published sources, which are listed at the end. Where a sentence is the author's opinion, it says so.

The Small Business AI Adoption Gap

Software vendors now ship AI features to every customer, whatever the customer's size or skills. Census Bureau data shows that use does not spread evenly. The Bureau reports that 37% of firms with at least 250 employees used AI. In the period ending 3 May 2026, 32% of firms with 100 to 249 employees did, and less than 20% of firms with four or fewer employees did (US Census Bureau, 2026).

The gap is also moving. Between December 2025 and May 2026, AI use rose among firms with at least 20 employees and did not change significantly among firms with fewer than 20 (US Census Bureau, 2026).

Forecasts of the prize are large. The McKinsey Global Institute (2023) estimated that generative AI could add $2.6 trillion to $4.4 trillion a year to the global economy across 63 use cases. A forecast like that says nothing about who captures the value. In this essay's view, the firms best placed to capture it are those that already have staff time, clean data and money to try tools and drop the ones that fail. A five-person shop has less of all three.

The question this essay asks is not whether AI will reach small business. It already has. The question is who benefits, who falls behind, and what a small owner can do about it.

Small Business AI Adoption Statistics

Measuring AI adoption is hard because the term is loose. Is an AI recommendation engine switched on inside a Shopify store AI adoption? Does a sole trader who asks ChatGPT to draft a reply to a complaint count? The answers depend on how a survey words its question.

Small business AI adoption survey: why the numbers differ

The Census Bureau's figures come from its Business Trends and Outlook Survey, which asks whether a business used AI in the past two weeks. In November 2025 the Bureau changed the question. It now asks about use in any business function, not only in producing goods or services (US Census Bureau, 2026). Figures from before and after that change are not strictly comparable, and private surveys that word the question differently will give different shares again.

Read any adoption number with three questions in mind. Who was surveyed, and how large were they? What exactly did the question ask? When was it asked? A headline share that hides firm size can mislead, because the smallest firms trail the largest ones in the Census data.

Small business is a wide category. A five-person legal practice, a 12-employee restaurant and a 180-worker manufacturer are all small businesses, and their needs differ. Averages across all of them hide more than they show.

Goldfarb and Tucker (2019) review how digital technology lowers five costs: search, replication, transportation, tracking and verification. A cheaper capability helps a firm only if it can use it. In this essay's view, that is where small firms lose ground. The tool costs little, but using it well takes data, time and skill.

A small business that does not use AI marketing or support tools competes against rivals that do. In this essay's opinion, the cost of that gap grows over time, because a rival that learns faster keeps learning.

The Skills and Literacy Deficit

The OECD (2021) found that many small firms still lag in digital transformation, especially the self-employed and micro-firms. Around 60% of those firms cited adaptation costs as a barrier. The OECD also lists remaining gaps in awareness, skills, investment, technology lock-in and data culture. The report covers digitalization in general, not AI alone, but the same gaps apply.

In this essay's view, AI literacy is not one skill. It covers four: what the tools can and cannot do, how to judge the quality of their output, how to fit them into daily work, and what can go wrong.

Agrawal, Gans and Goldfarb (2018) describe AI as a prediction technology, a way to make better guesses from data. That framing shows what AI needs. Good predictions need good data, and good data needs consistent record keeping. An owner whose customer records sit across email, a spreadsheet and an unconnected till has a data problem before an AI problem. The OECD's phrase for it is weak data culture.

Brynjolfsson, Li and Raymond (2023) studied the staggered rollout of a generative AI assistant to 5,179 customer support agents. Productivity, measured as issues resolved per hour, rose about 14% overall. It rose about 34% for novice and low-skilled workers, with minimal effect on experienced workers. The result suggests AI can spread expertise. In this essay's view, it describes a firm with a large support team and time to adapt. A solo operator rarely has those conditions, so the study should not be read as a promise for the smallest firms.

Language models write fluent text whether or not it is accurate. An owner who cannot check the output can publish a wrong claim on a website or send a customer an invented policy. Reading what the tool wrote before it goes out is the cheapest safeguard.

Vendor Lock-In and the Platform Dependency Problem

AI features often reach small businesses inside tools they already pay for, such as an ecommerce platform, accounting software or an email service. That removes a barrier. The owner switches on a feature and does not need to understand the model behind it.

It also creates a dependency. When a business's AI features live entirely inside one vendor's product, the business cannot move them. It cannot compare providers or negotiate price, and it accepts whatever terms and prices the vendor sets. The OECD (2021) lists technology lock-in among the gaps that hold back small firms' digital progress.

Terms of service differ on data. Before you enable an AI feature, check whether the vendor may use your customer data to train its models, and whether you can export your data and settings if you leave. If the terms do not say, ask the vendor in writing.

In this essay's view, lock-in grows with use. The longer a business runs on one platform, the more history and customization it builds there, and the harder that is to move. A platform you can leave keeps your options open. That is one reason some owners prefer to run a WooCommerce store without a developer on infrastructure they control, instead of adding features one subscription at a time.

Customer Expectations and Competitive Pressure

Pressure to adopt AI comes from customers as well as vendors. In this essay's opinion, a shopper who gets instant automated replies from a large retailer may start to expect quick answers from everyone, including the owner who answers email between jobs. Personalised recommendations work the same way. A customer used to them may find a plain product grid thin, even when the products are better.

Bessen (2018) showed that automation does not always cut jobs. Whether it does depends on demand. In textiles, steel and automobiles, demand was highly elastic at first and later became inelastic, and employment tracked that shift. Bessen treats the net effect of AI on employment as unclear. The same logic applies to firms. If AI-driven efficiency lets a business serve more customers at a lower price, that business can take share from one that did not adopt it. This essay offers that as an inference, not a finding.

Pricing is one place where the pressure shows up. Retailers use algorithms to set and change prices, and a small shop that cannot match that precision may lose on margin. The sister essay on AI dynamic pricing and dark patterns covers how those systems work and where regulators have stepped in.

How to Adopt AI: Start With One Task

Rigorous research on how small firms adopt AI is thin. Most published studies, including the support-desk study above, look at larger organizations. The advice below is the author's opinion, drawn from that evidence and from the gaps the OECD describes. It is not a research finding.

• Pick one task. Choose a job that is repetitive, low risk and easy to check, such as a first draft of customer replies or product descriptions.
• Measure before and after. Time the task for a week without the tool and a week with it. Keep the tool only if the saving is real.
• Check every output. Read what the tool produces before a customer sees it. Fluent text can still be wrong.
• Read the data terms. Find out what the vendor does with your customer data and whether you can export it.
• Expand slowly. Add a second task only when the first one runs without your attention.

If you run a store on WordPress, the small business hosting page explains what a managed host handles, and AI inside your hosting account describes one way to use AI for site tasks without adding another vendor.

Policy Responses and Support Gaps

The OECD (2021) reports that governments have prioritized digitalizing and reskilling their economies in recovery packages, including support for e-commerce, teleworking and digital security in small firms. It also says policies must address investment gaps, technology lock-in, digital skills and data culture. Those are the same gaps this essay has described.

In this essay's opinion, support aimed at awareness alone will not close them. A webinar does not help an owner with no time to attend, and a software subsidy does not help one who has not learned to use the software. Practical help that fits a specific trade and workflow looks more useful, whether it comes from trade associations, chambers of commerce or small business development centers. Whether those bodies have the resources to provide it is an open question.

A policy that treats AI adoption as always good misses a risk. A small business that adopts a tool it does not understand can leak customer data, depend on one vendor or act on wrong output. For some owners, waiting is the lower-risk choice.

There is no single fix. Barriers vary by trade, firm size and owner. This essay's view is that the gap between what AI offers and what a small owner can use today is wide, and it will not close on its own. Until training, data tools and exit options reach the smallest firms, the claim that AI levels the field remains unproven.