Access to AI-enabled capability has become inexpensive. Federal data suggests the harder constraint is now organizational: defining work, maintaining information, and integrating tools into real workflows.
Prefer the short version? Read the accessible article →For the first time, a ten-person company can rent capabilities that a few years ago required a research department, a development team, or an agency retainer: drafting, summarization, translation, code generation, document analysis, and first-pass research. The tools cost a monthly subscription and require no infrastructure to begin.
Yet access and advantage are not the same thing. If inexpensive access were sufficient, adoption should look roughly the same across firm sizes, and the firms that adopt should show similar depth of use. The federal data does not look like that. This paper asks why, and what it means for an operator deciding where to spend the next dollar and the next hour.
18%
of U.S. employer firms used AI in at least one business function, Nov 2025–Jan 2026 (32% employment-weighted).
Census CES WP 26-25
57%
of firms using AI in business functions use it in three or fewer of fifteen measured functions.
Census CES WP 26-25
64%
of AI-using firms report no organizational adjustments — training, workflow redesign, or data changes — to support it.
Census CES WP 26-25
Much of the public conversation collapses five different states into one word. We keep them separate throughout.
Figure 1
Access, adoption, usage, integration, and value
| State | What it means | What the federal data measures |
|---|---|---|
| Access | The business could use the capability if it chose to. | Not measured directly. Inferred from pricing and availability. |
| Adoption | The business reports using AI at all. | BTOS core question: AI used in any business function in the last two weeks. |
| Usage | Workers use AI for specific tasks. | AI supplement: worker task use, e.g., writing, search, document analysis. |
| Integration | AI is embedded in multiple functions, with workflows, training, and data adjusted to support it. | AI supplement: breadth of functions, organizational adjustments, operational investment. |
| Value | Measurable improvement in output, cost, quality, or revenue attributable to AI. | Only associations with self-reported performance. No causal estimate at population level. |
The distinction matters because each state has a different bottleneck. Access is limited by price. Adoption is limited by awareness and perceived relevance. Usage is limited by individual skill. Integration is limited by process, information, and ownership. Value is limited by whether anyone measures the result.
The change in access is real. Stanford's 2025 AI Index reports that the inference cost of querying a model performing at the level of GPT-3.5 on a standard benchmark fell more than 280-fold between November 2022 and October 2024. The capability that once required a dedicated vendor contract is now bundled into office software, accounting platforms, and customer-service tools most small businesses already pay for.
Controlled experiments also show that, on well-specified tasks, the capability is genuine. In a randomized experiment with 453 college-educated professionals performing writing tasks, access to ChatGPT reduced time taken by 40% and raised evaluated output quality by 18% (Noy and Zhang, Science, 2023). In a staggered deployment across 5,179 customer-support agents, access to a generative-AI assistant increased issues resolved per hour by 14% on average and by 34% for novice and lower-skilled workers (Brynjolfsson, Li, and Raymond, 2025).
The same body of research also shows the limit. In a field experiment with 758 consultants, AI access raised productivity and quality on tasks inside the model's capability frontier, but consultants using AI were 19 percentage points less likely to produce correct answers on a task designed to fall outside it (Dell'Acqua et al., 2023). The tool helped most when the task was defined and the human could tell whether the output was right.
The Census Bureau's Business Trends and Outlook Survey (BTOS) samples roughly 1.2 million employer businesses, excluding farms, in biweekly panels. It is the broadest nationally representative measure of business AI use in the United States. During the 2025–2026 AI supplement reference period (November 2025 to January 2026), 18% of firms reported using AI in at least one business function, rising to 32% on an employment-weighted basis. Firms expected adoption to reach 22% within six months.
The gap between 18% and 32% is the first important finding. A firm-weighted figure counts each business once; an employment-weighted figure counts each worker. When the employment-weighted rate is much higher, AI use is concentrated in larger employers. About three quarters of U.S. firms have fewer than ten employees, which is why firm-weighted national rates look low next to surveys that sample mostly large companies.
More recent core BTOS data, collected December 14, 2025 through May 3, 2026, shows national AI use holding between 17% and 20%, with 19.8% in the collection period ending May 3, 2026.
Figure 1
Current AI use by firm size, AI supplement period (firm-weighted)
Fewer than 20 employees
18%
100–249 employees
25%
250+ employees
31%
Share of firms reporting AI use in at least one business function. Figures are approximate values reported in the working paper text for the revised question.
Source: U.S. Census Bureau, CES WP 26-25 (Nov 2025–Jan 2026)
Before the revision, the size pattern was modest: firms with one to four employees drifted from roughly 4% to 10% current use, while firms with 250 or more employees rose from roughly 5% to 13%. After the revision, the gradient is clearer. The more consequential pattern is forward-looking.
Figure 2
Current versus expected AI use by firm size (revised question, firm-weighted)
| Firm size | Current use | Expected in six months | Expected change |
|---|---|---|---|
| 1–4 employees | ~18% | ~21% | +3 points |
| 100–249 employees | 25% | 34% | +9 points |
| 250+ employees | 31% | 40% | +9 points |
The authors describe this as an expectation gap that scales with firm size. Small firms are not far behind today on a firm-weighted basis; they are planning to move more slowly. In the core survey through May 3, 2026, AI use increased among firms with at least 20 employees but did not change significantly among firms with fewer than 20. By that period, 37% of firms with 250 or more employees and 32% of firms with 100 to 249 reported use.
Figure 2
Current AI use by sector, collection period ending May 3, 2026
Information
39.7%
Finance and Insurance
33.9%
National rate, all sectors
19.8%
Retail Trade (approx.)
14%
Source: U.S. Census Bureau, America Counts, May 26, 2026 (BTOS)
Adoption concentrates in knowledge-intensive work. Among very large firms in Information, Professional Services, and Finance, use reaches 50% to 60% (60% to 70% employment-weighted). Sector, as much as size, predicts adoption, and the two overlap: industries where information is already the product adopt first.
Figure 3
AI use by business function, among firms using AI in at least one function
Sales and marketing
52%
Strategy and business development
45%
Information technology
41%
Research and development
40%
Top four of fifteen functions measured. Firms may report more than one function.
Source: U.S. Census Bureau, CES WP 26-25 (firm-weighted)
At the worker level, 23% of firms (41% employment-weighted) report employees using AI for work tasks. Among those firms, writing and editing is the most common use, reported by 85%, followed by information search and document analysis. Roughly 65% limit worker use to three or fewer task types.
Use by workers and use by the firm are not the same thing. In 36% of firms where workers use AI, there is no formal firm-level adoption. In 19% of firms with formal adoption, there is no evidence of worker-task use. AI is arriving from both directions, frequently without the two meeting.
Figure 3
Organizational adjustments among AI-using firms
| Adjustment | Share of AI-using firms |
|---|---|
| No institutional adjustment reported | 64% |
| Trained staff | ~15% |
| Developed new workflows | ~15% |
| Changed data management or storage, or made complementary capital investments | ~7–8% |
| Hired staff trained in AI | Least common adjustment |
Using the paper's latent-class analysis, the largest group of AI users — 37% — are "minimalist adopters" with low use across all functions. "Comprehensive adopters" who use AI broadly are 4%. On the paper's composite integration index, 68% of all firms score zero; those firms average 14 employees. Firms in the highest range average 87 employees, use AI in 13 of 15 functions, and report about five types of organizational change.
Among firms reporting AI-driven task effects, 66% report augmentation only. About 5% of AI-using firms report any AI-driven headcount change, split roughly evenly: 2.3% increases, 2.0% decreases. Sixteen percent replaced existing software or equipment with AI-integrated tools.
The authors find that breadth of AI use across functions, breadth of worker-task use, and operational investment are each positively and significantly associated with above-average reported performance and current sales increases. Functional breadth shows the strongest association. The authors state explicitly that these regression summaries carry no causal interpretation.
This is where parts of the business press get ahead of the evidence. Headlines built on convenience samples of large firms report majority adoption; the national employer population does not. Claims of broad labor replacement are not supported by the 2% of AI-using firms reporting AI-driven headcount decreases. And claims that AI is already raising small-business productivity are, at population level, not yet measured.
If price were still the binding constraint, the most common reason for non-adoption would be cost. It is not. Among firms not planning near-term AI use, 65% say AI is not applicable to their business. Lack of knowledge of AI's capabilities follows at 22%, and privacy or security concerns at 20%.
Figure 4
Reasons for not using AI, among firms not planning near-term use
AI is not applicable to this business
65%
Lack of knowledge of AI capabilities
22%
Concerns about privacy or security
20%
Firms could select multiple reasons. Top three shown.
Source: U.S. Census Bureau, CES WP 26-25 (firm-weighted)
"Not applicable" is a reasonable answer for some businesses. For many, we read it differently: the owner cannot yet see which specific piece of work AI would attach to. That is not a technology problem. It is a problem of work that has not been described precisely enough to hand to anything — a person, a contractor, or a model.
Combine that with the adoption data. Most adopters use AI in a few functions, mostly for writing and search, without training, workflow changes, or data changes. The firms that report broad integration are larger and report more organizational change. The experimental literature shows gains when tasks are defined and outputs can be evaluated, and errors when they cannot.
Taken together, the evidence is consistent with a shift in the bottleneck — from buying the capability to the organizational conditions needed to use it:
None of these are new. They are the same disciplines that separated well-run small companies from struggling ones before AI. What has changed is their price relative to the tool. When the software was expensive, it was the obvious constraint. When it is nearly free, the organizational work becomes the visible one.
It would be easy to read the federal data as a story of small firms falling behind. On a firm-weighted basis today, the current-use gap between the smallest and largest firms is roughly 13 percentage points, and the pre-revision gap was narrower still. The larger divergence is in expectations and in depth of integration.
Some structural features of small organizations may make integration easier, not harder. We present these as operating hypotheses. The BTOS does not measure them, and we are not aware of population-level evidence that establishes them.
Shorter decision chains.
Hypothesis. In a firm where the owner approves spending and runs the workflow, a trial can start the same week. The data shows top-down adoption without worker use is common; small firms may be less exposed to that disconnect.
Less legacy infrastructure.
Hypothesis. Fewer integrated systems means fewer dependencies to unwind. Offsetting this, 16% of AI-using firms replaced software to adopt AI, a cost that is proportionally heavier for small budgets.
Proximity between decision-maker and workflow.
Hypothesis. The person evaluating whether an output is correct is often the person who knows the customer. The experimental evidence suggests that evaluative judgment is what separates useful from harmful AI output.
Faster, cheaper experiments.
Hypothesis. A small firm can test a narrow use case on one process without a formal program. Whether this translates into measurably better outcomes has not been studied at scale.
Framework
Liquid framework — from access to value
For operators, the practical sequence is narrower. AI should be deployed against a specific operating problem, not adopted because competitors are talking about it.
Framework
Liquid framework — the AI-ready operating sequence
The sequence is deliberately unglamorous. Most of the steps happen before the tool is chosen. That is the point: the federal data suggests the scarce input is no longer the software.
Technology can democratize access to capability without democratizing the organizational ability to use it.
The thesis in the title largely survives, with qualifications. The evidence supports three statements. First, access to AI-enabled capability has become inexpensive, and task-level gains under defined conditions are real. Second, adoption across U.S. employer firms is roughly one in five, rises with size and knowledge intensity, and among adopters is mostly shallow. Third, the barriers firms report are about relevance and understanding, not price.
The evidence does not support stronger statements. It does not show that AI is closing the small-business gap, nor that it is widening it in outcomes. It shows associations between integration breadth and performance, not causation.
The opportunity, stated as strongly as the data permits, is this: certain capabilities are becoming cheap enough that organizational discipline, rather than organizational size alone, may become a more important determinant of what a small company can accomplish. For an operator making technology decisions today, that suggests spending less time choosing tools and more time defining the work they will be asked to do.
References
Cite this research
Rebolledo, Alenn. "AI Did Not Eliminate the Small-Business Gap. It Moved It.." Liquid Research, 2026. https://liquidiq.io/research/ai-did-not-eliminate-the-small-business-gap.
Rebolledo, A. (2026). AI Did Not Eliminate the Small-Business Gap. It Moved It.. Liquid Research. https://liquidiq.io/research/ai-did-not-eliminate-the-small-business-gap
@misc{rebolledo2026ai,
author = {Rebolledo, Alenn},
title = {AI Did Not Eliminate the Small-Business Gap. It Moved It.},
howpublished = {Liquid Research},
year = {2026},
url = {https://liquidiq.io/research/ai-did-not-eliminate-the-small-business-gap}
}Liquid Research is independently published by Liquid IQ. It is not peer-reviewed and has no DOI.
How we research: Liquid Research methodology →
Author
Founder
Alenn founded Liquid after working across startups, technology, analytics, operations, and marketing strategy. His research focuses on where growth is actually constrained and how operating systems change business outcomes.
Strategy · Systems · Growth · Decision Intelligence
Related services
Have a similar problem?
Related Research
Smaller companies rarely lack ambition or information. They lack the analytical infrastructure that larger firms treat as ordinary overhead. A look at the gap and how to close it cheaply.
The Small Business Series
Automating an undefined process encodes its variation permanently. A sequence for deciding what is ready to automate, what should be standardized first, and what should stay manual.
Human + Machine
Tool sprawl is usually a symptom of undefined ownership. A framework for deciding whether a problem needs a purchase, a connection, or a decision.
The Operating System Series