Small-Business AI Adoption and Growth Expectations

Sources#

• AI Adoption and Employment Expectations: Evidence from a Survey of Small Business Owners

Summary#

The vault's firm-level AI-and-headcount evidence comes from large or tech-skewed populations: Ramp's card-spend customers (Firm AI-Spend Intensity and Headcount Growth), ChatGPT Enterprise buyers (The Enterprise AI Adoption Gradient), European ICT-survey firms with ten or more workers (AI and Market Power), and AI-citing public incumbents (Function-Selective AI-Era Headcount Cuts). Aarons & Sarkar's Liberty Street Economics post (Federal Reserve Bank of New York, 2026-10-08) adds US small employer firms, from the AI module of the Federal Reserve Banks' 2025 Small Business Credit Survey (SBCS), fielded September to November 2025 and weighted to the national population of employer firms.

The finding: firms using AI expect to grow more, in employment and in revenue, than firms that do not, and the gap survives controls. The authors go further and argue that it is not optimism-driven selection. That further claim is weaker than the measured result, for reasons set out below.

Evidence note. empirical, regulator-staff voice. A short research blog post by two New York Fed economists, not a working paper: no regression table, no list of controls beyond "firm and owner characteristics and location", and no standard errors in prose. Three limits apply to every number on this page. (1) The outcome is an expectation, a 12-month-forward self-report collected in late 2025, not a realized change, and the authors say outcomes are "a subject for future research". (2) Cross-section only, so every estimate is an association, with no instrument and no timing variation. (3) Expectations are about total employment and revenue, not change attributed to AI. A diffusion index of 32.7 does not mean AI users expect AI to create jobs. The regression estimates (Charts 2 and 3) were transcribed by reading points and confidence bars off the images and are approximate to about ±0.5 ppt; the prose rounds them and agrees. (Compiler background, not stated in the post: the SBCS is a convenience sample recruited through Reserve Bank partner organizations and then weighted, not a probability sample. Weighting corrects observed composition, not selection into responding.)

The raw gap#

Diffusion indexes, percent expecting an increase minus percent expecting a decrease over the next twelve months (Chart 1, bar labels viewed at compile):

46% of employer firms report that the business or its employees use AI tools, and another 15% plan to adopt within twelve months. Among users, 63% call AI somewhat or very important to production and 51% have integrated it partially or fully into business processes (SBCS 2026 report, cited in the post).

How much survives controls#

Net probability of an expected increase, adopters minus non-adopters, in percentage points (Charts 2–3, viewed; current-AI column, current-or-planned in parentheses):

Every 95% interval excludes zero; the tightest lower bound is ~3.7 (employment, current-or-planned, full controls). The controlled samples are 4,769 (employment) and 4,698 (revenue) firms. The "no controls" estimates sit about 1.5–2 ppt above the raw Chart 1 gaps. The post does not explain the difference, and the likeliest cause is the smaller controlled sample. Read the two charts as separate instruments rather than reconciling them.

The controls are informative in their own right. Younger and more profitable firms, and those with more than ten employees, hold more optimistic employment expectations. Past changes predict expected changes, and that control takes the most off revenue: the gap falls by more than a third.

The authors' optimism test, and what it does not exclude#

The post's strongest claim is that the association "is not due to optimistic owners being more likely to use AI." The test: proxy optimism with the owner's description of current financial condition. If optimism drove both AI use and rosy expectations, the AI gap should be larger among firms in good condition. It is the other way round:

The revenue differences are reported as statistically significant. This rules out one specific story, in which a good current position breeds both adoption and optimism. It does not rule out selection in general, for four reasons that go beyond the post:

• Financial condition is not optimism. A struggling owner who adopts AI may be the one with a turnaround plan, and a struggling firm can expect mean reversion. Either produces a large gap among fair/poor firms with no effect from AI. The post's own design cannot separate a growth-minded owner from a growth-producing tool.
• Planned adopters look like current users. Pooling firms that plan to adopt with current users leaves the gap almost unchanged (employment ~18.7 vs ~19.2 uncontrolled, ~9.3 vs ~9.9 fully controlled), and widens it on revenue under full controls (~17.5 vs ~13.8). Firms that have not used AI yet are about as optimistic as firms that have. That fits a forward-looking disposition selecting into AI better than it fits use generating optimism. (Compiler's reading. The post never shows planned-only firms alone, and the reference group changes between the two columns.)
• No dose-response. Among users, firms with AI fully or partially integrated, and those calling it important to production, are "equally likely" to expect growth as firms only experimenting, after controls. A treatment effect would be expected to scale with depth of use. A disposition would not.
• Realized effects are small in the same survey. 77% of respondents report no change in current labor costs from AI, and only 31% report increased sales from it. The expectations gap sits on top of modest reported effects so far.

None of this makes the association spurious. It means the measured result is "AI users are growth-expecting firms, beyond observables and past growth", and the causal half remains open.

The technology-challenge result#

Among AI users, those reporting an operational challenge "utilizing technology (e.g., website, social media, ecommerce, cybersecurity)" have a 16 ppt higher net revenue expectation than users who do not. The authors read this as AI helping firms past internal technology constraints. The comparison is within users only, with no matching figure for non-users who face the same challenge. A firm that reports trouble with ecommerce is also a firm trying to sell online, so the same selection reading applies. It is a hypothesis about mechanism, and the post presents it as one ("suggest").

Where it sits in the vault#

• Against Firm AI-Spend Intensity and Headcount Growth: a binary measure with a large gap, against a spend panel where only intensity moves headcount. Ramp × Revelio find low-intensity adopters indistinguishable from non-adopters on realized headcount. Here a yes/no adoption question finds a 10-ppt expectation gap, and depth of integration adds nothing. The two are consistent if expectations track who adopts while realized headcount tracks how intensively they adopt. They are inconsistent if the SBCS gap is read as a treatment effect. The same page's McKinsey grade adds a calibration caution in the other direction: the one forward AI-workforce expectation the vault can grade overshot its realization by about 2.3×. That survey asked large-firm respondents about AI-attributed change and found expected declines. The SBCS asks small-firm owners about total change and finds expected growth, so the two describe different questions and populations rather than disagreeing.
• Against AI and Market Power: the same add-controls design, a different outcome, a smaller drop. OECD's adopter market-share premium dies once broadband, digitalisation and lagged productivity enter. The SBCS expectations gap falls by about half under owner, firm and past-outcome controls and survives. The two papers control for different things. OECD's killing variable, lagged productivity, has no direct counterpart here. The closest one, past employment or revenue change, takes the most off the SBCS gap too.
• Against The Automation–Optimism Link: the firm-owner version of a worker-level pattern. The AEI Cadences survey finds the heaviest delegators the most optimistic, and cannot separate delegation-breeds-optimism from optimists-delegate. The SBCS has the same identification gap at the firm level, and the authors' financial-condition test is a partial attempt at it that page's survey does not make.
• Against Function-Selective AI-Era Headcount Cuts: the other end of the size distribution. AI-citing public incumbents cut 10–20%. Small employer firms using AI expect to hire. Different populations, and one side measures announced cuts while the other measures expectations, so this is not a contradiction.
• On the adoption level. 46% of US employer firms using AI is far above the firm-weighted Census BTOS benchmark (18%) and OECD's 20.2% of enterprises, and near Ramp's ~55% of eligible businesses with AI spend (Firm AI-Spend Intensity and Headcount Growth, AI and Market Power). The SBCS question counts use by the business or its employees, so informal employee use counts as adoption, which a procurement or spend instrument would miss.

Connections#

• Firm AI-Spend Intensity and Headcount Growth — realized headcount by spend intensity against expected headcount by a binary adoption flag, and the McKinsey expectation-versus-realization grade that says how far to trust a forward workforce expectation
• AI and Market Power — the same "watch the adopter premium under controls" sequence; there it dies on lagged productivity, here it halves on past growth and survives
• The Automation–Optimism Link — the worker-level adoption–optimism association with the same selection-versus-treatment problem, and no counterpart to the financial-condition test
• Function-Selective AI-Era Headcount Cuts — the incumbent end of the size distribution, cutting while small AI-using firms expect to hire
• The Enterprise AI Adoption Gradient — the large-firm adoption instrument; the SBCS supplies the small-employer-firm adoption rate and owner expectations that enterprise admin records cannot see
• Stanford Digital Economy Lab — the Canaries early-career payroll decline the post cites as its motivating contrast with flat economy-wide measures

Open Questions#

• Do the late-2025 expectations come true? Settler: the 2026 SBCS (fielded autumn 2026), reporting realized past-12-month employment and revenue change by AI-use status. If users' realized net change falls short of the 32.7 / 48.2 they expected by more than non-users' does, the gap was a disposition, not a forecast.
• Are planned-only adopters as growth-expecting as current users once separated out? Settler: the SBCS microdata, or an authors' follow-up, splitting current / planned-only / neither. A planned-only gap close to the current-user gap would point to selection over treatment.

Sources#

• AI Adoption and Employment Expectations: Evidence from a Survey of Small Business Owners — Will Aarons & Asani Sarkar, AI Adoption and Employment Expectations: Evidence from a Survey of Small Business Owners, Liberty Street Economics (Federal Reserve Bank of New York), 2026-10-08, empirical, ~1,500 words over three charts. Cited for Chart 1 (diffusion indexes, labels viewed at compile and matching the transcription), Charts 2–3 (point estimates and CIs read off the images, ±0.5 ppt; viewed at compile), the financial-condition split, the integration-depth null and the technology-challenge result (prose only, no table). Adoption shares, the 77% no-labor-cost-change and the 31% sales-gain figures are the SBCS 2026 employer-firm report as quoted in the post, not checked against the report.