Nearly half of small businesses now say they use AI somewhere in the company. Read past that headline number and the Federal Reserve's own data gets a lot less flattering. In the 2026 Report on Employer Firms, published March 3, 2026 and drawn from the 2025 Small Business Credit Survey, the twelve Federal Reserve Banks found 46% of employer firms with 1 to 499 employees currently use AI in some part of the business. Ask how far that use actually goes and the number that matters drops from 46% to 7%, the share of those AI users who told the Fed they'd fully integrated it into how the business runs. Everyone else is still experimenting or stuck halfway, which is another way of saying most of that 46% doesn't have software worth attaching AI to yet.
What the Federal Reserve actually asked
The Small Business Credit Survey (SBCS) is an annual survey run jointly by the twelve Federal Reserve Banks, fielded September 3 to November 14, 2025, with 6,525 employer firms responding across all 50 states. It's a different instrument than the Census Bureau's Business Trends and Outlook Survey, which asks a simple yes-or-no on recent AI use and which we covered in a look at where that gap sits by firm size. The SBCS asks something more useful for a founder deciding what to build next: not just whether the business uses AI, but how deep that use actually goes.
The 7% hiding inside the 46%
Of the firms that said yes to using AI, the Fed asked a follow-up on maturity. About half described their business as still experimenting with AI. Another 44% said they'd partially integrated it into business processes. Just 7% of AI users called it fully integrated into the business. Stack those against each other and the honest read of "46% adoption" is that roughly 3% of all employer firms in the survey have AI doing real, load-bearing work, with the other 43% somewhere between poking at a tool and half-wiring it in.
What that AI use actually looks like day to day
The survey also asked what AI users actually do with it. The three most common tasks were writing or marketing (83% of AI users), individual productivity (61%), and planning or analysis (51%). None of those are a core operational workflow. They're person-level tasks: someone drafting a product description, summarizing a spreadsheet, outlining a plan. That's not a criticism of the businesses doing it, it's a reasonably efficient use of a general-purpose tool. But it also explains the maturity numbers above. A person opening a chatbot to write ad copy doesn't need an operational system to plug into. A business trying to get AI into order fulfillment or customer support does, and most small, founder-run companies don't have that system built yet.
Why the number stalls at partially integrated
The Fed also asked AI users what's holding them back from going further. The top two answers were accuracy concerns (46%) and difficulty adapting the tool to their business (43%). That second number is the whole story in one line. Adapting a tool to a business assumes the business has an existing operational system for the tool to adapt into: an order database, a support queue, an inventory record. A company that's still running fulfillment through a shared inbox and a spreadsheet doesn't have anything there to adapt into, so the AI tool sits on top of nothing and stalls at partially integrated, sometimes for good. This is where I'll push back on the pitch a lot of AI vendors are selling right now, that any business can bolt on an AI feature and see it become part of how the company runs. The Fed's own numbers say that's true for the 7%, and it's the exception, not the rule.
A worked example
Picture a nine-person specialty pet-food subscription company, about eight months past a seed round, run by a founder with a retail background and no engineer on staff. The founder already uses a general chatbot for product descriptions and email copy, which is exactly the 83% use case above. Subscription pauses, address changes and refund requests all come through a shared support inbox, sorted by hand every morning. Adding an AI support tool sounds like the obvious next step, and the founder tries one. It answers generic questions fine and has no idea whether a given customer's subscription is paused, what their next ship date is, or whether a refund already went out, because none of that lives anywhere the tool can read from. Six weeks in, it's still "partially integrated": turned on, occasionally useful, not trusted with anything that matters. That's not a tool problem. It's that the operational system the tool needed to plug into was never built.
What actually gets a small business to fully integrated
The fix in that example isn't a better AI vendor, it's building the underlying system first: subscription state, order status, refund logic, as real software with a database behind it, scoped to what the business needs this year rather than a platform built for a company five times its size. AI becomes worth adding once that system exists for it to read from and write back into, not before. Worth saying plainly where this doesn't apply: if a business already runs on solid software (a proper e-commerce platform with clean order data, say) and the gap is narrower, wiring one AI feature into an existing system, that's a smaller integration project, not a ground-up build, and paying for the bigger one would be the wrong call.
If the honest read of your operation is closer to a shared inbox than a system, what it actually takes to scope a build like that is the more useful next read, and if it's already costing you time every week rather than sitting as a future risk, the signs it's become the actual bottleneck is worth a look first. We scope this kind of build, fixed-scope, so a founder without a technical background always knows what they're buying, under custom software development, with transparent pricing and an itemized estimate within 48 hours of a call.
Sources
- Federal Reserve Banks: 2026 Report on Employer Firms: Findings from the 2025 Small Business Credit Survey (published March 3, 2026)