Small Business AI Use Is Broad. Full Integration Is Rare

The adoption debate has moved

The Federal Reserve's 2026 Report on Employer Firms says 46% of small-employer firms reported that their business or employees currently used artificial intelligence for work. Another 15% planned to start within the next 12 months. Adoption is no longer confined to a thin early group.

That does not make every user advanced, or every nonuser late. It changes the useful question. Instead of asking whether a company tried AI, ask how much of a process the tool completes, how output is checked, and whether improvement is measurable.

What the Federal Reserve measured

Fielded from September 3 through November 14, 2025, the survey covers 6,525 responses from a convenience sample of employer firms with one to 499 workers across all 50 states and the District of Columbia. The Fed says the sample was not random.

The AI questions appeared in an optional end module, so not every respondent answered them. The definition included stand-alone applications and AI features inside existing software. Someone using an embedded feature therefore counts even without a formal company AI program.

Three camps

The Fed split is 46% currently using AI, 15% planning to start within the next 12 months, and 33% with no plans. These are positions, not grades. Each group includes deliberate choices and unresolved operating decisions.

Current users are not one operating class

Among users, writing or marketing led at 83%, followed by individual productivity at 61% and planning or analysis at 51%. Multiple selections were allowed. Broad adoption can coexist with shallow integration because an employee still supplies context, judges output, transfers results, and preserves continuity.

The next year group needs a process

Among firms planning adoption, 54% named finding suitable tools as a challenge, while 37% named implementation or training time. Begin with a recurring process, its input, expected output, approver, and error cost. Explicit boundaries narrow product selection and give demonstrations something concrete to prove.

No plans can be rational

One third reported no plans. Of that group, the report says more than half said AI was not applicable, and 30% preferred not to use it. Variable physical work, severe error costs, or little repeat volume may deserve other investment. Test only processes with repetition, digital information, and measurable output.

The Census Bureau reports a lower share

A federal cross check keeps the 46% figure from becoming a slogan. The Census Bureau release on its AI supplement says supplemental questions ran from November 17, 2025, through February 8, 2026. The BTOS data page describes a nationally representative survey of US employer businesses outside farming.

A later Census analysis of six months of BTOS data placed overall business AI use between 17% and 20% from December 2025 through May 2026, far below the Fed's 46% for small employers. The disagreement must be stated, not resolved by choosing the larger estimate.

Source	Population and timing	Use measure	Reported share
Federal Reserve SBCS	Small employer firms, surveyed in fall 2025	Business or employees using AI for work at survey time	46%
Census BTOS	US nonfarm employer businesses, measured repeatedly	AI used in any business function during the past two weeks	17% to 20%

Definitions travel with the number

The surveys differ in population, reference period, sample design, and wording. The Fed covers small employers and includes software features. BTOS covers employer firms of every size outside farming and asks about activity during the past two weeks. Their estimates are not interchangeable.

Depth is the sharper signal

Among Fed respondents using AI, the report says about half were experimenting, while 44% were partly integrated and 7% were fully integrated. The report states that first share in words, not as a number. The report illustrates experimentation as testing tools or pilots, partial integration as use in some processes, and full integration as AI embedded in business functions.

These are self-classifications, not technical audits. Still, they expose the operating gap. Adoption can occur when one person opens a tool. Integration requires a defined process, permissions, controls, ownership, and evidence that the redesigned path improves something valuable.

Experimenting depends on a person

During experimentation, someone usually carries the process. They paste material into a chat window, refine instructions, inspect the answer, and copy it elsewhere. Without that person, the old workflow returns. Shared instructions, logs, fallbacks, and acceptance standards may not exist.

Experimentation is useful for locating strengths and failures. It should end with a decision, not permanent theater. A serious test produces a candidate process, baseline, examples of acceptable and unacceptable output, and a choice to stop, revise, or enter controlled operation.

Partial integration has a process and a gate

At partial integration, AI handles a defined portion of recurring work, perhaps classifying a request, drafting a response, extracting fields, or preparing analysis. Someone reviews the result before customer delivery, a core record change, a payment, or another consequential action.

This can be a sound destination. Sensitive information, judgment, safety, or expensive mistakes can justify permanent approval. The relevant test is whether the gate has clear rules, sufficient capacity, and an accountable owner, not whether every human decision disappears.

Full integration closes the routine path

For a stricter operating test, full integration means a defined input travels through the routine path, reaches the next system, and produces a usable record or report without manual transfer. The process exposes status, preserves traceability, and routes exceptions instead of failing silently.

That does not require autonomous handling of every case. A low-risk path can run end to end while uncertain, sensitive, or high-cost cases stop for review. Full integration finishes the chosen path and its controls rather than maximizing automation.

Reported gains are signals, not audited results

Among AI users, 71% reported increased productivity, 39% reported improved quality of goods or services, and 31% reported increased sales. Productivity ranks first, quality second, and sales third. These self-reported outcomes are not audited results, and the survey does not establish that AI alone caused them.

Faster drafts may wait longer for review or require more correction. Sales can move for unrelated reasons, so every process needs a baseline.

Test productivity closest to the work

Productivity sits nearer the changed task than revenue, so it can usually be tested earlier. Count accepted units and all staff time required for completion. Generated drafts are not completed work when substantial correction, approval, or transfer remains.

Quality and sales need concrete definitions. Quality might be the share accepted without correction. A sales workflow might track qualified handoffs or accepted proposals, not messages produced. Select a downstream metric close enough to reveal whether the redesigned process helped.

Why pilots stall before full integration

The Fed found accuracy challenged 46% of current users, while 43% reported difficulty adapting tools to business needs. It does not assign reasons to adoption depth. Operationally, three design failures commonly prevent a promising test from becoming a measurable process.

Failure one: no baseline

Once a pilot starts, the old process may vanish before measurement. Nobody can determine whether cycle time, staff effort, rework, or downstream results improved. First observe two ordinary weeks, counting completed units, staff time, waiting, and corrections. This is practitioner guidance, not a Fed finding.

Failure two: no countable output

Goals such as help marketing or make the team efficient identify no finishable unit. Choose a countable object, such as an approved description, reconciled invoice, qualified request, or completed service summary. Define its start, finish, and acceptance standard in terms two employees apply consistently.

Failure three: no owner for approval

An unnamed review step creates a queue, not control. Outputs wait, standards vary, and exceptions circulate. Assign a role authorized to accept, reject, correct, and stop the process. Document acceptance and escalation rules, including errors always requiring review and routine cases eligible for narrower approval.

A depth ladder to run on Monday

Choose one recurring process, not a department or ambition. Keep the test narrow enough that the same completed unit is observable before and after the change. Then follow the ladder in order:

1. Name one process. Define its trigger, required input, completed output, destination, and owner. Exclude adjacent work that blurs the result.
2. Count current output for two weeks. Capture volume, elapsed time, staff effort, corrections, and the business result attached to each completed unit.
3. Run the tool in shadow mode. Produce an output beside the live process without sending, publishing, paying, or changing the system of record.
4. Add a human approval gate. Give a named role acceptance rules, access boundaries, an exception route, and authority to stop the pilot.
5. Narrow the gate where error cost is low. Advance routine cases automatically only when shadow results show stable performance. Keep consequential or uncertain cases in review.

Shadow mode separates learning from exposure

Shadow operation compares outputs without letting an untested result affect a customer or core record. Reviewers tag rejection reasons, identify missing context, and revise instructions. Sensitive data remains outside any tool until access, retention, and permitted use are understood.

When results stabilize, narrow approval by case type. Routine cases with reversible consequences may pass automatically. Ambiguous requests, unusual amounts, privacy concerns, or irreversible actions should stop. This exception map is part of integration, not an afterthought.

Four numbers to review after thirty days

Compare the pilot with its baseline using one definition of a completed unit. Review four numbers, not activity counts that rise merely because a tool is available:

• Completed output. Accepted units reaching the defined finish line.
• Human effort. Staff minutes per completed unit, including review and correction.
• Exceptions and rework. The share requiring correction, rejection, rollback, or escalation.
• Downstream result. The countable outcome the process should affect, such as an accepted proposal, qualified handoff, resolved request, or clean record.

If none moved usefully, the pilot failed and should stop rather than widen. If speed improved while rework or downstream performance worsened, the design also failed. A faster intermediate task is not a better process when its cost appears later.

Call failure clearly

Stopping one pilot is not a companywide verdict on AI. It shows this tool, process, data, or control design did not earn expansion. Record the failure, preserve safe examples, and consider a narrower test. Do not maintain weak workflow to protect the original decision.

Experimenters need a baseline and one named process. Partly integrated firms need an owned gate and exception map. Firms claiming full integration need evidence that the routine path closes, reports status, and improves a business measure.

Finish one process

The surveys neither say every business should adopt AI nor agree on one universal share. The Fed finds common use among small employers, with only 7% of users describing full integration. Census shows why every headline number needs its definition.

Trying is no longer scarce. Finishing is. Name one process, measure the old path, test the replacement without exposing the business, install an owned approval gate, and widen only what evidence supports. The next meaningful milestone is one completed, controlled process.