AI BUSINESS
TRANSFORMATION
REPORT
AI transformation is still rare.
The pilot era is over.
Small businesses did not wait for a permission slip. Employees opened the tools, found the useful edges, and started using them. Leadership often discovered the deployment after the fact. That is not a strategy. It is a reconnaissance report.
Nearly half of U.S. small employer firms now report current AI use. Yet only 7% of the users say AI is fully integrated. That implies roughly 3.2% of small employer firms have crossed the line from use to operating model.
Most firms are still at personal leverage: writing, analysis, customer response, and the low-friction work that fits inside one browser tab. Useful. Not transformational.
Transformation starts when AI owns a recurring workflow, touches the systems where the work lives, follows written rules, knows when to stop, and has a named human responsible for performance after launch. A prompt is not an operating model. Neither is a tool license.
This report maps that gap with current evidence from U.S. government surveys, Federal Reserve research, OECD studies, transaction data, worker surveys, and field evidence from real small-business workflows. The findings point in one direction: integration depth matters more than access.
Every pilot has a launch date. Transformation begins when someone owns the watch in month seven.
FOUNDER & CEO, MARSHAL
Small-to-medium sized businesses only. No enterprise benchmark theater. The field has enough of that already.
Six numbers that define the field.
The argument fits on one page. Human behavior has somehow made the rest necessary.
The adoption range
Different surveys measure formal firm use, employee use, digitally active SMEs, or regular platform users. The headlines conflict because the denominators do.
The implied transformation core
46% of small employer firms use AI. Only 7% of those users report full integration. Multiplication is cruel, but useful.
The integration dividend
In one OECD country sample, company-wide AI users were 11.5 times as likely as isolated-task users to report transformational value.
The content-process imbalance
Writing and marketing use outpaces process automation four to one among U.S. small employer AI users.
The augmentation-autonomy gap
Small-business workers use AI for personal productivity 10.7 times as often as they use it for workflows with minimal human input.
The exposure-control gap
Digitally active SMEs report cyber breaches 4.4 times as often as they report advanced cyber readiness.
A field report, not a vendor survey.
The research base combines six survey programs, government data, platform transaction data, and workflow-level field evidence.
Scope
Only small-to-medium sized businesses. Source definitions vary, but generally cover firms from self-employed through 499 employees. Large-company evidence is excluded from the analysis.
Time frame
Research published from late 2025 through June 22, 2026, with underlying fieldwork spanning 2024 to May 2026.
Evidence hierarchy
Government and intergovernmental datasets first. Then nationally sampled worker research, platform transaction data, and peer-reviewed or accepted field studies.
Derived analysis
Ratios and implied shares are labeled as Marshal calculations. They are arithmetic on published values, not new survey observations.
Comparability
Unlike samples are never treated as a single trend line. Where different adoption rates are shown together, the point is measurement architecture, not ranking.
Limitations
Several sources are self-reported. The Federal Reserve sample is weighted but nonrandom. The OECD D4SME sample is digitally active and nonrepresentative. Correlation is not causation.
Small business is not a side market.
It is the operating base of the U.S. economy. The transformation gap is therefore not a niche technology story. It is a national productivity story.
Adoption is broad.
Transformation is not.
The same market can honestly be 19% and 77% adopted.
Each number is defensible. None measures the same layer of behavior.
U.S. Census, all firms, AI used in any business function during the prior two weeks. Overall use hovered from 17% to 20%.
Federal Reserve survey of employer firms with 1 to 499 workers.
National probability sample of adults employed at businesses with 2 to 499 workers.
OECD platform-user sample across 12 countries, heavily weighted toward self-employed and micro firms.
QuickBooks survey panel, January 2026. This is a regular-use measure among a digitally engaged SMB audience.
AI use rises with organizational surface area.
More employees create more workflows, more specialization, more data, and more places where a time-saving tool can earn its keep.
of U.S. firms with 0 to 4 employees reported current AI use in the Census measure.
of U.S. firms with 100 to 249 employees reported current AI use in May 2026.
What this means
The transformation opportunity grows before the company becomes large. Once a business has a real division of labor, coordination costs become visible. That is where workflow-level AI starts to beat isolated productivity tricks.
Association, not causation. Industry mix, digital intensity, and management capability also matter.
The workforce is often ahead of the operating model.
AI spreads from the bottom because one worker can adopt it in an afternoon. Governance still arrives by calendar invite.
ADOPTION
USE
The transformation core is about 3.2% of small employer firms.
That figure is an inference, not a direct survey response. It is also a useful antidote to the ambient claim that everyone has already transformed.
U.S. small employer AI users
Digitally active SME users
The integration
dividend.
Company-wide integration reported 11.5x the transformational impact.
In the OECD Japan subsample, impact rose sharply with the breadth of integration. The pattern is descriptive, but it is hard to miss.
This comparison comes from one country subsample and relies on self-reported value. It should be read as directional evidence, not a causal estimate.
Small firms are four times more likely to use AI for writing than process automation.
Writing is the entry point. Process ownership is the operating model.
writing or marketing use versus process automation among U.S. small employer AI users
Augmentation outnumbers autonomy 10.7 to 1.
The market is mostly using AI beside a worker, not instead of a workflow owner.
Marketing leads. Operational work follows.
Two SMB surveys tell the same basic story with different instruments: the first wave sits close to content and communication.
Share of current AI users
Share of regular AI users
What changes next
Marketing remains a useful proving ground because the feedback loop is fast and the data is already digital. The next layer is harder and more valuable: lead routing, customer onboarding, receivables follow-up, reporting, and the administrative relay between systems.
When AI enters the accounting workflow, the close gets faster.
A field study of 79 SMEs and more than 200,000 transaction records shows what integration looks like below the survey headline.
The economics of
recovered capacity.
Productivity appears before sales.
That sequence is normal. It is also where weak measurement systems stop.
AI changes the task mix before it changes headcount.
The dominant near-term effect is workload relief, skill change, and less reliance on outside capacity.
Routine volume shrinks. Review, exception handling, system tuning, and customer-facing work grow. Small teams feel this faster because one person's calendar is the operating budget.
The rational question is not “How many roles disappear?” It is “Which tasks leave each role, and what higher-value work replaces them?”
Time saved is an input. Reinvestment is the outcome.
Small-business workers use recovered time in several ways. Only some of them flow directly into revenue or throughput, which is why every business needs a capture rate.
RATE0%-100%
Define it before calculating ROI.
The capture rate is the share of gross time savings that becomes measurable business value: more output, faster response, avoided hiring, lower contractor spend, improved quality, or protected owner time.
There is nothing wrong with better work-life balance. There is something wrong with calling every saved hour a dollar of ROI.
A simple capacity equation for an SMB.
This is an illustrative model, not a promise. Replace every input with your own baseline before anyone starts celebrating.
Put your own business into the equation.
The web version updates live. The PDF preserves the default 25-person scenario.
Do not stop at labor cost.
The largest gains often come from speed-sensitive workflows: responding to a lead while intent is high, shortening onboarding, collecting receivables earlier, or closing the books before decisions go stale.
Build the
operating model.
Five stages from access to compound operations.
Do not skip a stage because a vendor demo looked smooth. Demos are designed to avoid the terrain where businesses actually live.
Tools appear through individual accounts. No approved use, baseline, or data rules.
GATE: inventory tools and data exposureWorkers use AI for writing, research, analysis, and individual productivity.
GATE: measure hours, quality, and repeatabilityAI handles a defined recurring task with a named owner and documented handoff.
GATE: baseline cycle time, exceptions, and reworkCross-tool workflows operate with permissions, approval gates, evals, audit trails, and exception queues.
GATE: prove reliability at production volumeA portfolio of managed workflows shares data, ownership, budget, and a performance review cadence.
GATE: tie the portfolio to capacity and P&L outcomesStart where repetition is high and judgment risk is bounded.
A six-factor score prevents the usual selection method, which is choosing whatever the loudest software demo happened to show.
| Workflow | Freq. | Time | Speed | Data | Judgment | Change | Score |
|---|---|---|---|---|---|---|---|
| Inbound lead response | 89 | ||||||
| Receivables follow-up | 87 | ||||||
| CRM follow-up and notes | 79 | ||||||
| Customer onboarding | 76 | ||||||
| Support triage | 73 | ||||||
| Recurring reporting | 72 | ||||||
| Content drafting | 64 | ||||||
| Hiring decisions | 34 |
Use the score as a forcing function.
It is illustrative, not universal. A regulated firm should weight judgment and data risk more heavily. A high-velocity sales team should weight speed more heavily. The point is to choose with a model, then document the tradeoff.
The cyber gap is already wider than the agent gap.
Basic hygiene is common. Operational control is not.
reported breach exposure versus advanced cyber readiness among digitally active SMEs
The training gap is an operating gap.
Only one in ten small-business workers report formal AI training, while privacy, unclear use cases, and skill gaps remain the leading blockers.
Sets the business outcome, risk tolerance, and budget.
Defines policy, edge cases, and acceptance criteria.
Monitors integrations, exceptions, evals, and changes.
Reviews sensitive actions and incidents.
One person can wear multiple hats in a 20-person company. The mistake is not role compression. The mistake is pretending “the team” is an owner.
A 90-day path from useful tool to owned workflow.
One workflow at a time. Controlled scope. Real production data. Boring ownership. That is how systems survive contact with Tuesday morning.
Map the work
- Inventory current AI use and accounts.
- Baseline cycle time, volume, rework, and outcome.
- Choose two candidate workflows.
- Name the executive and workflow owners.
Design the system
- Document policy, handoffs, exceptions, and permissions.
- Build a 100-case evaluation set from real history.
- Define approval gates and kill conditions.
Build and test
- Connect only the systems required.
- Run simulations and regression tests.
- Inspect failure clusters, not average demos.
- Train the humans who handle exceptions.
Launch under control
- Start with limited volume and explicit handoff.
- Compare against the baseline weekly.
- Track completion, intervention, rework, and value.
- Write the month-seven ownership plan.
Measure the system, the economics, and the risk.
A dashboard that only shows usage is a vending-machine counter. It tells you the buttons were pressed.
| Layer | Metric | Definition | Cadence | Direction |
|---|---|---|---|---|
| System | Workflow completion rate | completed / started | Weekly | >90% |
| System | Human intervention rate | reviewed / completed | Weekly | Declining |
| System | Exception rate | exceptions / runs | Weekly | Stable or down |
| Quality | Rework rate | corrected / completed | Weekly | Below baseline |
| Quality | Policy adherence | passed eval cases | Release | >95% |
| Economics | Cycle time | start to accepted finish | Weekly | >20% faster |
| Economics | Cost per workflow | fully loaded cost | Monthly | Down |
| Economics | Captured capacity value | hours x cost x capture | Monthly | Positive |
| Outcome | Revenue affected | qualified sales or retained revenue | Monthly | Up |
| Outcome | Cash timing | days to invoice or collect | Monthly | Down |
| Risk | Sensitive-action errors | count and severity | Real time | Zero high-risk |
| Risk | Audit coverage | logged / consequential actions | Monthly | 100% |
Report both average performance and failure distribution. One dangerous edge case can hide inside a respectable mean.
Keep a baseline. “Feels faster” is not a metric, despite its remarkable popularity in steering meetings.
Assign every red metric to a person and a date. Dashboards do not repair systems through moral pressure.
Five SMB AI shifts likely by the end of 2027.
These are forecasts based on the evidence in this report. They are not survey results. Reality retains its traditional right to be inconvenient.
Fully integrated small employer firms
Up from an implied 3.2% in the 2025 Federal Reserve sample.
CONFIDENCE 65%Agentic use among digitally active SMEs
Up from 3.6% in the OECD D4SME sample as packaged workflow agents mature.
CONFIDENCE 70%Process automation across small employer firms
Up from an implied 9.7%, calculated from current use and automation share.
CONFIDENCE 70%Workers receiving formal AI training
Up from 10% as informal use collides with security, quality, and policy risk.
CONFIDENCE 75%Robust or advanced cyber readiness
Up from 16% among digitally active SMEs, pushed by insurance, customer requirements, and agent permissions.
CONFIDENCE 60%The managed-operations layer becomes a category
SMBs will increasingly buy owned workflow outcomes instead of standalone AI access. The software market is crowded. The accountability market is not.
STRUCTURAL CALLThe next advantage
is ownership.
AI access is becoming ordinary. The tools will keep improving, prices will move, and every software vendor will staple an agent to the homepage. That part is already underway.
The scarce capability is operational: choosing the right work, connecting the systems, defining the rules, controlling the exceptions, measuring the economics, and keeping a human name attached to performance after launch.
Small businesses do not need a grand transformation program. They need a sequence of owned workflows that earn the right to expand.
Kurt Fischman
Kurt Fischman is the founder and CEO of Marshal. He writes and operates at the point where AI systems meet real business workflows, existing software stacks, and the inconvenient details that arrive after a pilot goes live.
His work focuses on helping founder-led businesses turn AI from an individual productivity tool into a governed operating layer.
Managed AI Ops for small businesses.
Marshal designs, builds, deploys, and runs AI systems on top of the tools a business already owns.
Primary evidence and methodological notes.
Each source uses a different sample and definition. That is a feature of the analysis, not debris to sweep under the chart.
U.S. Small Business Administration, Office of Advocacy
Frequently Asked Questions About Small Business 2026, February 2026. U.S. small-business counts, employment, GDP, and payroll shares. Small business definitions vary by industry and can include firms up to 1,500 employees, but the headline national counts use the SBA framework.
U.S. Census Bureau
AI Use at U.S. Businesses, May 26, 2026. Biweekly, nationally representative BTOS estimates. Current AI use hovered between 17% and 20% from December 2025 to May 2026. The article reports less than 20% use among firms with 0 to 4 employees and 32% among firms with 100 to 249.
U.S. Census Bureau, Center for Economic Studies
The Microstructure of AI Diffusion: Evidence From Firms, Business Functions, and Worker Tasks, Working Paper CES 26-25, April 2026. Experimental BTOS AI supplement based on a 1.2 million-business sample frame. Used here for formal-versus-worker adoption overlap, augmentation patterns, functional breadth, and organizational change. Descriptive associations are not causal.
Federal Reserve Small Business Credit Survey
2026 Report on Employer Firms, March 3, 2026. 6,525 responses from employer firms with 1 to 499 workers, fielded September 3 to November 14, 2025. Nationwide convenience sample, weighted but nonrandom. Used for current use, integration stage, task mix, outcomes, and implementation barriers.
OECD Digital for SMEs Initiative
Empowering SMEs in the Age of AI: The 2026 OECD D4SME Survey, April 2026. 2,018 digitally active SMEs across 12 countries. Nonrepresentative platform-user sample: 74% self-employed or micro, 18% small, 8% medium. Used for maturity categories, integration depth, value, cybersecurity, and digital barriers.
Primary evidence, continued.
OECD
Generative AI and the SME Workforce: New Survey Evidence, November 2025. Representative survey of more than 5,000 SMEs in seven countries, fielded in late 2024. Used for adoption by firm size, workload, staffing need, skill need, contractor reliance, performance, and revenue effects.
Intuit QuickBooks
2026 AI Impact Report, 2026. Seven survey waves covering 34,364 SMB owners and operators in the United States, Canada, United Kingdom, and Australia, plus administrative payment data for 5.335 million businesses. Small is defined as 0 to 9 employees and midsize as 10 to 99, except Australia where midsize extends to 50. Used for regular use, paid adoption, retention, productivity, revenue, workday, task mix, and barriers.
U.S. Chamber of Commerce Foundation
Main Street AI Monitor: Half of Small Business Workers Use AI, June 17, 2026. National probability sample of 1,070 U.S. adults employed at businesses with 2 to 499 workers, fielded May 8 to 11, 2026 through Ipsos KnowledgePanel. Used for worker adoption, bottom-up leadership, training, task type, time reinvestment, and barriers.
Choi and Xie
Human + AI in Accounting: Early Evidence from the Field, accepted for publication in the Journal of Accounting Research. Survey of 277 accountants and field data from AI accounting software serving 79 SMEs, covering more than 200,000 transactions. Used for client support, task reallocation, ledger granularity, monthly close, and confidence-based human intervention.
Marshal calculations and forecasts
Derived values: 3.2% = 46% current use x 7% fully integrated; 9.7% = 46% current use x 21% process automation; 21.1x = 76% novices / 3.6% champions; 11.5x = 23% company-wide transformational impact / 2% isolated-task impact; 4.0x = 83% writing / 21% process automation; 10.7x = 64% personal productivity / 6% minimal-human workflows; 4.4x = 22% breach exposure / 5% advanced readiness; 1.94x = 45.8% adoption among 50 to 249 employee SMEs / 23.6% among one-person firms. Forecasts are planning scenarios, not observed results.
All analysis, calculations, copy, charts, and design execution in this report are original to Marshal.
THE 2026 AI BUSINESS TRANSFORMATION REPORT | MARSHAL RESEARCH