AI Adoption Is Broad, but Deployment Is Narrow
Executive summary
The business AI story is neither “almost nobody” nor “everyone.” Census research estimates that 18% of employer firms used AI during the November 2025 to January 2026 reference period. Employment weighting raises the figure to 32% because large firms use it more.
Depth is the constraint: 57% of adopters used AI in three or fewer functions. Sales and marketing led at 52%, strategy and business development at 45%, and IT at 41%. Tool access is spreading faster than end-to-end redesign. Source
The market in context
Census broadened its question from AI in producing goods or services to use in any business function, so older and newer estimates should not be spliced casually. The May 2026 summary reported national use near 19.8%, Information at 39.7%, Finance and Insurance at 33.9%, and Retail near 14%.
These are shares of firms, not model calls or workers personally exposed. One team using an assistant and a company rebuilding many processes can both answer yes.
The difference between firm-weighted and employment-weighted adoption is economically important. It means a worker is more likely to sit inside an adopting organization than the headline share of firms suggests, because large employers account for more jobs. It also means diffusion can look broad in labor-market terms while remaining difficult for a large number of smaller businesses. Those firms often lack a security team, dedicated data engineering or a program manager who can redesign work across departments.
Adoption measures answer a participation question, not a productivity question. A firm counts as a user whether AI supports a handful of employees or a core operating process. That breadth is useful for tracking diffusion, but it should not be interpreted as equivalent doses of technology. The next stage of measurement needs intensity, workflow coverage and realized outcomes.
Firm size creates a diffusion advantage
Firm size creates a diffusion advantage. Large firms can spread security review, integration, procurement and training across more workers. The 32% employment-weighted rate versus the 18% firm rate quantifies that tilt; it does not prove large firms use AI well. Source
For small firms, implementation capacity can matter more than model price. A packaged narrow workflow may outperform a broad assistant that requires the buyer to invent the operating model.
Large firms also have more opportunities to learn across functions. A security review completed for one division can lower the cost of a second deployment. Shared data infrastructure and procurement can spread fixed implementation expense. Yet scale introduces its own drag: fragmented systems, approval layers and inconsistent local processes can keep a licensed tool from becoming a common workflow.
For smaller firms, the winning product may be less configurable and more opinionated. A tool that arrives with a defined workflow, clear data boundary and measurable output can create value faster than a general assistant. The tradeoff is dependence on the vendor’s process. Buyers need exportability and a way to keep operating if pricing or product scope changes.
Adoption clusters where language is already digital
Adoption clusters where language is already digital. Sales, strategy and IT produce text, code, research and plans accessible to general models. Yet faster drafts may not shorten campaign launch; faster code may leave testing as the bottleneck.
The 57% concentration in three or fewer functions maps unfinished integration. Quote-to-cash, returns or onboarding cross data and authority boundaries. Those are harder than deploying a copilot and more likely to move cycle time. Source
Function counts can understate cross-functional difficulty. A customer-return workflow may touch service, inventory, payments and finance even when one team owns the interface. Automating the first response is easy; deciding eligibility, approving an exception, updating stock and reconciling a refund requires shared data and authority. This is where narrow deployment shows up as an operating constraint rather than a lack of employee enthusiasm.
The right unit of redesign is therefore an accepted outcome. In marketing, count campaigns approved and launched, not copy generated. In software, count changes shipped without higher incident rates, not code suggested. In service, count resolved contacts and repeat-contact rate, not summaries produced.
Sector gaps will shape competition
Sector gaps will shape competition. Information and finance report much higher use than retail. Task composition, firm size and regulation explain part of the difference. Retail should prioritize catalog, forecasting, associate assistance and exceptions rather than copy a software company’s office use cases.
High adoption in finance does not imply unrestricted autonomy. Privacy and model risk raise the cost of mistakes, favoring controlled explanation before autonomous decisions.
Sector comparison should begin with task mix and cost of error. Information businesses produce digital artifacts that models can read and change. Retail combines digital decisions with physical inventory, labor schedules and store execution. Finance has abundant language and data, but a mistake can trigger regulatory, credit or customer harm. Similar adoption rates would not imply similar autonomy or economics.
This creates room for vertical systems. A retailer may need product-attribute normalization and exception routing; a financial institution may need traceable evidence and approval; a small professional-services firm may need document intake and billing. General models supply capability, but sector-specific workflow determines whether that capability reaches the income statement.
Implications for operators
Build a workflow inventory rather than a tool inventory. For each process record volume, time, error cost, data, decision rights and exceptions. Assign a business owner and include downstream reviewers.
Measure task speed, workflow performance and business outcome separately. A drafting tool can save minutes while launch time stays flat. Small firms should favor clear workflows, transparent pricing and exportable work over complex stacks.
An AI portfolio should be managed as a set of workflow bets. For each one, record the starting cycle time, labor minutes, error rate, backlog and customer outcome. Name the person who can change the process, not merely the person who bought the software. Include the reviewers and exception handlers whose workload may rise when generation becomes cheaper.
Expansion should follow evidence at three levels. First, does the model perform the task accurately enough? Second, does the workflow move faster or cost less after review and exceptions? Third, does the business outcome improve? A strong task benchmark with no change in launch time or resolved cases is a deployment warning, not a success.
Leaders should also examine concentration. If adoption remains limited to a few functions, identify whether the blocker is data access, risk approval, integration, change management or weak economics. Each requires a different response. Buying another model will not solve an ownership problem.
What would change the view
Self-reported use is broad and wording changed. Firm-level adoption does not reveal intensity or quality. Later estimates may rise, but expected use can exceed realized deployment.
The thesis would weaken if broad assistant access produces sustained cross-functional margin or output gains without formal redesign and without higher error or review cost.
The Census results may change as question wording and respondent understanding evolve. Firms can interpret “use” differently, and rapid experimentation makes intensity hard to classify. The sector figures also reflect differences in establishment size and task composition, so they are descriptive rather than a clean measure of managerial quality.
A second risk is that narrow deployment may be temporary. General tools can diffuse informally before organizations formalize processes, and employees may create value that surveys or finance systems do not capture. The thesis should be revisited if broad access produces persistent gains in output, margin or service quality without corresponding workflow integration.
Methodology
Detailed function figures use the November 2025-January 2026 AI supplement. Later biweekly figures use the revised question. Source
A chart should show firm-weighted and employment-weighted use by size and sector, plus adopters using one, two, three and four-plus functions. Units are percentages of employer firms and employment; significance notes should accompany comparisons.
The 18% estimate refers to employer firms reporting AI use during the November 2025 to January 2026 reference period. The 32% figure weights by employment, giving larger firms more influence. The 57% figure describes adopting firms using AI in three or fewer business functions. These denominators are not interchangeable. A defensible update should keep the revised BTOS wording separate from earlier production-focused questions and report confidence intervals where available.
A useful operator dataset would add the number of employees with access, active weekly users, workflows in production, share of cases touched by AI and share completed without correction. Outcome fields should include cycle time, accepted output, error severity, customer impact and manual review. Report results by firm size, function and workflow complexity rather than collapsing experimentation and production into one adoption rate.
Longitudinal analysis should preserve the date of first use and each material process change. That makes it possible to distinguish a temporary learning period from a durable productivity effect and to avoid attributing an improvement to AI when pricing, staffing or demand changed at the same time.
Cost should include licenses, integration, data preparation, review, incident handling and vendor management. Benefit should use accepted work and business outcomes. This prevents a cheap model call from obscuring an expensive process and makes comparisons meaningful when usage shifts from occasional assistance to production volume.