Do Small Businesses Using Generative AI Need Fewer Staff?

Do small businesses using generative AI reduce staff needs?

Usually, surveyed small and medium-sized businesses (SMEs) using generative AI report changes in workload or skill needs before they report a change in total staff need. In the OECD’s late-2024 survey across seven countries, 83% said GenAI had no effect on how many staff they needed; 9.1% reported lower need and 5.5% higher need. Those are firm-reported directions, not counts of workers or a prediction for any employee. The practical answer is therefore conditional: capacity released by a tool can become more output, less overtime, changed responsibilities, reduced contractor use, slower hiring, or fewer staff. A time saving alone does not tell you which path a business will choose.

The distinction matters because two reasonable interpretations compete. One says that if software completes a task faster, the business will need fewer people. The other says that people will use the recovered time to serve more customers, check work, clear a backlog, or take on tasks the firm could not previously manage. Both can happen. Current SME survey evidence more often records no total staff-need change alongside workload and skill shifts, while a minority report lower need. A U.S. Census working paper adds a caution: deeper integration across business functions is associated with employment decreases, even though worker-level task use alone is not. The answer is not that small businesses are immune to cuts; it is that work redesign and headcount are separate stages.

For a worker, the useful early signal is not that someone has tried a chatbot. It is whether a recurring part of the work is now reliably handled inside an operating process, what happens to review and exceptions, and whether the employer is redirecting the resulting capacity or reducing labor need. Exposure, observed use, productivity, adoption, demand, and displacement are related but different facts. Begin by examining your own task bundle and the business’s choices rather than converting an industry headline into a personal job-loss estimate.

Sources: Generative AI and the SME Workforce: New Survey Evidence

What does the strongest SME-specific survey say about staff needs?

The OECD’s report, *Generative AI and the SME Workforce: New Survey Evidence*, presents the clearest recent cross-country answer specific to SMEs and generative AI. In a survey of more than 5,000 SMEs in Austria, Canada, Germany, Ireland, Japan, Korea, and the United Kingdom, conducted in late 2024, 83.0% of GenAI-using firms said it had no effect on the overall number of staff the company needed. Another 9.1% said staff need decreased, and 5.5% said it increased. The survey asked a knowledgeable business respondent, often an owner or manager, to report the direction of change. It did not count individual employees who lost or gained jobs.

The reduction share is larger than the increase share, and that difference should not be hidden behind the 83% no-change result. A genuine minority of surveyed users perceived lower labor need. But the survey did not ask how large an increase or decrease was. A small change in one small firm and a large change in another both count as a firm reporting a direction. The percentages therefore cannot be subtracted to estimate net jobs, expressed as a share of workers displaced, or used to calculate an individual’s odds of redundancy. The OECD itself says the missing magnitude prevents a net-effect calculation. The report also notes that slower hiring and attrition can be ways firms adjust labor needs, which may not appear as immediate layoffs.

The survey is informative because its topic and population fit this question unusually well. It asks about generative AI rather than AI as a broad category and includes SMEs in seven different national settings. The underlying results can be compared across country, size, and sector. Yet it is a voluntary survey, and the OECD notes that voluntary participation can attract technology-forward employers. Its headline use measure also begins with whether anyone at the firm, including the respondent, ever uses GenAI for work. A manager may see changes across a company more clearly than an individual worker, but retrospective perceptions are not payroll records or a causal experiment.

The denominator is firms that use GenAI, not all SMEs and not every employee in an adopting company. OECD reports an average 30.7% use rate among all SMEs in its covered countries under its broad measure, and firms with more employees were more likely to report use. One-person businesses deserve special care in interpretation: they more often reported reduced overall staff need, but for an owner working alone, “staff need” can describe the owner’s own workload or whether to hire, rather than an employee layoff. This category is evidence that the technology can substitute for some labor the business otherwise might buy; it is not interchangeable with a redundant payroll job.

The result supports a measured verdict: among the GenAI-using SMEs surveyed, no change in total staffing need was the modal response and reported reductions were a minority, though more common than reported increases. It does not support either an assurance that no small firm will cut staff or a claim that AI is already removing small-business jobs at a known rate. For an employee, the survey is a useful context signal. The more immediate question is whether the business has moved from trial use to recurring workflow change and what it does with the capacity that change creates.

Sources: Generative AI and the SME Workforce: New Survey Evidence

Why does ‘uses AI’ conceal several different business situations?

“A business uses AI” can describe anything from an owner drafting a one-off announcement to a system embedded in order processing or customer support. Those cases imply different exposure and different chances that the work process will change. OECD’s SME survey asks whether anyone at the company ever uses generative AI for work, including owner-initiated use. That measure is intentionally broad: a worker may use a public tool without formal company adoption, regular use, or a redesigned workflow. It is useful for counting exposure to the technology, but it does not say how much work has been automated.

The OECD report describes a concentration in text generation and reports that only 28.7% of GenAI-using SMEs used it in core revenue-producing activities. Use was often for simple, one-off, or less important tasks. The distinction between core and peripheral work does not by itself settle staffing effects: the OECD found core status made little difference to reported overall staff need. But recurrence and complexity did matter in its results. SMEs using GenAI mostly for recurring tasks were more than 50% more likely to report a decrease in staff need than those using it mostly for one-off tasks. Firms using it mostly for complex tasks were almost twice as likely to report an increase in staff need as those using it for simple tasks. This pattern fits a plausible mechanism: repeated automation may substitute for a stable queue of work, while complex use may complement skilled staff. It remains a survey association, not proof that task frequency caused the outcome.

The 2026 U.S. Census working paper makes the measurement issue especially clear by distinguishing firm use across business functions from worker use in tasks. For its November 2025 to January 2026 supplement period, the authors report that 18% of firms used AI in a business function, while 23% reported worker AI use in work tasks; those measures are not identical. Worker task use can occur without formal firm adoption, and firm-level use can occur without employees reporting task use. The study reports that writing, document analysis, and information search were common task uses, with most firms limiting use to a small number of tasks. Because the measure covers AI broadly rather than GenAI alone and firms of all sizes, it complements rather than replaces the OECD SME evidence.

A practical classification should therefore ask several separate questions. Who uses the system: an owner, a few staff members, or a business function? How often is it used: occasionally or in a daily process? Which task does it touch: a draft, an internal lookup, a customer transaction, or an end-to-end service? Does that task recur, and is it sufficiently important that the business will maintain and monitor the new process? Who checks the result, handles exceptions, and owns the outcome? Finally, has the workflow actually been changed, with approved data, access, training, and steps for failure, or is use still informal experimentation? Each answer narrows what an adoption headline can tell you.

Imagine a small retailer where one person uses a text tool to draft seasonal product descriptions. A second retailer uses a connected system to ingest product specifications, draft descriptions for hundreds of listings, route them for review, and publish approved text. The first case may save a short burst of writing time. The second may reduce a recurring production queue, but it also creates work in checking product claims, handling unusual products, maintaining data, and deciding who can approve publication. The example is illustrative, not a claim about an observed employer. Its point is that the word “use” compresses the very details that determine what changes.

This distinction also prevents a common analytical mistake: treating technical capability as business adoption. A system may be able to generate a plausible reply, but the firm may not permit it to see customer data, may find the errors too costly, or may prefer a person to handle the interaction. Conversely, a business can deploy a narrow tool with human oversight that makes a task faster without automating the whole role. For the reader, classify the workflow before looking for a career response. Isolated drafting may justify learning a permitted tool and checking its output. A recurring workflow being standardized deserves attention to task ownership, review, and whether the business is planning capacity or staffing changes.

Sources: How are SMEs using generative AI? Generative AI and the SME Workforce; The Microstructure of AI Diffusion: Evidence from Firms, Business Functions, and Worker Tasks

Why can workload change without a headcount change?

Workload is the amount, pace, or mix of work people experience; headcount is the number of workers a business says it needs. One can change while the other remains stable. In the OECD SME survey, 32.7% of users said GenAI reduced staff workload, while 11.8% said it increased workload. Both reports were more common than a reported change in total staff need. That pattern is consistent with time being absorbed in different ways, though the survey did not measure individual time use or prove why workload changed. Its respondents reported direction, not an audited accounting of hours saved and added.

A tool that helps produce a first draft can shorten one stage and leave the full task cycle unchanged. A worker who prepares product information may spend less time composing initial descriptions and more time checking sizes, materials, compatibility, and prohibited claims. A customer-service worker may draft replies faster but still need to inspect the account history, confirm the policy, decide whether a complaint needs escalation, and take responsibility for a promise made to the customer. These examples illustrate task mechanics. They do not establish that every business has adopted such systems or that review always cancels the time saved. Depending on the error rate, stakes, and workflow, review can be short, substantial, or itself partly assisted.

Workload can also rise when faster production changes expectations. If a business can prepare more proposals in a day, it might ask staff to handle a larger pipeline. If response times fall, more customers may start conversations or expect replies outside the old rhythm. Faster processing can expose a backlog that was previously too costly to tackle, so the next task is not a cut but clearing that backlog. A firm may also discover new tasks: configuring approved data, maintaining templates, investigating inconsistent outputs, or training colleagues. These are plausible routes from capability to work organization, not measured explanations for the OECD percentages.

Workload reductions can matter even when staffing is unchanged. A business may use a tool to reduce overtime, make a seasonal peak more manageable, or free staff for service work. The OECD discusses the possibility that GenAI reduces overtime but says the survey cannot test that outcome. That is a useful boundary: “less workload” could mean fewer hours, less pressure during the same hours, lower effort for the same output, or a perception that a task has become easier. It does not automatically mean more leisure, safer employment, or higher pay. Whether saved time benefits workers depends partly on how managers allocate it and what workers can negotiate.

The survey’s sector differences underline why an average is not a universal description. The OECD reports that agriculture and education SMEs were more likely to report workload increases, while finance SMEs were more likely to report decreases. These are patterns among surveyed firms, not predictions for an individual agriculture, education, or finance worker. Their task mixes differ, and the report’s question captures an employer-level judgment. In a service firm, for example, a drafting aid might speed routine written material; in a hands-on business, the main work may involve physical presence, safety, or customer interaction that a text system does not perform. Both might still use AI in supporting tasks, but exposure and workload changes are not the same.

The International Labour Organization’s June 2026 empirical review offers a broader check on the time-saved story. It synthesizes experiments, firm data, platform studies, and representative worker and firm surveys from several countries. Its summary says productivity gains are real but uneven and often not yet verified; worker-reported time savings of a few percent of working hours have not consistently turned into measured output, earnings, or employment. This matters because perceived minutes saved, observed task throughput, and business-level productivity are different measures. The ILO review is not SME-specific, so it helps explain why a workload claim should not be treated as a staffing result, rather than supplying an SME staffing estimate.

To understand your own workload, follow a task from intake to completion. Record the starting work, tool-assisted step, checking, rework, exceptions, handoff, and final accountability. Then ask what changed in practice: did overtime fall, did the queue shrink, did output rise, did the task move to another coworker, or did the firm add another service? A narrow tool trial that saves three minutes at drafting but adds two minutes of review may still be useful; it simply answers a different question from whether a job is disappearing. A whole-cycle view gives you concrete evidence to discuss workload boundaries, training, and responsibilities.

Sources: Generative AI and the SME Workforce: New Survey Evidence; The impact of GenAI on jobs, productivity and work organization: a review of the empirical evidence

When does task substitution become a plausible staffing change?

A staffing reduction becomes more plausible when a business can perform a recurring set of tasks reliably at scale, has integrated that work into a process, needs fewer labor hours after review and exceptions, and decides to capture the capacity as lower labor need. This is a chain of conditions, not a scoring formula. A tool’s capability is only the start. The output must be dependable enough for the stakes; someone must handle failures; the business must change its workflow; the saved capacity must exceed added checking and maintenance; demand must not call for more output; and managers must choose reduced hours, fewer contractors, slower replacement hiring, or lower headcount instead of expansion or workload relief.

That chain helps make sense of the OECD result. Most GenAI-using SMEs said total staff need had not changed, but a larger share said workload or contractor reliance changed. The report says 14% of users reduced their reliance on external contractors. That suggests one route by which a small firm may capture value without changing employee headcount: work that used to be outsourced moves in-house. In the opposite direction, some businesses reported increased reliance on contractors, including for custom systems or advice. The business may need outside support to install and govern a tool even if it later reduces a different kind of outsourced task. “Labor need” can therefore shift across employees, owners, contractors, and new specialist work.

OECD’s task-frequency finding gives a more specific reason to watch recurring work. Firms that mostly applied GenAI to recurring tasks were more likely to report decreased staff needs than firms that mostly applied it to one-off tasks. A recurring task forms a stable queue that can be measured, improved, and potentially consolidated. But frequency is not enough: the process may contain judgment, relationship, legal, or quality requirements that keep a person involved. And the survey does not show that every repeated task will be automated, only that firms reporting recurring use had a different association with staffing outcomes.

The 2026 Census working paper offers evidence about the business integration stage. In its nationally representative U.S. data, employment decreases attributed to AI were reported by 2% of firms. Most users, 66%, said they used AI solely to augment tasks. In its regression analysis, wider functional integration and operational investment were positively associated with employment decreases, while worker-task integration by itself had no significant link to headcount reduction once the other measures were considered. The finding is important because it distinguishes an employee using a tool from a firm building AI into its operations. But the paper covers AI broadly, firms of all sizes, and associations rather than causal effects; it cannot establish an SME-only GenAI rate or prove that integration caused a reduction.

Canada’s 2024 business survey shows the gap between task reduction and employment change from another angle. Statistics Canada asked businesses about AI used in producing goods or delivering services, not GenAI alone. Of businesses reporting this kind of AI use, 44.1% said it reduced tasks previously done by employees to a small extent and 39.2% to a moderate or large extent. Yet 84.9% of all businesses reported no change in employment after adopting AI for that purpose, and 6.3% of AI users said total employment decreased. The survey’s AI-use base was 6.1% of all businesses. These are retrospective firm reports and not a direct SME GenAI comparison, but they show why task change does not mechanically translate to headcount change.

There are several possible business responses after capacity is released. The firm can make the same staff produce more, serve customers faster, reduce overtime, catch up on a backlog, expand into new work, avoid a planned hire, use fewer contractors, or reduce employees. It can mix those choices across teams. A business with demand waiting and limited capacity may use the gain to expand service. A business with fixed or falling demand may decide the same output now requires fewer labor hours. Financial pressure, seasonal sales, owner priorities, customer standards, and the cost of mistakes all influence the choice. These mechanisms are economic reasoning that explains alternative outcomes, not evidence that a particular employer has selected one.

For a worker whose role includes repetitive digital tasks, the sensible action is to map where those tasks sit in the whole process. Look for a queue that is shrinking, a responsibility that has moved, fewer contractor assignments, a cancelled vacancy, or a repeated change in staffing plans. Alongside routine work, identify useful contributions in review, customer context, exceptions, process improvement, or accountable decisions. These activities can strengthen a role when they meet a real business need, but they are not permanent shields from change. The point is to make your skills legible within the redesigned workflow and to prepare a feasible alternative if the evidence turns against your current role.

Sources: Generative AI and the SME Workforce: New Survey Evidence; The Microstructure of AI Diffusion: Evidence from Firms, Business Functions, and Worker Tasks; Analysis on artificial intelligence use by businesses in Canada, second quarter of 2024

Which skills may become more important, and for whom?

The OECD survey suggests that surveyed SMEs more often saw GenAI increasing than decreasing the relative need for highly skilled staff. It reports 19.7% saying the need for highly skilled workers relative to lower-skilled workers had increased, compared with 9.4% saying it had decreased. Respondents also named data analysis and interpretation, and creativity and innovation, among skills that had gained importance. These are employer perceptions about relative skill needs. The OECD did not measure actual vacancies, wage premiums, credential requirements, or the hiring return to learning a particular skill. Its notes say respondents were not given a shared definition of “highly skilled,” so the term may have meant different things by firm, sector, and country.

A useful interpretation is that the value of a skill depends on where it sits in the work process. A person who knows a domain can frame the question, recognize missing context, check whether an output makes sense, explain its limits, and decide when it is unsafe to use. Data literacy helps someone inspect source quality and uncertainty. Communication and customer knowledge help turn a technically plausible answer into an appropriate service. Exception handling matters when the ordinary case is easy but the unusual one carries cost. Process ownership helps a worker see whether a faster step makes the complete workflow better or merely moves a queue downstream. These foundations are transferable because they apply across tools and sectors; they do not guarantee employment.

By contrast, a vendor’s button locations, prompt syntax, and interface labels can change. Learning a specific system is worthwhile when it is approved and tied to a recurring task, but it is a weaker long-term bet if separated from the business context. The goal need not be to become an engineer. A bookkeeper who wants to reduce time on routine reconciliation may need to understand data handling, output checking, and the firm’s accounting process. A communications worker may need to compare drafts against tone, factual accuracy, accessibility, and legal claims. A student aiming to build software products has a different need from a worker who only wants to use an approved tool in an existing profession.

The national Canadian results offer a practical clue about how smaller employers may adapt. Among businesses using AI in producing goods or services in the second quarter of 2024, 38.5% reported training current staff and 35.2% developing new workflows, while 8.2% reported hiring staff trained in AI. These data cover AI more broadly than GenAI and are not limited to SMEs, so they do not establish that a particular business will train rather than hire. They do indicate that the reported organizational response in that survey more often involved existing employees and process changes than hiring AI-trained staff. A small employer may distribute new responsibilities among people who already understand its customers and operations, rather than create a separate AI role.

That choice has both opportunity and risk. Existing workers can gain useful responsibility when they receive time, training, clear authority, and recognition. They can also inherit monitoring work and accountability without enough control over the system or workload. If routine assignments that helped junior colleagues learn are removed, the firm may save time now while weakening how future expertise is developed. The ILO’s 2026 review identifies unequal effects, potential erosion of younger workers’ opportunities, and changes to worker autonomy and job quality as concerns in the wider evidence. The review includes multiple methods and countries, but it is not a small-business causal estimate. It is a reason to ask who gets training, who checks results, and who carries the consequences when the process fails.

A proportionate learning experiment begins with the actual goal. If you want to use AI in your current field, choose one low-risk recurring task you are allowed to test. Compare a tool-assisted result with a known-quality baseline; record time for the entire cycle, including review and correction; and note which judgments still depend on your knowledge. If your goal is to build AI-enabled software, you may need a more technical project sequence in programming, data, product design, and testing. If you aim for software practice or machine-learning engineering, formal courses or a degree can provide deeper foundations and feedback, but a credential alone does not demonstrate readiness. For an existing worker, a small project that documents a real workflow improvement can be more relevant than an unrelated certificate.

The choice among degree, short course, certificate, project, and self-study should reflect the desired outcome, prerequisites, learning time, cost, need for feedback, and whether employers in your target location ask for that signal. A university degree can offer depth and structured foundations, but it takes substantial time and may not fit salary or care obligations. A course can fill a bounded gap quickly, but completion is not proof that you can apply the material. A certificate can signal exposure to a curriculum, while its value varies by employer and field. A project demonstrates application when it uses appropriate data and shows evaluation, limitations, and a clear result. Self-study can be low cost and flexible, but may need a peer, mentor, or work sample to provide feedback. None is a universal answer.

Keep the next step bounded. Choose a task whose failure is reversible, use only tools and data permitted by your employer, and check whether the experiment improves quality or time across the full process. If you cannot run it at work, a synthetic or public-data project may help you learn, but it will not prove the same workplace value. If your role is changing, also consider salary needs, location, training hours available, health, and family responsibilities before committing to a long program. Practical AI literacy should strengthen a coherent career path, not become an excuse to abandon hard-won experience or to buy training on the promise of a job outcome.

Sources: Generative AI and the SME Workforce: New Survey Evidence; Analysis on artificial intelligence use by businesses in Canada, second quarter of 2024; The impact of GenAI on jobs, productivity and work organization: a review of the empirical evidence

How can AI raise output while changing the value of experience?

A field study of customer-support agents illustrates how assistance can raise measured output and affect workers differently by experience. The working paper *Generative AI at Work* studied roughly 5,000 agents at one large software company. Agents given access to a conversational assistant resolved about 14% more customer issues per hour on average, with larger measured gains for newer and lower-skilled agents. The study’s setting was a specific customer-support operation and its rollout took place in 2020–21, before current GenAI tools were widespread. It measures task productivity, not whether the company reduced staff because of the tool. It cannot supply an expected productivity or headcount effect for a small business.

The result helps explain one way experience can be redistributed. A support assistant can surface response suggestions or procedures drawn from stronger prior examples, helping newer workers handle familiar cases. If that support reduces the time needed to learn routine cases, some know-how becomes easier to access. Experienced workers may still be valuable for unusual situations, policy interpretation, coaching, quality review, and service design, but firms may also change how many people need deep expertise in routine cases. The study did not establish which of these staffing choices follows from higher output. It shows why a productivity result is not itself evidence of job displacement.

The OECD SME report contains a related but different signal: 65% of GenAI-using SMEs said it improved employee performance, and 39.1% of SMEs with a skill gap said it helped compensate for that gap. These are employer-reported benefits, not randomized effects. They can coexist with increased skill needs: a tool may let existing workers do a task better while increasing the value of people who can interpret data, judge quality, or direct more complex work. The survey also found reported skill-need increases alongside reductions. Together, these findings caution against assuming that automation always lowers skill requirements or that assistance always protects a role.

Higher output can be spent in more than one way. Suppose a small service firm can prepare customer responses more quickly. If customers are waiting, the firm may handle more cases, shorten turnaround time, or extend opening hours. If the number of requests is fixed, it may reduce overtime or labor hours. If the market is competitive, it could lower prices and attract more demand; if quality assurance is expensive, it could instead keep staff levels stable and assign people to review. Those possibilities follow from basic business choices, not a measured outcome in the field study. The actual outcome depends on the firm’s demand, service promises, cost structure, and management decisions.

This is why “experience” should be made concrete. Years in a role can contain valuable pattern recognition, client context, knowledge of exceptions, and an ability to detect a misleading result. They can also contain routine habits that a new process makes less necessary. A worker can preserve the parts that solve current business problems by documenting how they evaluate a difficult case, which inputs matter, where errors occur, and when escalation is needed. Such documentation can support process improvement and make a person’s contribution visible. It cannot guarantee that the employer will preserve a job or reward the contribution.

An early-career worker may face a different exposure point from a senior colleague. If routine drafting, research, or first-line support tasks are often used to develop judgment, automating a portion may reduce opportunities to practice even if it does not immediately reduce total headcount. The ILO review flags younger workers’ employment opportunities as one risk in the broader empirical evidence. The NBER support study’s larger productivity gains for newer and lower-skilled agents are not proof that entry-level hiring will fall; the field study evaluated output among agents who were already there. Hiring, training, progression, and current productivity are distinct outcomes.

A realistic response is to build evidence of contribution across the whole workflow. A support worker can track not only response speed but accurate resolution, appropriate escalation, customer understanding, and repeat-contact patterns if those measures are available and fair. A marketing worker can show how review prevented a factual or brand error, or how faster drafting supported more tested campaigns. A small firm may have no formal performance dashboard, so a simple before-and-after record can help a manager conversation. Keep measures narrow and transparent; do not claim the tool caused an improvement if demand, staffing, or process changed at the same time. The durable asset is not familiarity with one interface but the ability to use domain knowledge to improve work and verify what changed.

Sources: Generative AI at Work; Generative AI and the SME Workforce: New Survey Evidence; The impact of GenAI on jobs, productivity and work organization: a review of the empirical evidence

What can national business surveys add—and what can they not prove?

National surveys add context, but their numbers should not be pooled as if they measured the same thing. The OECD asked SMEs across seven countries about generative AI and allowed broad workplace use to count. Statistics Canada asked about AI used in producing goods or delivering services. U.S. Census surveys have used changing definitions and periods, and its 2026 supplement separates AI use in business functions from worker task use. Each dataset can illuminate a different stage of adoption, but a comparison across them is not a single trend line in small-business job losses.

Statistics Canada’s second-quarter 2024 analysis used responses to the Canadian Survey on Business Conditions, collected online from April 2 to May 6, with 10,173 responses from a sample of 20,996 business establishments. Its 6.1% AI-use figure refers to businesses using AI to produce goods or deliver services over the prior year, not all occasional employee use and not GenAI alone. In that group, most businesses reported no employment-level change after implementation. At the same time, 44.1% said AI reduced some employee tasks to a small extent and 39.2% said tasks were reduced to a moderate or large extent. Training current staff and developing workflows were more commonly reported than hiring AI-trained staff. This offers corroboration that tasks and workflows can change while employment stays stable, but its population and measure differ from the OECD survey.

An earlier U.S. Census Bureau working paper using the Business Trends and Outlook Survey examined AI use from September 2023 to February 2024. Its findings, summarized in the saved evidence plan, describe uncommon use, organizational changes such as training and new workflows, and few firms reporting employment reductions; AI use was also associated with expansion. Because it predates later diffusion and measures AI broadly, it is best treated as an earlier snapshot rather than a direct forecast. Business surveys also depend on wording: one asks whether staff use a tool at work, another whether AI is part of producing a service. Broader questions often produce a different adoption rate because they count more activities.

The 2026 Census working paper provides a later, nationally representative U.S. view based on a supplement covering November 2025 through January 2026. The authors report functional AI use at 18% of firms and worker task use at 23%, with employment-weighted rates higher. Two-thirds of users said they used AI solely to augment tasks, and AI-related employment decreases were reported by 2% of firms. The analysis associates employment decreases with breadth of business-function integration and operational investment, while worker-task use alone had no significant link after these factors were considered. This is useful evidence that depth of organizational integration matters. It is a working paper, covers all AI rather than GenAI alone and businesses of all sizes, and its statistical associations do not establish causation or an SME-specific outcome.

The Federal Reserve’s March 2026 analysis asks a different question by linking online job postings with Census business survey adoption measures. It reports no evidence, in its study window, that firms or industries with higher AI adoption had lower total job postings. The authors explicitly caution that they study total postings rather than the occupations most susceptible to automation; a difficult job search in particular occupations could be offset by different hiring priorities elsewhere. Postings are not the same as hires, layoffs, or payroll employment, and an aggregate total can conceal shifts among roles. The result therefore weighs against a simple claim that AI adoption has already reduced overall hiring broadly in the measured U.S. data, while leaving room for concentrated effects.

Taken together, these sources point in a consistent but limited direction. In the observed periods, firm respondents often report stable total employment or staff need alongside changed tasks, training, or workflows. Some report lower need, and the newer Census study finds associations between deeper integration and employment decreases. None establishes a universal causal effect for small businesses using generative AI. Differences in dates, countries, survey populations, definitions, and outcomes prevent arithmetic pooling. Nor do U.S. and OECD-wide results predict the labor market in every city or country. Local demand, sector mix, business size, labor law, and the reader’s own workflow still matter.

Use official data to set context and employer evidence to guide near-term judgment. If your country publishes a business survey or labor-market measure, check its date, definition of AI, unit of analysis, and whether it measures reported plans, observed use, postings, headcount, or task change. Then compare it with what is happening at work: are tasks being removed, are queues growing, are coworkers being trained, are vacancies being replaced, are contractors being used differently? Aggregate evidence can warn against overconfidence in either direction. It cannot tell you what your employer will do next or whether a particular role is likely to be cut.

Sources: Analysis on artificial intelligence use by businesses in Canada, second quarter of 2024; The Microstructure of AI Diffusion: Evidence from Firms, Business Functions, and Worker Tasks; AI Adoption and Firms' Job-Posting Behavior

What evidence would make the staffing verdict change?

The verdict should become more concerned about staff reductions if repeated, comparable studies show that firms with deeper GenAI integration experience sustained reductions in payroll employment, hours, or replacement hiring relative to similar firms that did not integrate it, after accounting for demand and economic conditions. The strongest current exception is already visible in the evidence: OECD respondents reported decreased staff need more often than increased need, and firms using GenAI mostly for recurring tasks were more likely to report a decrease than firms using it for one-off tasks. The 2026 Census paper also links broader functional integration and operational investment with employment decreases. These findings deserve attention. They do not yet show a causal SME effect or tell us how many workers were affected.

A stronger test would follow businesses over time and distinguish planned changes from realized ones. Researchers would need to identify when and how GenAI enters a workflow, what tasks it changes, how much human review and exception work remains, and whether demand for the service changes. They would then observe hiring, exits, payroll, hours, contractor use, output, and service quality separately. A reduction in contractor spending is not the same as an employee layoff; a vacancy not filled is different from a current worker losing a job; higher output with stable employment is different from lower labor need. Comparison firms should be similar in size, sector, location, demand, and baseline digital capability. Even then, a firm may adopt AI because it is already growing or restructuring, so causal interpretation requires care.

The temporal horizon also matters. A firm can report no current staff-need change while redesigning tasks or slowing future hiring. Conversely, a business may reduce employment for reasons unrelated to AI even after adopting a tool. Retrospective surveys are useful for perception and early changes, but they cannot by themselves disentangle these paths. Linked employer-employee records can help show employment and hiring outcomes; worker surveys and interviews can reveal task quality, autonomy, hours, and training access. Neither source alone is enough. Public evidence should make clear whether it observes task capability, employee use, organizational adoption, employer plans, vacancies, or actual employment.

The broader ILO review provides a balanced current frame: it finds real but uneven productivity gains, limited large-scale displacement in the evidence reviewed so far, and potential risks that are unevenly distributed. Its warning about younger workers’ opportunities is particularly relevant because entry-level work can be altered before a whole occupation shows a large employment decline. If future linked data show persistent drops in early-career hiring in particular SME sectors after workflow integration, the practical advice should become more specific for those roles. That would still not establish a universal risk for all small businesses or all workers with similar job titles.

For an individual, the evidence worth watching locally is concrete and repeated. Has a stable task queue been eliminated rather than shortened? Are people no longer asked to do it, or does it reappear as checking and exception handling? Are fewer contractors retained for the same work? Have open roles been cancelled or replacement hiring slowed over several cycles? Is workload redistributed, has output expanded, or have managers explicitly consolidated responsibilities? Is the change limited to a pilot or built into the process? One isolated event may reflect seasonality, cost pressure, or a manager’s temporary experiment; a pattern across work cycles is more informative.

It is also reasonable to examine the quality of change, not only job counts. Stable employment can coexist with faster pace, weaker autonomy, more monitoring, or responsibility for errors without decision authority. Conversely, the same headcount can coexist with less overtime, training, and more useful work. The OECD workload results and ILO synthesis both make clear that employment totals alone do not describe the employee experience. A good workplace conversation asks how success will be measured, what review is required, how data is handled, who may override a result, and whether the time saved will reduce pressure or simply increase throughput.

What would change this article’s answer? A credible body of longitudinal evidence that comparable SMEs which integrate GenAI into recurring work reduce staff hours or hiring more than non-adopters, with demand and broader restructuring accounted for, would justify shifting from “workload and skills often change first” toward naming the settings where staffing reduction is demonstrable. Evidence that reductions are mostly avoided through output expansion, overtime reduction, or new services would strengthen the current interpretation. Until those comparisons exist at scale, treat today’s reduction reports as a signal to monitor, not a universal employment forecast.

Sources: Generative AI and the SME Workforce: New Survey Evidence; The Microstructure of AI Diffusion: Evidence from Firms, Business Functions, and Worker Tasks; The impact of GenAI on jobs, productivity and work organization: a review of the empirical evidence

What is a realistic next move if your small business is changing work?

Start with a task inventory, not a career label. List the recurring work you do, then note which steps are being tested or integrated with AI, what review and exceptions remain, and whether time saved changes overtime, throughput, contractor use, or replacement hiring. Choose one low-risk, permitted task for a bounded learning experiment. If your responsibilities or the firm’s direction point elsewhere, compare an upgrade in place, an adjacent role that uses your existing experience, and a larger change against salary, location, training time, cost, health, and family constraints. The evidence favors neither panic nor passivity: small businesses often report workload and skill change before changing total staff need, but recurring workflow substitution can matter. The free [task-level change-pressure checker](/ai-job-risk-checker) can help organize the first review; it reports task signals, not a validated probability of displacement. An optional [career roadmap](/career-roadmap) can compare scenarios against your constraints, without guaranteeing work or income.

Sources: Generative AI and the SME Workforce: New Survey Evidence