Key findings

• A strict definition prevents chatbot use from being mislabeled as agent adoption.

• Observed 2026 evidence shows broad AI use but early agent deployment.

• The base case reaches 15%-25% of SMBs with at least one production agent by 2028.

• Security, identity, data access, and exception handling are the binding constraints.

• Support and sales operations are likely to lead because work is digital and measurable.

What counts as an agent in this forecast

This forecast counts a production agent only when it chooses among permitted steps, calls a business tool, and advances a recurring workflow under a named owner.

The deployment can still require human approval. In fact, supervised action is likely to be the dominant SMB pattern through 2028. Examples include an agent that researches an inbound account and prepares a CRM record, a support agent that resolves approved issue types, or a finance agent that matches routine documents and routes exceptions. The definition focuses on operational action, not marketing terminology.

The denominator is all SMBs, making the ranges more conservative than surveys of AI users. They are planning assumptions, not measured market share.

Observed 2026 baseline

Representative firm data show the foundation for agents but do not directly measure this strict definition. The U.S. Census AI supplement found 18% of firms used AI in a business function in late 2025 and early 2026. Among adopters, 57% used it in no more than three functions. The 2026 Stanford AI Index, drawing on organizational survey evidence, reports agent deployment in single digits across nearly all business functions.

The OECD's 2026 D4SME survey found that most participating SMEs used off-the-shelf applications and that some were experimenting with tailored applications, including agents. The OECD explicitly says its sample of more than 2,000 SMEs across 12 countries is non-representative. That makes it evidence of behavior among engaged SMEs, not a population estimate.

Taken together, the observed baseline supports a low-single-digit to low-teens scenario range for production agents among all SMBs in 2026. This range is an assumption, not a reported statistic. The lower end reflects strict production criteria; the upper end allows for advanced digital sectors and broad vendor definitions. Confidence in the range is low, while confidence that adoption is still materially below general generative AI use is high.

Base, upside, and downside paths

Base case: approximately 5%-10% of SMBs operate at least one qualifying agent in 2026, 10%-18% in 2027, and 15%-25% in 2028. Assumptions include continued embedding of agents in major software suites, gradual improvement in reliability, clearer permission controls, and no severe economy-wide backlash. Adoption remains concentrated in firms with cloud systems, clean data, and a manager who owns process improvement.

Upside case: approximately 8%-15% in 2026, 15%-28% in 2027, and 25%-40% in 2028. This requires agent reliability to improve faster than integration complexity, common protocols to reduce connection work, and vendors to package monitoring and identity well enough for smaller teams. It also assumes visible ROI in support and revenue operations without a wave of damaging security incidents.

Downside case: approximately 3%-6% in 2026, 5%-10% in 2027, and 8%-15% in 2028. This path follows if high-profile failures reduce trust, tool costs remain unpredictable, legal uncertainty delays deployment, or agents require more maintenance than SMBs can support. Confidence is low for all point ranges, medium for the ordering of scenarios, and medium that adoption rises each year.

• Observed data: general AI adoption is much higher than strict production-agent adoption.

• Base 2028 assumption: 15%-25% of SMBs operate at least one bounded agent.

• Upside 2028 assumption: 25%-40% with stronger reliability and packaged controls.

• Downside 2028 assumption: 8%-15% if risk and maintenance dominate.

• Confidence: low on levels, medium on upward direction.

Why capability growth is not the same as adoption

Technical capability provides an upside driver. METR's time-horizon research found an approximately seven-month doubling in the length of software tasks frontier agents could complete at 50% reliability over its measured period. The authors caution that external validity and future trend changes dominate forecast uncertainty. A benchmark task with a clear success condition is not the same as a live business process with ambiguous requests and changing data.

Adoption depends on the last mile. An agent needs authenticated access, least-privilege permissions, reliable source data, deterministic checks, spend limits, monitoring, and a safe handoff. It must also fit the way people actually work. A model can be capable of composing a refund response while the business remains unable to determine which policy applies or who can approve the payment.

The forecast rises more slowly than capability headlines because each company must expose systems and accept operational risk.

Likely leading workflows

Customer support is likely to lead because tickets provide a queue, knowledge can be retrieved, actions can be bounded, and outcomes such as resolution, reopen, escalation, and satisfaction can be measured. Early agents will handle narrow intents, collect missing information, update records, and escalate. High-risk complaints, vulnerable customers, refunds above a threshold, and unclear policies should remain with people.

Sales operations is another leading category. Agents can research accounts, enrich records, summarize interactions, prepare follow-up, route inbound leads, and flag stale opportunities. Autonomous outbound messaging is riskier because errors are customer-facing and scale quickly. Strong systems will use suppression lists, frequency limits, approved claims, identity disclosure where required, and human review for sensitive or high-value communication.

Finance administration and internal IT may follow. Consequential decisions will move more slowly because errors require stronger evidence and oversight.

Security and identity set the pace

NIST's 2026 work found broad agreement that agent security needs adapted controls. A system that acts needs a clear identity, limited authority, and an audit trail.

Treat each agent like a constrained service account. Give it only the systems and actions needed for the workflow. Separate read, draft, approve, and execute permissions. Require step-up approval for money movement, account changes, external publication, deletion, or access to sensitive records. Rotate credentials, log tool calls, and test whether instructions hidden in documents or messages can redirect the agent.

Security maturity is therefore a forecast trigger. The base case assumes mainstream platforms make identity, approval, logging, and policy controls accessible to small teams. If those controls remain expensive or inconsistent, adoption stays near the downside path. If common standards and secure defaults mature, the upside becomes more plausible.

How SMBs should plan through 2028

Plan around a portfolio of bounded workflows rather than a general digital employee. Classify candidate actions by reversibility and consequence. A draft can be reviewed; a CRM note can be corrected; a payment or rejection may be difficult to undo. Start with read and draft permissions, then add actions only after evidence shows the system is reliable for defined cases.

Use a stage gate for every deployment: baseline, offline test, shadow mode, limited live pilot, and controlled expansion. Track completion rate, accepted-output rate, exception rate, correction time, customer impact, security events, and total cost per completed task. A high task-completion rate is not success if humans spend more time finding subtle errors.

Revisit the forecast twice a year. Move toward the upside after robust evaluations and net-positive pilots; move down when failures, review, or legal exposure dominate.

Sources and methodology

This article synthesizes the primary sources below as of the publication date. Forecasts and recommendations are directional scenarios, not guarantees; they should be tested against your workflow, data, risk tolerance, and current vendor documentation.

U.S. Census Bureau: The Microstructure of AI Diffusion (accessed 2026-07-30)

OECD: Empowering SMEs in the Age of AI (accessed 2026-07-30)

Stanford HAI: 2026 AI Index Economy Chapter (accessed 2026-07-30)

METR: Measuring AI Ability to Complete Long Tasks (accessed 2026-07-30)

NIST: AI Agent Standards Initiative (accessed 2026-07-30)

NIST: Security Considerations for AI Agents (accessed 2026-07-30)