AI adoption – Small firms are increasingly adopting AI, yet most have not integrated it into core operations. Meanwhile, researchers question AI ethics and benchmark value.
A genuine “AI transformation” won’t be measured by dashboards at the largest employers—it will be decided by millions of small businesses trying to fit new tools into day-to-day work.
For much of the past year. the conversation around business AI has been dominated by enterprises. typically defined as companies with more than 500 employees.. The focus is understandable: landing a major customer can translate into a dependable stream of recurring revenue for AI and cloud vendors.. But if the goal is productivity for “everyone,” small and medium-sized businesses deserve far more attention.
In the U.S., the scale is hard to ignore.. The Small Business Administration estimates there are around 36 million small businesses, employing about 46% of private-sector workers.. Federal data further suggests that most of these firms are very small—roughly 88% have fewer than 20 employees.. That matters for AI adoption, because the barriers to implementation often rise sharply as budgets, staffing, and internal expertise shrink.
Existing research has offered a mixed picture.. Studies published in 2024 pointed to low levels of meaningful adoption among small businesses.. Yet survey results collected in 2026 complicate that storyline.. A Goldman Sachs study covering 10. 000 small businesses found that three-quarters are now using AI. with 84% linking their use to productivity and efficiency gains.. However, only 14% reported integrating AI into core operations.. Another survey. conducted by the National Federation of Independent Business (NFIB). found that only about a quarter of small businesses said they use AI tools at all—though the surveys differ in the types of firms they tend to capture. with NFIB often focusing on very small. traditional businesses.
This divergence helps explain why AI’s “arrival” and AI’s “integration” can look like different stories.. Many small firms may be experimenting—using AI for particular tasks—while still not changing how work is planned. executed. and managed across the business.. In practice, adoption can start with a tool, but transformation usually requires process changes, training, and repeatable workflows.
Behind the scenes, software vendors are also tailoring AI to fit smaller organizations.. Intuit, Zapier, HubSpot, Lindy, and Microsoft are among the companies competing in the small-business AI market.. Some have embedded AI copilots and automation directly into products customers already use. such as accounting platforms. CRM systems. office suites. customer support tools. and workflow automation.. Microsoft, for example, integrated Copilot into its productivity suite, while Google is weaving its Gemini model into Google Workspace.
Meanwhile, the big AI labs are moving downmarket with offerings explicitly packaged for smaller teams.. OpenAI offers ChatGPT for Business/Teams. which can support tasks like drafting marketing copy and analyzing spreadsheets. and it also provides “skills”—described as reusable. shareable workflows that bundle instructions. examples. and code.. Anthropic. in a move aimed at small business operations. launched Claude for Small Business. a package of AI workflows. skills. and integrations built for functions commonly handled by smaller firms.
What makes these products notable is the way companies describe their approach to adoption barriers.. Anthropic’s small business go-to-market lead Lina Ochman said research indicates around 32% of SMB employees don’t really know how or when to use AI. particularly beyond basic chatbots.. A separate challenge is more forward-looking: Ochman said 64% want to move past chat and use agents that can help run real workflows.. Even when employees encounter agent-style tools capable of handling more complex tasks. they may still struggle to translate that capability into specific applications within their own business.
That is why Anthropic is betting on a plug-and-play model. offering pre-built workflows meant to reduce the work required to get started.. How easily these workflows can be customized for unique business functions remains to be seen. but the underlying goal is clear: make AI usable without turning small business owners into AI engineers.
The alternative path—building and managing custom AI systems—can be far more demanding. especially when the business itself is already stretched.. A vivid example comes from an Austin-based vegan cheese-maker. Rebel Cheese. which reportedly spent time deep in customization to tackle an expensive operational problem: excess shipping charges costing the company $50. 000 a month.. The company used Anthropic’s Claude to investigate the issue and map out a solution. then relied on Manus. an agentic orchestration tool. to automatically dispute suspected carrier overcharges.
Rebel Cheese’s cofounder. Kirsten Maitland. said the effort took months. involving testing multiple AI agents and working long nights to develop and refine the system.. For many smaller firms, that level of effort would be difficult to replicate.. Yet the case also signals what can become possible when small businesses gain access to the same categories of tools that larger companies can more easily experiment with and support.
Over time. tools from Anthropic and OpenAI are likely to evolve toward more specialized and customized builds that are less burdensome to implement.. For now. though. many small businesses appear to use AI in less sophisticated ways than larger counterparts—often focusing on discrete tasks rather than fully re-architecting operations.
That gap between “using AI” and “integrating AI” is likely to shape competitive dynamics across industries.. If small firms remain stuck in trial modes, productivity gains may stay uneven.. If tool vendors succeed in packaging AI workflows that fit common small-business operations. adoption could shift from sporadic assistance to repeatable business processes—where value compounds.
Even as adoption grows, researchers are also raising questions about what AI tools actually do when they appear to reason.. A new paper published in the journal AI and Ethics. from researchers at Harvard Kennedy School’s Allen Lab. argues that leading AI models often give the appearance of deliberating over moral complexities without performing the kind of nuanced reasoning users might expect.
The study. titled “Crocodile Tears: Can the Ethical-Moral Intelligence of AI Models Be Trusted?” tested four models—Claude. GPT. Llama. and DeepSeek—on ethical dilemmas drawn from moral psychology. including scenarios where both available options carry genuine moral costs.. In 87% of the so-called tragic tradeoff trials. the models converged on the same choices. and the researchers reported that these choices often didn’t follow from what the reasoning would suggest.. The study describes the behavior as “shedding crocodile tears. ” performing moral anguish while carrying out what the researchers characterize as an implicit. opaque value hierarchy.
The researchers argue this could create trust problems for users who turn to AI for guidance on hard decisions.. Lead author Sarah Hubbard said in a statement that if a model appears to grapple with an ethical dilemma while reducing it to a predetermined answer. it may be earning trust under false pretenses.
Separately, another debate in AI development focuses on how well benchmarks measure intelligence.. In AI research. benchmark tests are a common way to evaluate models across different skills. from coding to instruction-following and reasoning.. But the concern is that labs can effectively “game” the tests.. Jerry Tworek. a former OpenAI researcher who helped build the o1 and o3 reasoning models. said that once training runs start after a benchmark is released. it stops being a reliable measure of intelligence—because models may be trained on those exact scenarios.
Tworek said that sample questions and answers tend to surface online. enabling model developers to train with that information to score higher.. He also argued that meaningful benchmarks must be able to generate new questions or scenarios for every test so the model being evaluated has never seen them before.
A recently released benchmark tied to those concerns is ARC-AGI-3 by François Chollet.. The benchmark presents an AI agent with novel gaming environments and challenges it to determine the game’s point and how to win. forcing the agent to draw on past experience while making judgments about how to apply it in situations it hasn’t been trained on.
Taken together. these threads point to a broader theme in AI’s business moment: adoption is accelerating. but the quality of outcomes—and the credibility users attach to those outcomes—depends on how tools are integrated. how trust is earned. and whether evaluations measure genuine capability rather than memorized test patterns.
AI adoption small businesses AI transformation SMB productivity Claude for Small Business AI ethics benchmarks