Middle Market AI Structure Research Note

Middle Market AI Structure Research Note

Research conclusion

Current evidence does not prove that AI is squeezing the middle of professional services. It supports several mechanisms that could produce that result and several mechanisms that could strengthen mid-sized firms.

The public article should remain a falsifiable positioning analysis, not a market forecast.

Defining the middle

There is no single cross-industry definition. The OECD and Eurostat commonly treat 50 to 249 employees as medium-sized and 250 or more as large. US Small Business Administration standards vary by industry and may use employees or receipts.

The article should therefore separate statistical size from strategic position. A strategically middle firm has more coordination cost than a small specialist but less proprietary data, technology capacity, distribution, or service breadth than a scaled platform. A 40-person niche firm can be strategically specialized. A 500-person regional generalist can be strategically middle.

US Census SUSB data can measure firms, establishments, employment, payroll, and receipts by enterprise size and industry. It does not currently isolate AI-caused changes in market share or margins.

Evidence for scale advantage

The OECD reported in January 2026 that 52 percent of large firms used AI in 2025 compared with 17.4 percent of small firms across countries with available data. Its 2025 SME report identifies data, skills, finance, connectivity, and digital maturity as adoption enablers.

Eurostat's 2025 data show AI use was more common in large EU enterprises than in SMEs. Professional, scientific, and technical services had a 40.43 percent adoption rate, while real estate activities were at 24.82 percent. Lack of expertise, legal uncertainty, and privacy concerns were leading reasons considered adopters did not proceed.

These statistics support an adoption gap. They do not show that large firms capture the resulting profit or that medium firms lose clients.

Evidence against an inevitable squeeze

The OECD also finds that natural-language-generation adoption gaps are smaller than gaps for some capital-intensive AI applications, which it links to easier access. Among AI-using firms, small enterprises have a slightly higher share using AI for marketing or sales. General tools and embedded software can lower the fixed cost of experimentation.

The Henry case-study library documents small and mid-sized commercial real estate firms using a shared product for offering memorandums, research, and marketing materials. The cases are vendor-selected and cannot establish average ROI. They do show that access is not exclusive to the largest firms.

The peer-reviewed Organization Science study of 758 BCG consultants found large gains on tasks inside the tested model's capability frontier and worse accuracy on a task outside it. The mechanism is task-specific, not firm-size-specific.

NBER's Generative AI at Work study found larger productivity gains for novice and lower-skilled customer-support agents than for the most experienced group. That finding comes from one company and workflow, but it demonstrates a plausible compression of experience advantages.

What would prove or reject the thesis

Longitudinal firm-level data is needed. Adoption surveys and anecdotes cannot prove causation.

Sources