MIT finds enterprise AI adoption remains shallow despite rising agent use

Deep integration is concentrated in technology

MIT IDE and MIT FutureTech researchers analyzed 10 years of U.S. Securities and Exchange Commission 10-K filings from 510 companies to assess how major businesses were adopting AI. The researchers used the filings because they are federally mandated public disclosures that are held responsible for accuracy and completeness.

By the end of 2025, 11% of S&P 500 firms had AI deeply embedded in their business. Two-thirds of that deep integration was at technology companies. Another 45% of S&P 500 companies were running AI pilots, although the research says many of those pilots may not reach full production.

The divide between industries was especially wide. Half of technology firms had reached deep integration, compared with about 4% of financial services firms. At the same time, 68% of financial services firms were heavily piloting AI.

Returns may depend on reorganizing work

The research describes an AI adoption J-curve, in which companies can initially see lower profit margins before benefits emerge. Technology firms with deep AI integration eventually saw a 3% margin improvement, while non-technology firms using AI to make goods or deliver services saw a 5% improvement. However, productivity gains had not yet appeared in the S&P 500 research.

MIT researchers point to possible explanations including organizational friction, bottlenecks elsewhere in a business and delays before gains become visible. In the working paper Chaining Tasks, Redefining Work: A Theory of AI Automation, John Horton and his co-authors argue that small improvements in AI quality may have limited early effects. Once AI passes a quality threshold, it can handle longer sequences of tasks, potentially forcing a sharp change in how work is organized.

Agent adoption at OpenAI offers one example. According to the working paper The Shift to Agentic AI: Evidence from Codex, Codex rose from under 10% of the average worker’s AI output in August 2025 to 99.8% by June 2026. Legal and recruiting teams moved from near-zero Codex use to nearly 90% of their AI output in about five months. The shift coincided with company-wide training sessions and regular feedback.

Human oversight and trust remain central

MIT’s research presents AI adoption as a question of task design and accountability, rather than simple replacement. A study of four decades of automation across 303 U.S. occupations found different outcomes depending on whether automation took over expert or routine tasks. The researchers say this distinction raises a key question for businesses adopting AI: which work should remain with people?

Another MIT IDE study found that managers reviewing work from a named AI colleague reduced their monitoring intensity by 16% compared with managers reviewing work from an unnamed AI tool. Accountability also shifted from people to the AI, while treating agents as employees did not increase adoption.

Customer trust is lower than professional trust. At Salesforce Connections 2026, Renee Richardson Gosline said 81% of professionals trust agentic AI with customer data, compared with 36% of customers. Her research identifies clearer reasoning, some human review and an exchange of value with customers as ways to help close that gap.

For businesses, the findings suggest that scaling AI will require workflow redesign, training, review processes and deliberate decisions about accountability. The MIT research also says workforce effects will diverge by occupation, industry and employer, rather than producing one uniform outcome.