MIT researchers count 95 percent of corporate generative AI pilots as failures: no measurable impact on profit and loss. We have built AI systems since 2014, and the failed programs we inherit share one trait, long approval chains. This paper lays out the evidence that mid-market companies hold a structural advantage in getting AI to production, and the operating model that converts that advantage into working systems.
The enterprise playbook produces pilots, not production
In August 2025, Fortune reported the finding that defined the year in enterprise AI. Researchers at MIT's Project NANDA analyzed 300 public AI deployments alongside interviews with 150 leaders and a survey of 350 employees. Their conclusion: 95 percent of corporate generative AI pilots fail to produce measurable profit-and-loss impact. Only about 5 percent achieve rapid revenue acceleration.
The report's lead author, Aditya Challapally, told Fortune what separates the successful few: they "pick one pain point, execute well, and partner smartly." The same research found that AI tools purchased from specialized vendors reached deployment about 67 percent of the time, while internal builds succeeded about 33 percent of the time. The failure mode is not model quality. It is organizational: tools that never meet the workflow, and workflows that never meet a decision.
We have shipped more than 200 production AI and automation deployments since 2014, and the failed programs we inherit share a shape. Scope was set by a committee that does not own the workflow. Success criteria were written after the build, if at all. The pilot was designed to be presentable, not acceptable. The vendor was paid regardless of outcome. Each is a symptom of the same disease: too many people who can say no, and no one who must say yes.
A pilot is a decision deferred. Production is a decision made. Much of the enterprise AI program apparatus is a machine for deferring decisions.
This paper is addressed to the leaders of companies between roughly $50 million and $500 million in revenue. If that is you, the MIT numbers are not a warning. They are an opening. The capabilities behind the failed enterprise pilots are available to you at commodity prices, and the organizational defects that sank them are defects you do not have. What follows is the evidence, then the operating model.
Mid-market appetite has caught up with the Fortune 500; integration has not
RSM's 2025 Middle Market AI Survey, fielded in early 2025 among 966 executives with influence over technology investment, finds that 91 percent of mid-market organizations now use generative AI. Adoption is no longer the differentiator. Depth is. Only 25 percent report generative AI fully integrated across core operations, and another 43 percent report integration across some workflows. The rest use it at the edges (Exhibit 1).
A maturity ladder from RSM's 2025 survey of 966 mid-market technology decision makers: 91 percent use generative AI in some form, 43 percent have it integrated across some operations and workflows, and 25 percent have it fully integrated across core operations. OECD analysis shows the same pattern internationally: among smaller firms using generative AI, only 29 percent apply it to core activities.
The international data show the same shape with a sharper size gradient. OECD analysis published in December 2025 finds that 40 percent of firms with 250 or more employees use AI, against 20.4 percent of firms with 50 to 249 employees and 11.9 percent of firms with 10 to 49. Among smaller firms that do use generative AI, only 29 percent apply it to core activities. In the United States, Federal Reserve economists tracking Census Bureau data report that about 18 percent of firms had adopted AI by year-end 2025, with the adoption rate up 68 percent in the year ending September 2025 and adoption rising steadily with firm size.
Read those gradients carefully. They measure resources, not returns. Large firms adopt more AI because they can fund more attempts and absorb more failures. Nothing in the adoption data says the attempts pay off; the MIT data say most do not. The mid-market's shallower integration is a capacity problem, and capacity can be bought one workflow at a time.
RSM's respondents say as much. Sixty-two percent found generative AI harder to implement than expected. Ninety-two percent hit implementation challenges, with data quality the top concern for 41 percent of those who did. Seventy percent say they need outside help to get full value from AI. In our delivery experience, most of what registers as an implementation problem is a scoping problem: the build started before anyone wrote down what passing looks like.
One more number from the same survey deserves attention, because it cuts against the failure narrative. Eighty-eight percent of mid-market respondents say AI has affected their organization more positively than expected. Hold that beside MIT's 95 percent enterprise failure rate and the picture sharpens. Where the technology actually meets a working process, it overdelivers. Where it meets a program office, it stalls. The mid-market's problem is not disappointment with AI. It is unfinished integration, and integration is a buildable thing.
Four structural advantages let mid-market companies out-ship larger rivals
The mid-market cannot outspend the Fortune 500 on AI. It does not need to. Production speed is governed by structure, and on structure the advantage runs the other way. We see four advantages repeat across our mid-market engagements.
- Concentrated decision rights. In a $50 million to $500 million company, the person who owns a workflow's P&L can usually approve a build on it. One sponsor signs acceptance criteria. In an enterprise, the same decision routes through security, architecture, procurement, legal, and a steering committee, each with veto power and none with delivery accountability.
- Single-tenant estates. Most mid-market firms run one productivity tenant and one ERP. There is one place data lives and one set of permissions to respect. Enterprise estates carry decades of acquisitions: parallel tenants, shadow systems, contested data ownership. Integration effort scales with estate fragmentation, not with headcount.
- Leadership proximity to the workflow. Owner-operators and their executives have usually done the job the AI is meant to assist. They can judge a build against reality in minutes. Enterprise sponsors judge a build against a slide.
- Capital discipline. Mid-market CFOs price technology the way they price equipment: fixed cost, defined output, commissioning test before payment. That discipline is usually framed as conservatism. Applied to AI, it is exactly the contract structure the MIT failure data argue for.
| Stage | Common enterprise path | Common mid-market path |
|---|---|---|
| Scoping | Platform selection and program charter, written by a committee | One named owner picks one measurable workflow |
| Approval | Sequential reviews across security, architecture, procurement, and legal | Sponsor signs executable acceptance criteria at the start |
| Build | Program teams, often multi-quarter, success defined as a demo | Fixed scope in the client's own environment, measured in weeks |
| Acceptance | Steering committee review, criteria negotiated after the fact | Pass or fail against criteria signed on day one |
Speed at mid-market scale is not hypothetical. For WS Audiology we delivered an employee experience portal with intelligent search, serving more than 10,000 employees across more than 30 countries, in eight weeks. The decisive factor was not team size. It was a sponsor with the authority to accept the build against agreed criteria.
Enterprise scale can ship too. We built an AI-driven digital workplace for Rockwell Automation that serves more than 28,000 employees in over 80 countries; content findability improved 60 percent and support tickets fell 40 percent. But the disciplines that made it work, named owners and explicit acceptance gates, are the mid-market's default operating mode. Mid-market buyers get them without the committee tax.
Fixed-scope economics fit how mid-market CFOs already invest
The mid-market is putting real money behind AI. The National Center for the Middle Market's year-end 2025 Middle Market Indicator, a survey of 1,000 C-suite executives at US companies with $10 million to $1 billion in revenue, finds AI is now the leading destination for investment dollars: 28 percent of the incremental investment dollar, ahead of information technology at 18 percent and plant and equipment at 17 percent (Exhibit 2). Fifty-three percent of companies say they will invest in intelligence tools in the near term, up from 44 percent six months earlier. And this is a healthy segment: the same survey reports 11.7 percent year-over-year revenue growth, with 85 percent of companies growing.
Allocation of the marginal investment dollar among 1,000 surveyed mid-market C-suite executives, fielded December 2025: AI 28 percent, information technology 18 percent, plant or equipment 17 percent, facilities 9 percent, additional personnel 9 percent, training and development 8 percent, risk mitigation 5 percent, acquisitions 5 percent. Source: National Center for the Middle Market.
A CFO who would never buy a machine tool without a commissioning test should not buy AI without one. The commissioning test for software is a set of executable acceptance criteria: specific, measurable statements of what the system must do, written and signed before the build starts. The MIT finding that vendor-built tools succeed about twice as often as internal builds is, in our reading, a finding about contracts. A vendor paid on acceptance has one path to revenue: make the criteria pass.
We structured our entire delivery model around this logic. We scope one workflow. The client signs executable acceptance criteria on day one. We build in the client's environment over two weeks. The client pays $10,000 only after every criterion passes. Larger programs run as repeated two-week sprints on the same terms. The structure removes the pilot trap by construction: there is no deliverable except a working system, and there is no payment for anything else.
Run the CFO math on that structure. Downside exposure on a single sprint is the cost of your team's attention for two weeks; the cash payment is contingent on acceptance. Compare that with the standard enterprise alternative: a six-figure discovery phase followed by a platform commitment, with a benefits case that cannot be tested until the program is too large to cancel. Sequenced two-week builds give a finance leader something the program model never can: a kill switch after every sprint, exercised at zero stranded cost.
If a vendor cannot tell you before the build starts exactly what passing looks like, you are funding their learning, not your workflow.
Governance is a scoping decision, not a committee
The honest objection to mid-market speed is governance. Enterprises route AI through committees partly because customers and regulators demand evidence of control. Mid-market firms face the same demands with thinner staff. RSM's survey shows the exposure: data quality is the top implementation concern, and most firms track AI regulation through trade publications and conferences rather than through an internal review function.
The answer is not to clone the enterprise committee. It is to put governance inside the acceptance criteria. We build to SOC 2, HIPAA, and GDPR standards, and every implementation we ship is aligned with the EU AI Act. In a fixed-scope build, those obligations become testable statements: where data may flow, what the model may see, what gets logged, who can override. A control written as an acceptance criterion gets built. A control that lives in a policy binder gets audited later, after the incident.
Compliance work itself is a workflow, and workflows can be shipped. A European pharmaceutical regulator runs an AI compliance scanner we built: more than 620 marketing assets scanned to date, 11 rules applied per scan, about two minutes per asset, down from 2 to 3 hours of manual review. If a regulator can put AI into production on its own compliance workload, a mid-market manufacturer can put AI into production on order intake.
The lesson for mid-market leaders in regulated supply chains is direct. Governance and speed are not opposites. Governance is a workflow with unusually well-documented acceptance criteria, and a fixed-scope build is the cheapest place to satisfy it, because the surface area under review is one workflow rather than an entire platform.
The window is open now, and the adoption data say it is narrowing
Federal Reserve economists report that US firm-level AI adoption grew 68 percent in the year ending September 2025, with growth strongest in the most recent quarters of their data. The National Center for the Middle Market finds 65 percent of mid-market companies very confident their industries will grow in 2026. Both numbers point the same way: competitors are moving, and the segment has the cash flow to fund the move.
The structural advantages in this paper do not expire, but their value is relative. When your enterprise competitor finally grinds a copilot through its committees, your head start in shipped workflows and accumulated process data is the moat. Every two-week build compounds: the second workflow is cheaper than the first because the data boundaries and the acceptance discipline already exist. The expensive part of AI is not the model. It is the first decision. Mid-market companies are built to make that decision faster, and the moves below are how to make it this week.
What to do with this on Monday morning
- Pick one workflow where cycle time or error rate is measurable today. Record the current numbers before you talk to any vendor.
- Draft acceptance criteria as pass-or-fail statements, and put the governance controls in them: data boundaries, logging, human override.
- Name one sponsor with authority to accept or reject the build against those criteria. If the decision needs a committee, the scope is too big.
- Require the build to run in your own environment, on your own tenant, against your own data. Nothing leaves your estate.
- Cap the first engagement at a fixed price and a fixed calendar, with payment contingent on every criterion passing. Refuse open-ended discovery phases.
- Hold AI to the same hurdle rate as plant and equipment. A proposal with no measurable baseline is not an investment; kill it.
- Fortune, MIT report: 95% of generative AI pilots at companies are failing (2025)
- RSM US, Middle Market AI Survey 2025 (2025)
- National Center for the Middle Market, Year-End 2025 Middle Market Indicator (2026)
- Federal Reserve Board, FEDS Notes: Monitoring AI Adoption in the U.S. Economy (2026)
- OECD, AI Adoption by Small and Medium-Sized Enterprises (2025)