Recent AI research keeps pointing to the same problem: pilots are easy, operating impact is hard. Here is what appears broken about SMB AI strategies, and the patterns that work.
The noise around AI adoption makes it easy to believe everyone is winning. The research says otherwise. Here is what recent AI research and industry analysis point to, and what the pattern looks like when you stack them side by side.
Number one: most pilots are not turning into operating impact. Fortune's coverage of MIT's Project NANDA report said about 5% of enterprise generative-AI pilots achieved rapid revenue acceleration, while the vast majority stalled or delivered little to no measurable P&L impact. The study was grounded in 150 leader interviews, a survey of 350 employees, and analysis of 300 public AI deployments.
Number two: expected ROI remains the exception, not the rule. IBM's 2025 CEO Study (May 2025) reported that only 25% of AI initiatives delivered expected ROI over the last few years, even as 61% of surveyed CEOs said they are actively adopting AI agents and preparing to scale them.
Number three: abandonment risk is real. WorkOS summarized S&P Global Market Intelligence survey data showing 42% of companies abandoned most of their AI initiatives in 2025, and cited RAND analysis that more than 80% of AI projects fail. The exact failure rate varies by source, but the directional lesson is consistent: adoption alone is not proof of value.
Stack those sources. The story is consistent: many AI deployments do not turn into measurable operating value.
When a study calls a project a failure, it usually means one of four things. The research identifies these patterns consistently.
Pattern one: the pilot that never leaves the pilot. The team runs a demo, the demo impresses, budget gets spent, then the project disappears because nobody scoped how it would actually live in the operation.
Pattern two: adoption without impact. WorkOS's enterprise-AI failure analysis notes that successful programs start with measurable business pain, redesign workflow, and treat deployment as a living product. Employees can use the tools and the company can pay for the tools while the P&L remains unchanged.
Pattern three: missing data foundation. The sources above point to a recurring culprit: organizations launch AI on top of data that is not ready. The models produce output that cannot be trusted because the inputs are inconsistent, incomplete, or wrong. IBM's CEO Study also emphasized integrated enterprise data architecture and proprietary data as central to generative-AI value.
Pattern four: no defined outcome. The project was approved because leadership wanted "to be doing AI," not because a specific problem had been scoped. Without a defined success metric, there is no way to evaluate ROI, no way to iterate, and no way to justify the next round of investment.
Most of the research above focuses on enterprises. SMBs face the same failure patterns with less margin for error, for three structural reasons.
Smaller margin for error. An enterprise can burn $10M on a failed AI initiative and absorb it. A $5M SMB cannot burn $200K on a failed deployment without consequences. The downside of picking wrong is proportionally much larger.
Less in-house expertise. SBA Office of Advocacy research from September 2025 found small firms are closing the AI-adoption gap. Many still lack a dedicated AI owner, which means the person picking the tools, integrating them, and measuring their impact is often someone whose primary job is something else.
Barriers compound. NSBA's 2025 summary of small-business AI adoption research found that many small firms still need simpler, more secure tools, clearer value, trusted partners, and practical training. Any one barrier is manageable. Together, they slow deployment before value is visible.
The patterns that separate successful AI deployments from stalled pilots are not mysterious. The sources point to the same operating disciplines.
Data foundation first. IBM's 2025 CEO Study (May 2025) reported that 68% of surveyed CEOs identify integrated enterprise-wide data architecture as critical for cross-functional collaboration, and 72% view proprietary data as key to unlocking generative-AI value. The companies most likely to get AI ROI are the companies that build the data foundation before the model layer.
Narrow scope, measured outcome. The deployments that work start with one function, one metric, one timebox. A lead-scoring model that either improves conversion by a measurable amount in 90 days or gets killed. A collections agent that either reduces days-sales-outstanding by a measurable amount in one quarter or gets killed. No vague "productivity gains."
Integration into an existing workflow. The tools that deliver value are the ones embedded in a workflow someone already runs. The tools that fail are the ones that require a new workflow nobody was asking for.
Someone accountable for the outcome. Not just the technology. The business outcome. When a named person owns the ROI of an AI deployment, the work has a better chance of producing one.
If you are an SMB leader considering AI, the research tells you something specific. A common outcome is stall or failure. The path away from that outcome requires discipline the enterprise playbooks do not always teach and vendor marketing rarely emphasizes.
That is why strategy work has to come before tool selection, not after. A ninety-day pilot with no defined outcome is not a strategy. A stack of subscriptions chosen because competitors use them is not a strategy. A shiny demo from a vendor who will not own the P&L outcome is not a strategy.
A real AI strategy names the problem, defines the measurable outcome, specifies the data requirements, identifies the accountable owner, and commits to kill the project if it does not deliver. That is what separates operating impact from a pile of pilots.
The research is clear enough to change the sequence. The path is narrow. Do the strategy first.
Five governance failures that quietly kill AI initiatives — and the operating model that prevents them.
6 min read · April 28, 2026
Five responsibilities every AI governance forum should claim — and how to staff them in mid-market companies.
5 min read · April 22, 2026
What 13 weeks at MCRD San Diego taught me about transformation, and why most AI transformations never finish. Four principles from boot camp that every stalled program violates.
3 min read · April 17, 2026
Find out where your organization stands and what to do next.