Why Most Companies Aren't Getting ROI From AI

Ask most founders whether their business is "using AI" and the answer is yes. Ask whether it's paying for itself, and the room goes quiet.

That gap isn't a feeling. It's now one of the most well-documented findings in enterprise research. MIT's NANDA initiative studied 300 public AI deployments, 52 executive interviews, and a survey of 153 leaders in 2025 and landed on a number that's since become shorthand for the whole industry: 95% of generative AI pilots delivered zero measurable financial return (MIT NANDA, "The GenAI Divide," 2025).

McKinsey's most recent State of AI research says the same thing in a different key. Only 37% of organizations attribute any earnings impact to their AI use — flat compared to a year earlier — and just 6% qualify as genuine "high performers" with AI moving 5% or more of EBIT. Meanwhile, spending keeps climbing. McKinsey's own framing: "organizations' conviction in AI is growing faster than the immediate financial returns" they can point to (McKinsey via The Register, August 2026).

So what's actually going wrong? Three things, and none of them are "the AI isn't good enough."

1. Individual productivity is getting confused with business results

MIT's research draws a sharp line here: 90% of workers report using personal AI tools daily for job tasks, but only about 40% of companies have official LLM subscriptions (MIT NANDA, 2025). McKinsey found the same shape of problem on the enterprise side: 80% of AI users report individual productivity gains, but those gains aren't translating into organizational-level performance (McKinsey via The Register, 2026).

MIT has a name for this: the "shadow AI economy" — employees quietly using personal AI tools because they're more reliable than whatever the company officially rolled out. Individuals are getting faster. The business isn't. Those are not the same thing, and a lot of AI budgets are being justified by conflating them.

2. Companies are pointing AI at the wrong layer of the business

This is the finding that should reframe how most founders think about their AI budget. The pilots that failed in MIT's study were concentrated in sales and marketing — customer-facing tools meant to interact directly with prospects and clients. The pilots that worked were almost all back-office: document processing, procurement, risk review. Those delivered real, measurable savings — $2 million to $10 million a year for the companies that did it well (MIT NANDA, 2025).

Front-office AI is the exciting pitch. Back-office AI is where the return actually lives. Most companies built the exciting one first.

3. Tools got deployed. Change didn't

McKinsey's coauthor Michael Chui put this plainly: real ROI requires more than adopting a tool. It requires organizational change alongside it — new workflows, new ownership, new accountability for the process the AI now touches (McKinsey via The Register, 2026). That's consistent with what I see building AI systems for founder-led firms: the tool is rarely the bottleneck. The bottleneck is that nobody rebuilt the workflow around it, so the AI became one more thing bolted onto a broken process instead of a replacement for the broken part.

An AI notetaker doesn't fix a sales process where leads sit unassigned for three days. It just produces a very well-formatted transcript of the delay.

What the 5% are doing differently

MIT's research isn't just a story about failure — it's a map of what's working. The common thread among the initiatives that did produce ROI: they targeted a specific, bounded, back-office workflow with a clear before-and-after metric, rather than a broad "let's use AI in sales" mandate. They measured cost or time saved, not sentiment. And they were built by teams close enough to the actual work to know where the friction was.

That's a smaller, less glamorous ambition than "AI-powered growth." It's also the one with a return attached.

The question worth asking before the next AI purchase

Not "what can AI do here," but "what workflow, specifically, is this replacing, and how will we know if it worked in 90 days." If you can't answer both halves of that question, you're building toward the 95%, not the 5%.

Sources

• MIT NANDA, "The GenAI Divide: State of AI in Business 2025", July 2025
• McKinsey & Company, State of AI research, as reported by The Register, August 2026