Is Enterprise AI Actually Working?

Enterprise spending on AI has grown faster than any software category in history. So why are 95% of pilots failing to deliver financial returns?

In 2023, the average large company spent a modest sum experimenting with generative AI. By 2025, according to Menlo Ventures’ annual tracking study[1], collective enterprise spending on generative AI had reached $37 billion – more than tripling in a single year. It is the fastest growth of any software category ever recorded. Andreessen Horowitz[2] found that what enterprise chief information officers (CIOs) spent on AI in all of 2023, they now spend in a single week.

That’s a lot of money being spent quickly. So, boards should ask: what’s the benefit?

The honest answer, based on the best available evidence, is: less than most organisations expected.

The Number That Should Concern Every Board

MIT’s influential GenAI Divide study[3], based on 150 executive interviews, surveys of 350 employees, and analysis of 300 public AI deployments, found that about 95% of enterprise AI pilots are not delivering measurable returns on the profit-and-loss statement.

That’s an interesting number, and we should think about what it means. It does not mean AI doesn’t work. A Google Cloud survey[4] found 74% of enterprises say they are seeing a return on their AI investments, and Gartner’s early-adopter data[5] showed average revenue increases of 15.8% and cost savings of 15.2%. These findings are not necessarily contradictory; they measure different things. The Google and Gartner figures capture perceived value and benefits at the level of individual tasks. The MIT figure measures something harder: whether the investment shows up in the financial results that boards actually review.

This difference is very important. More than 80% of companies have tested tools like ChatGPT or Microsoft Copilot, and almost 40% are using them. These tools really help people do things faster, like writing documents, finding info, and coding. But being more productive and making more money are not the same. Most companies haven’t yet figured out how to make these tools increase profits.

This Is Not a Technology Problem

The most important finding in the current research is that the barriers to getting value from AI are overwhelmingly organisational, not technical. The technology, generally, works. What fails is how organisations choose, deploy, measure, and govern it.

The evidence shows three clear trends.

1. Most AI budgets are pointed in the wrong direction. MIT’s data shows that roughly 70% of enterprise AI spending goes to sales and marketing. Yet the strongest evidence of financial return comes from back-office automation – eliminating outsourced business processes, reducing agency costs, and streamlining operations. Organisations that targeted these areas reported savings of £2 -10 million. The pattern is consistent: the high-profile, customer-facing projects attract the budget, but the unglamorous operational ones deliver the returns.
2. Too many organisations are trying to build what they should buy. Purchasing AI solutions from specialist vendors succeeds roughly 67% of the time. Building internally succeeds about a third as often. In 2024, 47% of AI solutions were built in-house. By 2025, that figure had dropped to 24%, as organisations learned expensive lessons. The exceptions are few and far between. Internal builds make sense only where you have a genuinely unique competitive advantage that no vendor can replicate. For most needs, specialist vendors have deeper expertise, better training data, and faster development cycles.
3. Resistance to new tools remains the single biggest barrier. This was the top-rated adoption challenge in the MIT study[3], ahead of concerns about model quality, data readiness, and cost. It is a finding that should resonate with anyone who has led a significant technology change programme. AI adoption is, at its core, a change management challenge, and most organisations are underinvesting in the people side of the equation.

Your Employees Already Know What Works

One of the most revealing findings is what MIT calls the “shadow AI economy.” Only 40% of companies have an official AI subscription, but workers from over 90% of surveyed companies reported using personal AI tools (ChatGPT, Claude, Gemini, etc.) for work tasks, often on personal devices and without organisational oversight.

This is typically framed as a governance risk, and there are legitimate concerns about data security and compliance. But the more interesting signal is this: employees are voting with their feet. They are finding AI tools that genuinely help them do their jobs, often faster and more effectively than the officially approved corporate tools. The smartest organisations are not trying to stamp this out. They are studying it, understanding which tools people choose, which tasks they use them for, and what that reveals about where AI can add the most practical value.

What the 5% Do Differently

The 5% of enterprises that are delivering measurable returns from AI share several common traits:

• They buy from specialist vendors rather than building internally.
• They target back-office operations first, where the return on investment is clearest and most measurable.
• They empower line managers – the people closest to the work – to identify where AI can help, rather than running adoption from a central AI team.
• They learn from shadow AI usage rather than fighting it.
• And they invest seriously in data quality and governance before attempting to scale.

None of these traits are bizarre or require cutting-edge technical expertise. They are, basically, good management practices applied to a new technology. The gap between the 5% and the rest is not a gap in technical sophistication. It is a gap in organisational discipline.

What This Means for Boards

AI investment should not slow down, but it should be redirected. The technology is proven. The problem is targeting. Organisations that shift AI budgets towards back-office operations and away from speculative customer-facing experiments are seeing returns. This is not a call to stop innovating; it is a call to fund what works while you experiment with what might.

Your AI strategy is really a change management strategy. The 95% failure rate does not reflect a failing technology. It reflects that it’s hard to fit new tools into established processes, incentive structures, and cultures. Boards that treat AI as purely a technology investment are missing the point. The organisations succeeding are those that invest as much in adoption, training, and process redesign as they do in the tools themselves.

If you’re patient and careful, you’ll get good results. Investment experts think companies will truly benefit from AI in 2026. But they’ve said similar things each year since ChatGPT launched. Really, the companies that focused on good data, rules, and testing are now best able to grow their AI use. Those that ignored these basics for fast results are still having problems. There’s no easy way out.

This article draws on findings from my current research on The Real-World State of AI, which examines in detail what separates the 5% of organisations succeeding with AI from the rest, and what the evidence says about where this technology may be heading next.

References

[1] Menlo Ventures, “The State of Generative AI in the Enterprise (December 2025)”, 2025, https://menlovc.com/perspective/2025-the-state-of-generative-ai-in-the-enterprise/

[2] Andreessen Horowitz (a16z), “How 100 Enterprise CIOs Are Building and Buying Gen AI in 2025”, 2025, https://a16z.com/ai-enterprise-2025/

[3] MIT NANDA, “The GenAI Divide: State of AI in Business 2025”, 2025,https://mlq.ai/media/quarterly_decks/v0.1_State_of_AI_in_Business_2025_Report.pdf

[4] Google Cloud, “ROI of AI 2025”, September 2025, https://cloud.google.com/resources/content/roi-of-ai-2025

[5] Gartner, “GenAI Business Value Survey”, July 2024, https://www.gartner.com/en/newsroom/press-releases/2024-07-29-gartner-predicts-30-percent-of-generative-ai-projects-will-be-abandoned-after-proof-of-concept-by-end-of-2025