Full disclosure: Google signs my paychecks. AI sits at the core of everything the company does. So understand the position this comes from — the enterprise GenAI ROI story ranks, for most organizations, as pure fiction.

Not for every organization. And not forever. But right now, in March 2025, the gap between investment and measurable return staggers. The industry’s collective unwillingness to talk about it honestly? That starts to feel irresponsible.

The Numbers That Should Worry You #

Last June, Goldman Sachs published a report that landed like a grenade in the AI hype cycle. Jim Covello — their head of global equity research — asked a straightforward question: will the projected $1 trillion in AI spending generate appropriate returns? Covello’s answer landed blunt. For most enterprise use cases, AI was solving tasks at roughly six times the cost of doing them manually.

Six times. Not six percent more expensive. Six times.

I read that figure three times. Then I sent it to a few colleagues and got back a lot of “yeah, but…” responses.

Which is kind of the problem.

Around the same time, David Cahn at Sequoia Capital identified what he called the “$600 billion question” — a gap between AI infrastructure investment and the revenue that infrastructure was actually generating. Sequoia runs no alarmist reputation; the firm backed some of the most successful AI companies on the planet. When they flag a discrepancy that large, it deserves attention.

Most people in the industry gave it a few days of LinkedIn discourse and moved on.

And then MIT dropped a study that deserved front-page coverage but somehow got buried. Across 300+ initiatives representing $30-40 billion in investment, 95% of organizations had achieved zero measurable ROI on their GenAI spend. Zero. Not “below expectations.” Not “too early to tell.” Zero.

The Paradox Nobody’s Resolving #

Here’s the part that makes my head hurt. Deloitte’s 2025 survey found that 85% of organizations increased their AI investment in the past twelve months — despite the mounting evidence that returns are elusive. Companies double down. Hard.

Why?

Partly FOMO drives it: the fear that competitors figure it out first and leave everyone else behind. Partly sunk cost explains it; once an organization drops $50 million on AI infrastructure, the political cost of admitting that premature stands enormous. And partly the narrative machine runs powerful enough to drown out doubt. Every vendor, every consulting firm, every conference speaker (and yes, that speaker role landed on me more than once) has incentives to tell the optimistic version.

But there’s something else going on — something genuinely confusing. The top performers report returns of 3.7x to 10.3x on their AI investments. No small companies running toy projects populate that group; large enterprises with mature data practices and clear use cases get those returns.

So the technology works. Works, demonstrably. For about 5% of organizations.

That 5% figure should change how we talk about GenAI entirely. No technology gap explains the failure. Organizational readiness masquerades as a technology problem — and selling a board on fixing culture costs far more than buying a platform.

Why Most Enterprises Are Failing #

Watching enough enterprise AI rollouts — from inside Google and from outside as a CTO at DreamFlare AI — makes pattern recognition unavoidable. The organizations getting zero ROI tend to share a few characteristics. None qualify as secrets. All of them are ignored anyway.

First, they bought the platform before defining the problem. Meetings at DreamFlare surfaced potential customers who asked “what do we do with AI?” instead of “here’s a specific bottleneck — does AI help here?” The first question leads to expensive experimentation with no success criteria. The second leads to measurable outcomes. The first question also makes for a more enjoyable meeting — probably why executives reach for it so often.

Second, they underestimated the data work. GenAI products look magical in demos because the demo data is clean. Enterprise data never arrives clean. It’s fragmented across systems, inconsistently labeled, full of edge cases that nobody documented. The RAG architecture looks elegant on a whiteboard; getting it to work reliably against a messy knowledge base takes months of unglamorous data engineering. Engineers call this “the boring part.” That boring part determines whether an organization succeeds or fails.

Third, they measured activity instead of outcomes. “We deployed Copilot to 10,000 employees” is not a success metric. “Developer cycle time decreased 18% in teams using Copilot for code review” is a success metric. Most organizations track license counts with precision. Almost none track what changed as a result.

The Honest View from Inside an AI Company #

Writing from inside Google creates a strange tension. The technology earns genuine belief. Google’s own products show remarkable things — things that genuinely surprised a skeptic. The AI tools Google builds for developers prove useful in ways that compound over time; real hours saved, real bugs caught, real value generated.

The gap between what works inside Google — massive data infrastructure, world-class ML teams, decades of operational maturity — and what happens when an average enterprise tries to replicate those results on a fraction of the resources and none of the institutional knowledge stays impossible to ignore.

The comparison never held up as fair, yet the industry keeps making it anyway.

The industry needs to stop pretending that buying an AI platform is equivalent to capturing AI value. Buying a platform and capturing AI value differ by an enormous gap of organizational work that nobody funds because unglamorous process work beats no AI headlines. Buying a Ferrari and wondering why races stay unwinnable captures the same logic. The Ferrari never was the problem.

What the 5% Actually Do #

The organizations seeing 3.7x-10.3x returns share a pattern, observed consistently across sectors — annoyingly unsexy every time.

The 5% start with a narrow, well-defined use case where the current process cost already has a number attached. They invest heavily in data quality before touching any AI tools. They define success metrics upfront and hold themselves accountable to them (this part sounds obvious; almost nobody does it). And they treat AI adoption as an organizational change management project, not a technology deployment.

None of that makes for a compelling keynote — which explains why conference agendas never lead with data cleanup sprints. But it’s the difference between burning $40 million and earning $400 million.

One company from a consulting stint (unnamed for NDA reasons) spent four months cleaning and structuring customer support data before feeding any of it to a model. Four months. Rival companies, moving fast, deployed chatbots in week two without doing that groundwork. Six months later, the slow company that did the boring work had cut support costs by 22%. The fast company had decommissioned their chatbot and gone back to the old system.

I’ve told that story probably a dozen times. It lands differently depending on who’s in the room. Engineers nod. Executives look mildly pained.

The Uncomfortable Forecast #

My read: 2025 marks the year the GenAI ROI reckoning gets loud. Not because the technology stops improving — the pace stays honestly remarkable, and nothing slows that down. But because boards and CFOs now ask hard questions that “transformational potential” no longer deflects.

The $600 billion gap Sequoia identified isn’t closing. If anything, it’s widening as infrastructure costs grow faster than revenue generation. The organizations that figure out how to cross that gap will build genuine competitive advantages. Organizations that skip the hard work end up with expensive AI infrastructure and nothing to show for it — plus the awkward board meeting where someone finally asks why.

I’m optimistic about AI’s long-term value. Working at Google and being pessimistic about AI’s future would be incoherent. But optimism without honesty is just salesmanship. The enterprise AI market has had enough salesmanship for one decade.

GenAI’s ability to deliver ROI no longer needs debating. The 5% proved it. The real question: whether the other 95% commit to the boring, expensive, organizational work required to get there. From everything observed lately, most prefer buying another platform license and hoping for the best.

Not a technology strategy.

That’s a prayer.