The labs say AI is working. Independent research says it isn't.

If you read what the AI companies publish about enterprise adoption, you'll conclude that the transformation is well underway and delivering. If you read the independent surveys, you'll conclude that almost nobody is getting value. These are not small discrepancies. They point in opposite directions.

We track both, because the gap between them is more informative than either number alone.

What the labs are reporting

Google Cloud

Google Cloud's ROI of AI study surveyed 3,466 senior leaders across 24 countries. Among their findings: 52% of executives said their organizations had deployed AI agents in production, 74% reported achieving ROI within the first year, and 39% had already deployed more than ten agents. Among executives reporting productivity gains, 39% said productivity had at least doubled.

Their stated hurdles are worth noting, because they're the same ones the independent research identifies: data security and systems integration.

Anthropic

Anthropic's 2026 State of AI Agents report, based on a survey of over 500 technical leaders, found that roughly eight in ten organizations believe AI agents have already delivered measurable ROI. The obstacles named were integration challenges (46%), data quality requirements (42%), and change management (39%).

Separately, Anthropic's Economic Index — which analyzes actual usage rather than self-report — found that around 49% of jobs have seen at least a quarter of their tasks performed using Claude.

OpenAI

OpenAI's State of Enterprise AI report describes enterprise users saving 40–60 minutes per day, with adoption growth of roughly 6× across the median sector over twelve months. The company has reported that enterprise now accounts for more than 40% of its revenue.

Meta

Meta's contribution is different in kind and arguably more telling. In February 2026, according to Bloomberg reporting, Meta became the first major technology company to formally tie employee performance reviews to AI usage — making "AI-driven impact" a core expectation for every employee.

That's not a claim about ROI. It's a company restructuring its incentives around the assumption that AI fluency is now a job requirement.

What independent research reports

Against all of that: McKinsey's State of AI finds around six percent of organizations capturing significant enterprise value despite near-universal adoption. PwC has found a majority of CEOs reporting no measurable ROI. Gartner forecasts that over 40% of agentic AI projects will be cancelled by 2027.

Same technology. Same eighteen months. Opposite conclusions.

Four things that explain the gap

1. Who gets surveyed

Google Cloud surveyed leaders at organizations that had already deployed generative AI. Anthropic surveyed technical leaders. These are populations selected for having gotten somewhere. A survey of companies that adopted successfully will find that adoption succeeds.

Broader surveys catch the companies that bought licenses, ran a pilot, and quietly stopped. Those companies rarely appear in a vendor's research sample, and they're the majority.

2. Who's answering, and what they can see

An executive reporting "we achieved ROI" and a CFO confirming it in the accounts are doing different things. Perceived productivity gains are real experiences — people genuinely feel faster — but time saved only becomes money when the freed capacity gets reallocated to something that generates revenue or when headcount changes.

Most organizations never do the reallocation step, which is why individual productivity gains and enterprise financial return can both be honestly reported and still not connect.

3. What counts as success

"Delivered measurable ROI" and "captured significant enterprise value" are different thresholds by a wide margin. One can be met by a team saving four hours a week. The other requires a change visible in the P&L. Both are legitimate measures; they're just not the same measure, and the headline numbers get compared as though they were.

4. Incentive, stated plainly

The AI labs sell AI. That doesn't make their research dishonest — the Google Cloud and Anthropic studies disclose their methodology and sample sizes, which is more than can be said for a lot of what circulates. But a company's own research will reliably ask questions whose answers it can live with, and you should read it accordingly.

Consulting firms are not exempt from this. McKinsey sells transformation services, and "only six percent are capturing value" is a number that happens to describe a large addressable market. We sell modernization consulting, so treat this article the same way.

The finding we'd actually plan around

Buried in Anthropic's Economic Index is a result more useful than any of the ROI headlines. Analyzing actual usage rather than opinion, they found that users with longer tenure had a meaningfully higher success rate in their interactions — roughly four percentage points, and the effect held after controlling for the tasks being attempted, the models chosen, language, and country.

The interpretation the report offers is learning-by-doing: people get better at extracting value from AI through practice. Experienced users also brought harder, higher-value work and used the tools more collaboratively rather than just delegating.

If that's right, it changes what the adoption question is. Capability isn't something you procure on a Tuesday. It accumulates. Which means the cost of waiting isn't the value you didn't capture this quarter — it's the compounding you didn't start.

Google Cloud's own read on their top performers points the same direction: the organizations achieving the strongest results invested in the capability, not just the technology.

What to take from all of it

• The technology works. Enough independent and vendor data now agrees on this that treating AI as unproven is no longer a defensible position.
• Value capture is genuinely hard, and the failure points are consistent. Integration, data quality, and change management appear at the top of every list — the labs' own research included.
• Perceived productivity is not financial return. The step almost everyone skips is reallocating the freed capacity to something that shows up in revenue.
• Capability compounds. The strongest argument for starting isn't this quarter's ROI. It's that the learning curve is real and it doesn't start until you do.

The honest summary: the optimists are describing what's possible and the pessimists are describing what's typical. Both are accurate. Which one describes you in eighteen months is a function of decisions you make now — and almost none of those decisions are about which model to buy.