For the past two years, the enterprise AI story has been about adoption. Who’s using it, how many pilots are running, which departments are experimenting. That chapter is closing. The next one is about results, and the numbers are not what anyone expected.
At some point in the last eighteen months, AI spending crossed a threshold. It stopped being an innovation line item that flew under the CFO’s radar and became real money with real expectations attached. The timing varies by company, but the pattern is the same everywhere: once a department’s AI bill passes a certain point, someone in finance asks to see the returns. And for roughly two-thirds of organizations, nobody can produce them.
WRITER’s 2026 AI Adoption in the Enterprise survey, fielded across 2,400 C-suite executives and employees in nearly 30 industries, delivers the headline figure: 59% of organizations now spend more than $1 million annually on AI. But only 29% can point to measurable ROI. More than two-thirds of enterprise AI spending is happening without a clear financial outcome attached to it.
This isn’t a failure of the technology. It’s a failure of measurement. And it’s about to reshape who controls AI budgets.
The data comes from some of the most trusted research organizations tracking enterprise technology. Gartner found fewer than one-third of corporate leaders can connect AI spending to financial results. Forrester predicts a quarter of planned 2026 AI spending will slide into 2027 because proof-of-concept projects can’t show a credible path to value. And KPMG’s tracking shows strategic priority climbing even as measurable gains decline. In other words, companies want AI to work more than ever while getting worse at demonstrating that it does. That’s a problem that won’t solve itself.
The CFO Is Now the Gatekeeper
Forrester’s 2026 predictions forecast that enterprises will defer a quarter of planned AI spending into 2027. Not because the tools don’t work, but because the proof-of-concept phase has run its course and the people writing the checks now want to see returns.
Gartner’s parallel research found that fewer than one in three corporate decision-makers can tie their AI investments to a specific financial outcome. The predictable result: AI purchasing decisions are moving up the chain to CFOs, who aren’t asking whether the technology is impressive. They’re asking what it costs and what it’s producing.
This shift is bigger than it sounds. For two years, AI adoption was driven by department heads and innovation teams with budget flexibility and a mandate to experiment. That era is ending. When the CFO’s office starts treating AI spending the way it treats any other capital allocation, the bar for continuing a project gets a lot higher than “the team seems excited about it.”
KPMG’s quarter-over-quarter data captures the disconnect directly. Strategic priority for AI keeps climbing, from 74% to 79%, even as measured productivity and cost-reduction gains have declined in the same period. Companies want AI more than ever. They’re just getting worse at showing it helps.
This is an uncomfortable trend to square. If AI is getting better and adoption is going up, why are measurable gains going down? The most likely answer isn’t that AI stopped working. It’s that the early wins were in the easiest places — software development, basic customer service — and the harder problems don’t yield to plug-and-play deployment. You can’t just roll out a tool and wait for the numbers to improve. You have to build the measurement system first.
Where the Money Is Going and What’s Actually Working
The spending isn’t random. WRITER’s survey breaks down where the money flows, and three categories dominate.
First, software development and IT operations. This is where ROI is easiest to measure, and it’s where the most consistent returns show up. Code generation, automated testing, infrastructure management — the gains here are concrete and quantifiable.
Second, customer service and support. Chatbots and agent-assist tools have matured past the “frustrating deflection machine” phase. The organizations reporting positive ROI in this category tend to be the ones that measured baseline metrics before deployment, which turns out to be the single biggest predictor of whether a project gets called a success.
Third, sales and marketing. Content generation, lead scoring, personalization engines. This category shows the widest variance in reported returns, partly because marketing attribution was already messy before AI arrived. A company that can’t tell you whether its last billboard campaign worked probably can’t tell you whether its AI-generated email subject lines are moving the needle either. The problem isn’t the AI. It’s the measurement infrastructure, or lack of it.
Fourth, and this one doesn’t get discussed enough, internal knowledge management. Enterprise search, document summarization, meeting transcription and analysis. These tools have quietly become the most widely deployed AI applications in large organizations, partly because the ROI case is straightforward: if you can save every employee an hour a week finding information, the math works at almost any scale. The catch is that most companies don’t actually measure how much time people spend searching for things, so even the easy cases end up unmeasured.
The pattern is clear enough that you can predict whether an AI project will survive the coming budget scrutiny by asking one question: did someone establish a baseline measurement before turning it on? Most didn’t.
There’s a specific reason baselines matter so much, and it’s not just about having numbers to show the CFO. When you deploy AI without a baseline, every claim about its impact becomes unfalsifiable. Did productivity improve after the chatbot went live? Maybe. But without knowing what the baseline was, you can’t tell whether the improvement came from the tool or from the fact that you also hired two people and changed the workflow at the same time. The 29% of companies reporting clear ROI didn’t deploy AI and then look for evidence. They measured first, then deployed, then measured again. It’s a protocol so basic it would be standard in any other domain. Somehow, in AI, it became optional.
The survey methodology itself is worth understanding here. WRITER’s study surveyed 1,200 C-suite executives and 1,200 employees across companies ranging from 100 to more than 10,000 workers, spanning the United States, United Kingdom, Ireland, Benelux, France, and Germany. The field dates — December 2025 through January 2026 — matter because they capture enterprise sentiment right as AI budgets for the current fiscal year were being finalized. These aren’t predictions. They’re snapshots of what decision-makers actually committed to spending.
The Price War Nobody’s Talking About
While enterprise buyers wrestle with ROI measurement, a parallel shift is happening on the supply side. Chinese AI models from DeepSeek, Z.ai, and ByteDance now handle nearly a third of enterprise token volume, at roughly one-tenth the cost of their U.S. competitors.
This price differential is quietly reshaping vendor relationships. Companies that signed enterprise deals with U.S. providers at 2024 prices are now watching colleagues run production workloads on Chinese models for a fraction of the cost. The quality gap between top Chinese and U.S. models has narrowed to the point where, for many enterprise use cases, the cost difference overrides any residual performance gap.
This doesn’t mean U.S. providers are losing. It means the enterprise AI market is becoming a commodity market faster than anyone predicted. When the product is API tokens and the differentiator is price, the economics start looking a lot more like cloud storage than bespoke software.
The Governance Gap Nobody Wants to Talk About
There’s a third dimension to the spending problem that gets less attention than ROI or pricing: governance. As AI spending has scaled, the frameworks for managing it haven’t kept up. The WRITER survey shows that while 59% of enterprises crossed the million-dollar spending threshold, the share with formal AI governance policies is significantly lower.
The practical consequence is that AI purchasing is happening in silos. Marketing buys one set of tools, engineering buys another, customer support buys a third, and nobody at the enterprise level knows the total spend or whether the tools overlap. This is the kind of problem that only becomes visible when someone with budget authority — typically a CFO — starts asking questions that individual departments can’t answer.
Salesforce’s $1 billion commitment to its Agentforce platform in Switzerland shows what the other side of this looks like. That kind of investment happens at the strategic level, with governance baked in from the start. It’s a fundamentally different approach from the department-level experimentation that drove the first wave of enterprise AI adoption. The question for most companies isn’t whether they should centralize AI governance. It’s whether they’ll do it proactively or get forced into it by a budget review they didn’t see coming.
What Happens Next
The numbers point to a shakeout in the next twelve months. Companies that can’t show AI ROI will cut spending, not because they’ve given up on the technology but because they can’t justify the line item to a CFO who’s now paying attention.
The organizations that survive this shift will share a few characteristics. They established baselines before deployment. They measure outcomes in business terms, not model performance metrics. They treat AI as infrastructure — something you budget for continuously rather than fund through one-off innovation budgets. And they’ve moved past the “what can AI do?” phase into the “what should AI do for our specific business?” phase.
The 29% who can prove ROI aren’t necessarily spending more than everyone else. They’re just measuring it better. That’s the uncomfortable lesson in these numbers: two-thirds of enterprise AI spending might be flying blind, not because the technology fails but because nobody set up the instruments to see whether it was working in the first place.
The takeaway for anyone managing an AI budget right now is simple and uncomfortable. If you can’t quantify what your AI tools are producing, assume someone above you is about to ask. The CFO isn’t going to be impressed by adoption rates or employee sentiment surveys. They want a number, and the 29% of companies that can produce one are about to have an much easier time in budget meetings than everyone else. The gap between 59% spending and 29% measuring is 30 percentage points of organizations that are about to have an awkward conversation they didn’t prepare for.