AI investment is not being rejected. Opaque AI investment is.
This week, the gap became hard to ignore. A survey of 396 enterprise organizations found that unexpected AI costs led one-third to impose emergency spending freezes and one-quarter to delay or cancel an initiative. At the same time, companies able to connect AI to workflow performance and margin improvement are giving investors something much more valuable than another adoption story: evidence.
That does not mean every reported AI gain is independently audited or that every cost surprise proves a failed strategy. It means the market is beginning to separate two very different kinds of AI programs: those built on visible unit economics and those still asking leaders to trust the activity.
For most organizations, the next AI budget decision will not be “Do we believe in AI?” It will be “Can we show what this workflow costs, what business outcome it produces, and whether the result is improving?”
Here’s the week’s data—and what it means for leaders and professionals trying to make AI pay.
The numbers worth knowing this week
74% are in production, but half still cannot consistently measure the return. Plug and Play’s 2026 Enterprise AI Strategy Pulse Survey found that nearly three-quarters of large enterprises surveyed have at least one AI solution in production, while half of production-stage organizations cannot consistently measure ROI. The sample skews toward Fortune 500 and Forbes Global 2000 companies, so this is a view of major enterprises rather than the full business market. The accessible public summary did not disclose a respondent count. (Plug and Play; Forbes)
Unexpected AI costs changed a material business decision at 62% of surveyed organizations. Mavvrik and Benchmarkit surveyed 396 enterprise organizations across several industries in April and May 2026. Forty percent reported board-level escalation after an AI cost surprise, 33% imposed an emergency spending freeze, and 25% delayed or canceled an AI initiative. Only 11% said they could forecast AI spending within plus or minus 10%. This is vendor-published research, but the sample and field dates are disclosed. (Mavvrik; CIO Dive)
Twenty-five S&P 500 companies have quantified an AI-related margin impact. According to 22V Research as reported by Briefs, those companies described an average 180-basis-point margin lift; the figure was 150 basis points when companies combining AI with other efficiency programs were excluded. This is a useful directional signal, not an independently audited causal study. It shows that some companies are beginning to put numbers beside the AI story, while the quality and comparability of those numbers still vary. (Briefs)
C.H. Robinson reports more than 60% productivity improvement in two major divisions since the end of 2022. In its Q2 update, the logistics company also reported a 20% year-over-year increase in adjusted operating income and substantial operating-margin expansion in those divisions. Management attributes the progress to a combination of Lean operating principles, custom-built AI, workflow integration, and disciplined execution—not AI in isolation. That qualification is precisely what makes the case useful: the return came from changing the operating system around the technology. (C.H. Robinson)
The pattern that matters
The divide is not between companies spending on AI and companies cutting AI. It is between organizations that can see the full economics of a completed workflow and those measuring only the easiest pieces.
Token prices can fall while the cost of a completed task rises. Agentic workflows may call several models, move data across systems, use multiple tools, retry failed steps, and require human review. The visible model charge is only one part of the cost. Data platforms, integrations, infrastructure, monitoring, rework, compliance review, and exception handling all belong in the denominator.
If those costs are invisible, a productivity claim can look profitable while the workflow quietly consumes more money than it returns.
That is why an accountable AI business case needs five things before scale:
● A pre-deployment baseline tied to a business outcome ● The full cost of completing the workflow—not merely the cost per token ● A defined human handoff for judgment, risk, and exceptions ● A financial threshold for deciding whether to scale, redesign, or stop ● A regular review cycle that turns real performance data into the next improvement
Willis Towers Watson recently offered a clear example of how a forward-looking case should be framed. Its Propel plan identifies approximately $625 million in expected cash investment, approximately $400 million in targeted run-rate savings, $350 million in targeted net savings after reinvestment, and a 2028 margin objective. Those are forecasts, not realized returns. But the company has made the cost, intended outcome, reinvestment, and timeline visible enough to evaluate later. (WTW)
The hard-accountability phase of AI is not asking leaders to predict the future perfectly. It is asking them to define what success would look like—and build the instrumentation to know whether it happened.
The EMDUIT take — for the professional putting AI to work
Consider a brokerage using AI to follow up with recruiting leads. Access alone produces activity. The AI can draft messages, summarize notes, and reduce response time. Those are useful outputs, but they do not prove the workflow is creating value.
The operating loop becomes measurable when the brokerage defines the outcome before launch: qualified conversations, appointments held, agents recruited, or recruiting cost reduced. It then captures the full workflow cost, establishes approved messaging and compliance guardrails, routes sensitive or high-intent conversations to a human, records the result in the CRM, and reviews which behaviors actually improved conversion.
Now the team can answer a serious question: What did it cost us to create one qualified recruiting conversation, and is that result improving?
This is why EMDUIT focuses on the full operating loop around AI: the task, the behavior, the guardrails, the handoff, the measurement, and the next improvement. The goal is not more AI activity. It is a workflow that becomes easier to trust, safer to scale, and more valuable over time.
The AI budget is not the asset. The measured workflow is.
Sources & methodology
Methodology note: Vendor-published and company-reported results are labeled. Forecasts are not presented as realized returns. Company margin improvements are not attributed to AI alone when management identifies other operating factors. Broad market performance is not treated as proof of AI causation.