The headline number comes with a warning
A report titled "Scaling AI agents with trustworthy data" says that within two years, 100% of respondents plan to be using agentic AI, meaning software that carries out tasks on its own rather than just answering questions. It adds that 69% expect to use it widely.1 Our summary of the same report puts the average share of company data that AI can actually reach at only about 45%. It also says only about half of organizations trust their agents' decisions.1 In organizations the report calls 'data laggards', AI access to company data falls to 30% or less.1
Before you rely on those numbers, note something we measured. Of 11 claims we checked from this publication, none held up against their sources, a 0% rate.1 That is a small sample, and it does not mean the survey is wrong. It does mean these figures are directional evidence, not settled fact, and you should read them that way.
What the verified numbers show: the machinery is being built
The firmest evidence in our records is not a survey. It comes from SEC filings for Nvidia, the chip maker whose products are widely used to run AI. Nvidia's cost of revenue, essentially what it spends to deliver what it sells, was $16.621 billion in fiscal 2024.2 It was $32.639 billion in fiscal 2025 and $62.475 billion in fiscal 2026.2 That is roughly 3.8 times the fiscal 2024 figure in two years.2
Our records hold no revenue figures to set against these, and we have no outside comparison to offer for a number this large. So we will not claim to show profit or demand. What the filings do show is a business whose delivery costs have grown to nearly four times their earlier size. That is consistent with heavy spending on AI infrastructure, though not proof of how it is being used.
The quarterly pattern points the same way. Cost of revenue was $17.394 billion in the first quarter of fiscal 2026 and $20.458 billion in the first quarter of fiscal 2027.2 Cash is more mixed. It was $15.234 billion in the first quarter of fiscal 2026 and $13.237 billion in the first quarter of fiscal 2027.3 Year-end cash rose from $7.28 billion in fiscal 2024 to $10.605 billion in fiscal 2026.3 These are Nvidia's own labels, and a company can hold less cash for many reasons, so we draw no conclusion from the dip.
The infrastructure story extends beyond one company. Our event records show Alibaba's T-Head unit scheduling its Zhenwu V900 AI chip for commercial release in Q3 2027, with a second chip, the J900, also on its calendar.4 These are announced plans, not delivered products.
Big companies are buying trust, not just tools
The announcements by established companies are mostly about control and safety. On July 22, 2026, Manulife and Microsoft announced a renewed five-year agreement. Manulife will adopt Microsoft's Frontier Suite and deploy Microsoft Agent 365, and it will expand Microsoft 365 Copilot to more than 30,000 employees.5 Shamus Weiland said the partnership is "a critical enabler of Manulife's continued evolution into a truly AI-driven organization."5
The Box announcement a day earlier has the same emphasis. Box introduced agent guardrails, oversight of third-party agent activity, prompt injection detection (a defence against hidden instructions that try to hijack an AI agent) and access policies based on agent classification.6 Nomura Research Institute's Tatsutoshi Murata said the firm expects Box "to provide the administrative features needed to safely leverage this new era of AI."6 Both product announcements come from newswire press releases, a source type for which we measured only 57% of 4,954 checked claims holding up.5,6 Treat the quotes as accurate attributions and the product claims as the companies' own.
The topic brief also lists a Siemens–NVIDIA partnership among these ties. Our dossier holds no detail on it, so we do not describe it.
Startups are pricing themselves against payrolls
Latitude, a payments platform founded by an ex-Stripe crypto team member, raised a $35 million Series A on September 10, 2026 for its stablecoin-to-local-currency payments product.7 It is a payments business rather than an agent company, so it shows that funding for enterprise-focused startups is still flowing, not that agents are being adopted.
Several vertical AI companies interviewed by CB Insights describe their markets as labor, not software. Covecta's Ben Thomas says its addressable market is "tens of thousands of financial institutions globally, and for them we are not just disrupting their software budget but their labor budget as well."8 He says Covecta deploys "seasoned banker agents" and currently serves customers across the US and UK.8
Penguin AI's Glenn Herzberg frames healthcare the same way. He says US healthcare administration "runs about a trillion dollars a year, about a quarter of the total health spend, and the published estimates put around $570 billion of that in work that has no effect on health outcomes."9 That is a company's marketing claim relying on published estimates it does not name, so treat it as a pitch, not a measurement.
Maisa AI's David Villalon defines his market as "the market of process automation of core business and production tasks at regulated industries."10 He says that work must be auditable, reproducible and resistant to hallucination, meaning an AI inventing false answers.10
Via News's analysis of these interviews is that the startups are not targeting software budgets. They are going after administrative work done by people. That is where the data bottleneck bites hardest, because an agent that cannot reach or trust a company's records cannot do that work.
What to watch
- The gap between intent and access. If 100% plan to adopt agents but only about 45% of data is reachable, watch whether companies report that figure rising. The source's weak record means we would want a second, independent survey.
- Whether trust follows tooling. Manulife and Box both sell oversight and control. About half of organizations trusting their agents is the number these products are meant to move.
- Nvidia's next SEC filings. Cost of revenue has risen from $16.621 billion to $62.475 billion over two fiscal years.2 The next filings will show whether that pace continues.
- Chip schedules. The Zhenwu V900 is scheduled for Q3 2027.4 A slipped date would be an early sign that the infrastructure build-out is meeting friction.
We cannot tell you from this evidence whether AI agents will pay off for the companies adopting them. What the evidence does show is that spending and partnerships are accelerating while the unglamorous question of data access remains unresolved.
