The claim that frames everything, and why we hold it loosely
The most quotable statistic in this story is also the least certain. MIT Technology Review's piece on scaling agents with trustworthy data says that within two years, 100% of respondents plan to be using agentic AI, and 69% expect to use it widely.1 It also says that in organizations categorized as 'data laggards', AI access to company data falls to 30% or less.1 Our topic summary attributes a further figure to Google Cloud research: companies average AI access to only 45% of their data. The excerpt we hold does not show that number, so treat it as unconfirmed.
Our own measurement of this source is poor. Only 0% of 11 checked claims from it held up against their sources.1 That is a small sample, but it is a reason to read the headline numbers as directional, not settled. The story is that adoption plans are running ahead of data readiness. The exact gap is not established.
What is actually shipping: governance before capability
The concrete evidence in the past two months comes from vendors selling control over agents rather than raw model power. On July 21, 2026, Box announced agent guardrails, oversight of third-party agent activity, prompt injection detection and access policies based on agent classification.2 A customer, Tatsutoshi Murata of Nomura Research Institute, said: "As we rapidly advance our utilization of AI agents, we expect Box—which has consistently led the development of security management capabilities for secure collaboration—to provide the administrative features needed to safely leverage this new era of AI."2 Nomura also values Box's multi-vendor support, which lets it switch between AI models.2
A day later, Manulife and Microsoft announced a five-year agreement. Manulife will adopt Microsoft's Frontier Suite, deploy Microsoft Agent 365 and expand Microsoft 365 Copilot to more than 30,000 employees.3 Shamus Weiland said the partnership gives Manulife "the trusted foundation to advance AI across our global operations."3 The trust language in both announcements matters. Buyers are asking for the safeguards before they scale up.
Both announcements come from NewsEOD, a source where 57% of 4,953 checked claims held up.2,3 That is better than the survey source above, but it still means about four in ten claims did not hold. Press-release language such as 'confidence' and 'responsibly' is the vendor's own description.
The startups: selling labor, not software
The CB Insights interviews show a consistent pitch. Covecta's Ben Thomas describes deploying "seasoned banker agents" for workforce productivity, workflow automation and portfolio management, serving banks, non-bank lenders, building societies, credit unions and private credit firms in the US and UK.4 He also says that for these institutions, "we are not just disrupting their software budget but their labor budget as well."4
Penguin AI makes the same move in healthcare. Glenn Herzberg says the company defines its market as administrative labor spend rather than the healthcare IT software budget. He says US healthcare administration runs about a trillion dollars a year, about a quarter of total health spend, with published estimates putting around $570 billion of that in work that has no effect on health outcomes.5 That is a founder-side market-sizing claim, not an audited figure.
Maisa AI's David Villalon defines his market as process automation of core business and production tasks at regulated industries, work that must be auditable, reproducible and hallucination resistant.6 The requirement is the same one that data readiness raises. An agent that cannot reach or trust the underlying data cannot be audited.
Investor Emily Man of Primary describes the origin of Casap, another company in this wave. The founders had "experienced the pain points of disputes firsthand at their respective large fintech companies."7 On the funding side, Latitude, a payments platform built around stablecoin-to-local-currency transfers, raised a $35M Series A on September 10, 2026.8 Its pitch is payments infrastructure, not agents, so it is a weaker data point for this story than the others.
The compute layer: Nvidia's verified numbers
The dossier links Nvidia to much of the agent tooling. Nvidia develops NeMo Agent Toolkit, NeMo Guardrails and NeMo microservices, and Mistral AI is a customer.9 Mistral's founders include Timothée Lacroix.9 Our dossier holds no detail on the Microsoft-Mistral or Siemens-NVIDIA deals named in the topic, so we say nothing further about them.
What we can state plainly is checked against SEC filings. Nvidia's cost of revenue, the direct cost of what it sells, was $16.621 billion in fiscal 2024, $32.639 billion in fiscal 2025 and $62.475 billion in fiscal 2026.10 That is roughly 3.8 times the fiscal 2024 level in two years. Cost of revenue rises with sales, so it is a rough proxy for how much product is moving, not a measure of profit. Cash, by fiscal year, was $7.28 billion, $8.589 billion and $10.605 billion.11 The latest quarter in the series, Q1 2027, shows $13.237 billion, against $15.234 billion in Q1 2026.11 Nvidia's cash therefore does not rise in a straight line.
Earnings per share were $11.93 in fiscal 2024, $2.94 in fiscal 2025 and $4.90 in fiscal 2026.12 That series is not smooth, and the dossier does not explain why. Do not compare the first figure to the later ones without checking how each was reported. We do not know whether the periods are on the same share basis.
If you hold a broad index fund, you may hold Nvidia. The dossier does not give its weighting, so we will not guess at it.
What the evidence does and does not say
Our topic summary also mentions insider selling at C3.ai as a sign of mixed investor conviction. We hold no verified data on that, so we cannot assess it. It should not be read as a finding.
Via News's reading of the material is that the gap between plans and data readiness is the story, and the vendor announcements are early evidence that buyers know it. That is our interpretation, and it rests on a survey source that failed our reliability check. It would be stronger with a second, independent measurement of how much company data agents can actually reach.
What to watch
- Whether a better-sourced study confirms the 30% figure for data laggards and the 45% average, since our current source does not.
- Whether Box, Manulife and Microsoft report results from their agent deployments, not just announcements.
- Whether Nvidia's cash and cost of revenue continue to climb in the coming filings.
- Whether the labor-budget pitch from Covecta and Penguin AI shows up as named customers and revenue, not market-size claims.
Sources:
1 MIT Technology Review, "Scaling AI agents with trustworthy data" (undated) — source reliability: 0% of 11 checked claims held up
2 NewsEOD, "Box Unveils New Controls to Secure AI Agents Operating Across Enterprise Content" (2026-07-21) — source reliability: 57% of 4953 checked claims held up
3 NewsEOD, "Manulife Expands Partnership with Microsoft to Accelerate Enterprise AI Governance and Innovation" (2026-07-22) — source reliability: 57% of 4953 checked claims held up
4 CB Insights, "Executive Interview: Covecta" (undated)
5 CB Insights, "Executive Interview: Penguin AI" (undated)
6 CB Insights, "CEO Interview: David Villalon, Maisa" (undated)
7 CB Insights, "Investor Interview: Primary on Casap" (undated)
8 Via News event record — Latitude raises $35M Series A (September 10, 2026)
9 Via News entity graph — Nvidia relationships (undated)
10 Via News verified data — Nvidia cost_of_revenue, SEC filing (FY 2024, FY 2025, FY 2026)
11 Via News verified data — Nvidia cash, SEC filing (FY 2024–FY 2026; Q1 2026 and Q1 2027)
12 Via News verified data — Nvidia eps, SEC filing (FY 2024, FY 2025, FY 2026)