The headline number is the least informative one
A survey reported by MIT Technology Review says that within two years 100% of respondents plan to be using agentic AI, and 69% expect to use it widely.1 A result of 100% usually means the question was too easy. The more useful figures are about readiness. Our dossier summarises the same survey as finding that AI can reach only about 45% of company data on average, and that only about half of respondents trust their agents' decisions.1 In organisations the survey classes as 'data laggards', AI access to company data falls to 30% or less.1
A caution before building on these figures. In our own measurement, only 0% of 11 checked claims from that MIT Technology Review document held up.1 The sample is small, so this is not a verdict on the publication. It does mean that the 45% and 'about half' figures should be read as survey claims we have not independently confirmed. Treat them as a direction of travel, not a measurement.
Where the money is going: controls before capability
The deals in our dossier are mostly about control. On July 22, 2026, Manulife and Microsoft announced a five-year renewal and expansion. Manulife will adopt Microsoft's Frontier Suite, deploy Microsoft Agent 365, and expand Microsoft 365 Copilot to more than 30,000 employees.2 The agreement is framed around governance as well as innovation, in its own title.2 Shamus Weiland said: "Our partnership with Microsoft is a critical enabler of Manulife's continued evolution into a truly AI-driven organization."2
Box made the same point from the vendor side on July 21, 2026. It announced agent guardrails, oversight of third-party agent activity, prompt injection detection and access policies based on agent classification.3 Nomura Research Institute's Tatsutoshi Murata said he expects Box to "provide the administrative features needed to safely leverage this new era of AI."3 Both are corporate announcements, a category our measurement scores at 57% of 4,956 checked claims holding up.2,3 Read them as statements of intent, not results.
On September 10, 2026, Latitude, a payments platform founded by an ex-Stripe crypto team member, Cyril Mathew, raised a $35M Series A for stablecoin-to-local-currency payments.4 It is a small round next to the others here, and our dossier does not tie it directly to agentic AI.
The cost side: what Nvidia's filings show
The clearest verified numbers in our dossier come from Nvidia, whose chips and software sit underneath much of this activity. Our graph links its NeMo Agent Toolkit, NeMo Guardrails and NeMo microservices to Nvidia as developer. It also lists Mount Sinai Health System, the State of Alaska Legislative Affairs Agency and Yum! Brands as customers, and Advanced Micro Devices as a competitor.5 Nvidia is therefore a supplier to both the build-out and the guardrails.
Its cost of revenue, meaning what it spends to produce what it sells, was $16.621 billion in fiscal 2024, $32.639 billion in fiscal 2025 and $62.475 billion in fiscal 2026, all from SEC filings.6 That is roughly a doubling each year, and the latest figure is about 3.8 times the first. Cost of revenue is not demand. Our dossier carries no revenue figures, so we cannot say what happened to margins, and we won't guess.
The quarterly series tells the same story. Cost of revenue was $2.544 billion in Q1 2024 and $20.458 billion in Q1 2027.6 Q1 2027 is eight times Q1 2024.
Nvidia's cash balance is a different picture. At year-end it was $7.28 billion (FY 2024), $8.589 billion (FY 2025) and $10.605 billion (FY 2026).7 In the first quarter it went from $5.079 billion (Q1 2024) to $7.587 billion (Q1 2025) and $15.234 billion (Q1 2026), then slipped to $13.237 billion in Q1 2027.7 One caution on how these figures are labelled: the year-end and quarterly series are not reconciled in our dossier. Q1 2026 cash of $15.234 billion exceeds the FY 2026 year-end figure of $10.605 billion, so cash fell over the rest of that year. We report the numbers as filed and do not explain the movement.
Earnings per share is $11.93 for FY 2024, $2.94 for FY 2025 and $4.9 for FY 2026.8 Read the first-to-second drop with care. Per-share figures depend on share count, and our dossier does not say whether these series are adjusted for changes in it. Do not read this as a collapse in profit.
Do you own this? If you hold a broad index fund, Nvidia may be in it, but our dossier contains no index weights and we won't invent one. Check your fund's holdings list.
The new pitch: you are buying labour, not software
The startups in our dossier size their markets by the cost of human work, not by software budgets. Ben Thomas of Covecta 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."9 Penguin AI's Glenn Herzberg makes the same move in healthcare. He says US healthcare administration "runs about a trillion dollars a year, about a quarter of the total health spend," with published estimates putting around $570 billion of it in "work that has no effect on health outcomes."10 Those are the speakers' own claims, not our findings. A trillion dollars is about a quarter of what the US spends on health, by Herzberg's account.10
David Villalon of Maisa AI defines his market as process automation for regulated industries, where tasks must be auditable, reproducible and hallucination resistant.11 That requirement lines up with the survey's trust gap. If only about half of respondents trust agents' decisions, a product that promises auditability is selling to the actual objection.1,11
A market being mapped as it forms
CB Insights published a run of enterprise-AI interviews in late September 2026. They included Shepherd's CEO Justin Levine, Arlo's CEO Uzair Hayat and LARX's CEO Tad Mielnicki on September 24, and Veridox's CTO Joey Clover and Binary World's CEO Farrukh Furqan Elahi on September 22.12 Our narrative also notes CB Insights distributes its datasets through Nasdaq Data Link. We have no source document for that, so it is context, not a verified fact.12 The interviews are founders describing their own companies, so they say little about whether the products work.
Investors say they back founders who have lived the problem. Primary's Emily Man said of Casap's founders that colleagues' feedback was "resoundingly clear that these were two exceptional builders."13 That is a venture partner talking about a portfolio company, which makes it an endorsement and not evidence.
What to watch
- Data access. If the 45% figure moves up in the next survey, the bottleneck is easing. Ask whether the sources behind it are better than the 0% we measured here.1
- Nvidia's cost of revenue against its other filed figures. Costs have roughly doubled each year in the filings.6 Revenue and margin data would show whether pricing is keeping up, and our dossier has neither.
- Governance deals. Manulife's 30,000-seat Copilot expansion and Box's agent controls are the kind of spending that follows a trust gap.2,3 Watch for results, not announcements.
- Quoted market sizes. Labour-budget claims like $570 billion are the speakers' own estimates.10 Look for sourcing before relying on them.
