The constraint has moved from models to data
Every respondent in a survey summarised by MIT Technology Review says it plans to be using agentic AI within two years, and 69% expect to use it widely.1 The same piece says that in organisations it labels 'data laggards', AI can reach 30% or less of company data.1 Via News's topic summary of the survey adds two figures: AI can reach only about 45% of company data on average, and only about half of respondents trust their agents' decisions. We could not tie those two figures to a checked claim in our records, so treat them as reported, not verified.
A caution on that source. Our measured reliability for the MIT Technology Review item, "Scaling AI agents with trustworthy data", is 0% of 11 checked claims held up.1 We cannot say from this record whether the survey is wrong or our checker handles survey prose badly. What we can say is that the direction of the finding (broad intent, patchy data access) is echoed by the deals below, while the exact percentages deserve scepticism.
Where the money and contracts are going: control, not capability
The clearest signals are about governing agents, not building smarter ones. On July 22, 2026, 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.2 Shamus Weiland said: "Our partnership with Microsoft is a critical enabler of Manulife's continued evolution into a truly AI-driven organization."2 The announcement's own headline frames it around enterprise AI governance.2
Box made the same bet from the content side on July 21, 2026. It announced agent guardrails, oversight of third-party agent activity, prompt injection detection and agent classification-based access policies.3 Tatsutoshi Murata of Nomura Research Institute said the company expects Box "to provide the administrative features needed to safely leverage this new era of AI."3 He also valued Box's multi-vendor support for switching between AI models.3
Both items are company announcements carried by NewsEOD, a source for which 57% of 4,956 checked claims held up.2,3 Read them as what the vendors say they are selling, not as proof of results.
The compute backbone: Nvidia's checked numbers
Agents run on hardware, and Nvidia's SEC-checked figures show how fast that bill is growing. Its cost of revenue, the direct cost of producing what it sells, was $16.621 billion in fiscal 2024, $32.639 billion in fiscal 2025 and $62.475 billion in fiscal 2026.4 That is nearly four times the fiscal 2024 level in two years. The fiscal 2026 figure is also higher than the whole of fiscal 2025's.4 The first quarter of fiscal 2027 came in at $20.458 billion, against $17.394 billion a year earlier, about 18% more.4 Growth is still strong, but the quarterly rate is far slower than the annual doublings.
Nvidia's cash tells a different story. Cash was $7.28 billion in fiscal 2024, $8.589 billion in fiscal 2025 and $10.605 billion in fiscal 2026.5 It was $13.237 billion in the first quarter of fiscal 2027, below the $15.234 billion at the first quarter of fiscal 2026.5 Our dossier has no revenue or margin figures for these periods, so we do not draw conclusions about profitability here.
Our knowledge graph links Nvidia to the agent tooling layer. It lists NeMo Agent Toolkit, NeMo Guardrails, A-IQ, NeMo microservices and the Nemotron 3 Super model as Nvidia products, and HPE's Agentic Trend Analyzer as built on Nvidia.10 It records Mount Sinai Health System, the State of Alaska Legislative Affairs Agency and Yum! Brands as customers, Advanced Micro Devices as a competitor and TD Synnex as a supplier.10 Guardrails sit in that product list beside the compute, which matches the governance theme above. On the roadmap, Alibaba's T-Head has two Zhenwu AI chips, the V900 and J900, scheduled for commercial release in Q3 2027 and Q3 2028.11
What the startups say they are replacing: labour, not software
CB Insights' enterprise AI interviews show how vendors pitch the scale-up. Ben Thomas of Covecta says it serves corporate and commercial banks, non-bank lenders, building societies, credit unions and private credit firms in the US and UK.6 He says its addressable market is "tens of thousands of financial institutions globally" and that "we are not just disrupting their software budget but their labor budget as well."6 Glenn Herzberg of Penguin AI sizes US healthcare administration at about a trillion dollars a year, about a quarter of total health spend, with "around $570 billion" of it in work that has no effect on health outcomes.7 That is his company's framing of published estimates, not our verified data.
David Villalon of Maisa AI defines his market as "process automation of core business and production tasks at regulated industries."8 In the dossier's summary, those tasks must be auditable, reproducible and hallucination resistant.8 That is the trust problem in the survey, restated as a product requirement.
Emily Man of Primary explains why such companies exist. Casap's founders had seen disputes at large fintechs and "the amount of internal effort and organizational work that it took to solve those challenges even with a really strong engineering team."9 Even strong engineering teams struggle with the unglamorous parts of deployment.9
Funding is moving in the same direction. Latitude, a payments platform founded by Cyril Mathew, a former Stripe crypto team member, raised a $35 million Series A on September 10, 2026 for stablecoin-to-local-currency payments.11 CB Insights also published CEO or executive interviews on September 22 and 24, 2026 with Veridox, Binary World, Shepherd, Arlo and LARX.11 The dossier gives only the headline of each, so we do not characterise their content.
Via News's read
Via News's analysis of this material suggests the pilot-to-scale story is a data-access and trust story. The evidence for that is indirect: governance purchases from Manulife and Box,2,3 and vendors pitching auditability.8 The survey percentages that would state it directly come from a source we could not corroborate.1 None of the interviews or press releases above reports measured returns from deployed agents.
What to watch
- Whether Manulife reports results from its Agent 365 and Copilot rollout to more than 30,000 employees.2
- Whether Nvidia's quarterly cost of revenue keeps growing at the recent 18% pace or returns to annual doublings.4
- Whether independent data confirms the 45% data-access and 'about half trust' figures.
- Whether Alibaba's Zhenwu chips ship on the Q3 2027 and Q3 2028 schedule.11