The plainest way to read this week's enterprise AI news is that companies are no longer just testing AI agents. They are signing multi-year contracts for them, and the software that manages them is being sold as a separate product. Manulife, a large insurer, announced a five-year agreement with Microsoft to adopt the Microsoft Frontier Suite and deploy Microsoft Agent 365.1 Manulife also said it will expand Microsoft 365 Copilot to more than 30,000 employees.1 Box announced new controls for AI agents that work with company files, including agent guardrails, oversight of third-party agents and prompt injection detection.2
A question sits underneath this: does the data behind these agents deserve the same trust as the announcements? The evidence here is mixed, and we show where it is thin.
What 'agents' means, and why buyers want them
An AI agent is software that does a task on its own, such as processing a dispute or routing a claim, instead of just answering a question. Three CB Insights interviews show what start-ups are selling.
Ben Thomas, Chief Revenue Officer at Covecta, says the company serves corporate and commercial banks, specialist non-bank lenders, building societies, credit unions and private credit organisations, currently across the US and UK.3 He describes the pitch this way: "We deploy seasoned banker agents that are able to take on the mission-critical tasks, workflows, and portfolio activities that generalize AI."3 The quote is reproduced as it appears in the dossier. He also says Covecta's market is "tens of thousands of financial institutions globally" and that it is disrupting "not just their software budget but their labor budget as well."3
Penguin AI frames its market the same way. Head of Marketing Glenn Herzberg says the company counts administrative labour spend rather than the healthcare software budget. He says US healthcare administration runs about a trillion dollars a year, about a quarter of total health spend, and that published estimates put around $570 billion of it in work that has no effect on health outcomes.4 That is a company's own market sizing, and the dossier does not say which published estimates he means.
Maisa AI's CEO, David Villalon, defines his market as process automation of core business tasks at regulated industries, work that must be auditable, reproducible and hallucination resistant.5
Investor interest is visible too. Emily Man, a Partner at Primary, says her firm met the founders of Casap, who came from Robinhood and Chime, before they had left their companies. She says they "had both experienced the pain points of disputes firsthand at their respective large fintech companies."6 On the funding side, Latitude, a payments platform for stablecoin-to-local-currency transfers, raised a $35M Series A on 10 September 2026.7
Note the shared logic. These firms are selling agents as a replacement for paid staff time, not as a cheaper software licence. That is a bigger claim, and it raises the cost of the agent getting things wrong.
The data problem, and a caution about the source
An MIT Technology Review piece, 'Scaling AI agents with trustworthy data', argues that returns depend on having the right data foundation. It reports that within two years 100% of respondents plan to use agentic AI, with 69% expecting to use it widely.8 It also says that in organisations categorised as 'data laggards', AI's access to company data falls to 30% or less.8 Our topic brief attributes this research to Google Cloud and gives an average of 45% of data accessible. We could not tie that 45% figure to a checked claim in the source material, so treat it as unconfirmed.
The reliability of this source is poor. Via News's own tracking found that 0% of the 11 claims we checked from this article held up.8 That is a small sample, and it does not prove every number is wrong. It does mean the 100%, 69% and 30% figures should be read as what the article says, not as established measurement. The direction of the argument is plausible, and the article's own framing is that agents fail without good data. But it is a claim, and we have not verified it.
The two press releases carry a caution as well. Our measured record for the wire service that carried the Box and Manulife announcements is that 57% of 4,953 checked claims held up.1,2 That is a reasonable but imperfect record. These are companies describing their own deals, and such announcements state intentions and capabilities, not results.
What the companies themselves say about trust
The most useful signal may be how buyers talk about risk. Manulife's Shamus Weiland said: "Our partnership with Microsoft is a critical enabler of Manulife's continued evolution into a truly AI-driven organization."1 The quote is truncated in our source, so we stop where it does. Tatsutoshi Murata of Nomura Research Institute said he expects Box "to provide the administrative features needed to safely leverage this new era of AI."2 The dossier also records that he values Box's multi-vendor support, which lets a company switch between AI models.2
Both buyers are asking for governance: who can see what, and what an agent may touch. Neither statement says how much of the company's data the agents can actually reach. Security controls and data readiness are different problems, and the announcements speak mostly to the first.
The chipmaker underneath: verified numbers
Nearly all of this sits on hardware. Our knowledge base links Nvidia to several agent tools: NeMo Agent Toolkit, NeMo Guardrails, NeMo microservices and A-IQ.9 It records Mistral AI as a customer of Nvidia, and Mount Sinai Health System as a customer too.9 Advanced Micro Devices is listed as a competitor.9
These Nvidia figures come from SEC filings and have been checked. Cost of revenue, the money a company spends to deliver 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 nearly four times the fiscal 2024 level in two years, a calculation from the two filed figures. Cash was $7.28 billion in fiscal 2024, $8.589 billion in fiscal 2025 and $10.605 billion in fiscal 2026.11
Quarterly cash moves less smoothly. It was $15.234 billion in the first quarter of fiscal 2026 and $13.237 billion in the first quarter of fiscal 2027.11 Cost of revenue over the same two quarters rose from $17.394 billion to $20.458 billion.10 Earnings per share were $11.93, $2.94 and $4.90 in fiscal 2024, 2025 and 2026.12 The dossier does not explain why that series falls and then rises, so we do not compare across years. We also have no figure for Nvidia's weight in any index fund, so we cannot tell you whether you already own it through one.
What we could not verify
The brief mentions insider selling at C3.ai as a sign of mixed investor conviction. The dossier contains no C3.ai transaction data, so we cannot confirm, size or interpret it. The Microsoft-Mistral and Siemens-NVIDIA partnerships named in the brief also have no supporting detail in our source material. We do not describe them beyond naming them. The interviews are also undated in our records, although the source pages carry August 2026 timestamps.
What to watch
- Results, not announcements. Manulife's five-year Microsoft deal and Box's controls are commitments. The test is whether any adopter reports how much of its data agents can actually reach.
- Replication of the 30% figure. If a source with a better measured record reports a similar data-access gap, the argument becomes much stronger.
- Whether the labour-budget pitch holds. Covecta and Penguin AI size their markets by staff spend. Watch for customers who publish savings.
- Nvidia's next filing. Rising cost of revenue and uneven cash are the checked numbers to follow as agent workloads grow.
