On August 11, 2026, Thomas M. Siebel, the founder, chairman and chief executive of C3.ai — one of the earliest and most prominent public companies built around enterprise AI — sold roughly 453,000 Class A shares for approximately $4.8 million, following the exercise of equity awards, according to an SEC Form 4 filing.1 Insider sales after an options exercise are routine and do not, by themselves, mean an executive doubts his own company. But the timing lands in the middle of the loudest run of enterprise-AI-agent announcements in months, and it is a useful place to start, because it is one of the few numbers in this story that is not in dispute.
A report on trust that itself did not hold up
The headline case for caution is a survey published August 12, 2026 by MIT Technology Review, built on a Google Cloud-commissioned poll of 300 data and technology executives, titled "Scaling AI agents with trustworthy data."2 It reports that within two years, all of the organizations surveyed expect to be running agentic AI, with 69% expecting to use it widely — and that at the weakest-performing companies, so-called "data laggards," AI systems are given access to 30% or less of the company's own data.2 Those are striking numbers, and they fit neatly with the wider industry narrative that adoption is outrunning readiness.
Here is the honest caveat Via News's verification process exists to surface: when this specific source's claims were checked against their underlying facts, only 0% of the 11 claims tested held up.2 That does not necessarily mean the survey's headline statistics are wrong — surveys of executive intentions are hard to falsify — but it means readers should treat this particular report as directional at best, not as an established fact the way a verified financial filing is. We are naming that plainly rather than repeating the numbers as settled truth, because knowing which claims have been checked, and how they held up, is the whole point of this kind of reporting.
What the corporate announcements actually commit to
Set against that shaky research is a run of concrete, named partnerships — though even these come from a source with a measured reliability of 57% on checked claims, so they are treated here as company statements, not independently audited facts.3
Manulife and Microsoft announced on July 22, 2026 a five-year renewal and expansion of their relationship, under which Manulife is adopting Microsoft's Frontier Suite and Agent 365, and expanding Microsoft 365 Copilot to more than 30,000 employees.3 "Our partnership with Microsoft is a critical enabler of Manulife's continued evolution into a truly AI-driven organization," said Shamus Weiland of Manulife. "Adopting the Microsoft Frontier Suite represents the next phase of that transformation, giving us the trusted foundation to advance AI across our global operations with confide[nce]."3 The emphasis on governance and a "trusted foundation" — rather than raw capability — is itself telling: a large regulated insurer is buying oversight tooling alongside the AI itself, which only makes sense if trust, not technology, is the bottleneck.
Box made a similar move a day earlier, on July 21, 2026, unveiling agent guardrails, third-party agent activity oversight, prompt-injection detection and classification-based access policies for enterprise content.4 A customer testimonial from Nomura Research Institute's Tatsutoshi Murata captured the same logic: "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."4 Two large enterprises, in two days, publicly prioritizing control over their AI agents ahead of expanding what those agents can do — that pattern is consistent with a trust gap real enough that vendors are building entire product lines around closing it, whatever the precise survey percentages turn out to be.
Via News's own knowledge graph adds a structural note here: Nvidia is the developer of NeMo Guardrails and A-IQ, tools built specifically for constraining and auditing AI agent behavior, and Mistral AI — the French model developer co-founded by Timothée Lacroix — is a paying Nvidia customer. That is our own connection drawn from the data, not a claim from any single source, but it illustrates how much of the current "self-verifying agent" push runs through a small number of infrastructure providers rather than through each enterprise building its own controls from scratch.
The builders closest to the problem tell a more grounded story
Away from the platform announcements, a set of narrower, more specific bets is being made by founders who lived the underlying problem before they tried to automate it. Emily Man, a partner at venture firm Primary, described how her firm found Casap, a company addressing payment disputes: "They had both experienced the pain points of disputes firsthand at their respective large fintech companies and saw like the amount of internal effort and organizational work that it took to solve those challenges even with a really strong engineering team."5 Man said the firm's interest wasn't abstract: "We were immediately really excited about them because of their backgrounds, and they talked to us about this opportunity that they were thinking about tackling," and that colleagues' feedback was "resoundingly clear that these were two exceptional builders."5
In banking, Covecta's chief revenue officer Ben Thomas frames the pitch in terms of replacing effort, not just software: "The problem that we solve sits across workforce productivity, workflow automation, and portfolio management. We deploy seasoned banker agents that are able to take on the mission-critical tasks, workflows, and portfolio activities that generalize AI [cannot]."6 Covecta currently serves corporate and commercial banks, non-bank lenders, building societies, credit unions and private credit firms in the US and UK, with plans to expand globally, and Thomas puts its addressable market at "tens of thousands of financial institutions globally," targeting labor budgets, not just software spend.6
Maisa AI's CEO David Villalon defines his market even more narrowly — regulated-industry process automation where mistakes are expensive: "I define my market like the market of process automation of core business and production tasks at regulated industries... it's all the core tasks that are today being manual or handled by humans that are part of the core product or the core services of the company."7 And in healthcare, Penguin AI's Glenn Herzberg sizes the opportunity against wasted administrative spend rather than software budgets: "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."8
What unites these four accounts is that none of them describe agentic AI as a general-purpose replacement for judgment. Each describes a narrow, auditable task — a dispute, a loan workflow, a regulated back-office process, a specific category of healthcare paperwork — where the case for automation rests on measurable waste, not on trusting an agent's open-ended decisions.
What to watch
Three threads are worth tracking, based only on what has actually been confirmed so far. First, whether Box's and Manulife's new governance tooling produces measurable adoption gains, or whether it remains a sales talking point — that will only be visible in future disclosures, not in this week's announcements. Second, whether the MIT Technology Review/Google Cloud survey's headline figures get corroborated by a source with a better track record; right now they stand alone with a 0-for-11 verification record, which is a reason for caution, not dismissal.2 Third, whether Thomas Siebel's August 11 stock sale is followed by other insider transactions at C3.ai; a single post-exercise sale tells us little, but a pattern would tell us more about how confident the people closest to enterprise AI actually are in its near-term returns.1
