For a European business reader, the question about enterprise AI agents is not how fast they are being sold. It is whether they can be audited. Our dossier contains no EU legislation, and we will not pretend it does. It does hold pitches aimed squarely at regulated industries, where an auditable process is a purchasing condition. David Villalon of Maisa AI defines his market as process automation of core business and production tasks at regulated industries, covering back office, operations, finance and similar manual work that must be auditable, reproducible and hallucination resistant.1 That requirement is the trust gap described below.
The UK angle: banking agents with a labour-budget pitch
Covecta, whose Chief Revenue Officer Ben Thomas was interviewed by CB Insights, says it serves corporate and commercial banks, specialist non-bank lenders, building societies, credit unions and private credit organisations. It operates "currently across the US and the UK, with aspirations to move globally thereafter."2 The UK is in Europe but outside the EU, and continental Europe is not in the stated footprint. Thomas describes the market as "tens of thousands of financial institutions globally," and says that "we are not just disrupting their software budget but their labor budget as well."2 He describes the product as "seasoned banker agents" that take on mission-critical tasks, workflows and portfolio activities.2
The labour-budget framing recurs in other interviews. Penguin AI's Glenn Herzberg says his company defines its market as administrative labour spend rather than the healthcare IT software budget. He puts US healthcare administration at about a trillion dollars a year, about a quarter of total health spend. He says published estimates put around $570 billion of that in work that has no effect on health outcomes.3 Those are US figures and company-supplied estimates, not audited data. They show how these vendors size themselves. They are selling against wage bills, not IT budgets. A buyer who replaces people with agents needs to be able to check the agents' work.
The evidence for a gap, and how far to trust it
The clearest statement of the gap in our material comes from MIT Technology Review's "Scaling AI agents with trustworthy data." It reports that within two years, 100% of respondents plan to be using agentic AI, with 69% expecting to use it widely. In organisations it labels 'data laggards', AI access to company data falls to 30% or less.4
Via News tracks how often each source's claims hold up. In our checks, only 0% of 11 claims from this MIT Technology Review piece held up.4 That sample is small, and the result does not prove the survey findings are wrong. It does mean we present them as reported, not as verified. The direction is plausible and matches what the vendors say. The numbers should be read with that caveat.
How incumbents are hedging
Larger, established firms are buying governance alongside capability. Manulife and Microsoft announced a five-year agreement in which Manulife adopts Microsoft's Frontier Suite, deploys Microsoft Agent 365, and expands Microsoft 365 Copilot to more than 30,000 employees.5 Manulife's Shamus Weiland said: "Our partnership with Microsoft is a critical enabler of Manulife's continued evolution into a truly AI-driven organization."5 The headline of the release itself puts "AI governance" first.5
Box's July 2026 announcement points the same way. It added agent guardrails, oversight of third-party agent activity, prompt injection detection and agent classification-based access policies.6 Tatsutoshi Murata of Nomura Research Institute said of the company's plans, "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."6 Both releases come from NewsEOD, where 57% of 4,952 checked claims held up.5,6 Both are company announcements, so treat them as statements of intent.
Nvidia builds the tools for this problem as well as the chips. Our entity graph lists NeMo Guardrails and the NeMo Agent Toolkit as Nvidia products, and lists Mount Sinai Health System and Yum! Brands among its customers.7
What the hardware numbers show, and what they don't
Nvidia's SEC filings give the most solid figures in this dossier. 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, roughly 3.8 times the fiscal 2024 level in two years.8 The quarterly figure shows the same pattern. Q1 2024 was $2.544 billion and Q1 2027 was $20.458 billion, about eight times as much.8 That is a large increase in what it costs Nvidia to deliver its products.
Cash tells a less simple story. Nvidia held $10.605 billion in fiscal 2026, against $8.589 billion in fiscal 2025 and $7.28 billion in fiscal 2024.9 Across the first quarters it went from $5.079 billion in Q1 2024 to $15.234 billion in Q1 2026, then down to $13.237 billion in Q1 2027.9 These numbers show a company scaling fast. They do not show whether the agents running on its hardware are trusted or productive. That is a separate question that cost of revenue cannot answer.
Our dossier also lists two scheduled chip launches from Alibaba's T-Head. The Zhenwu V900 is scheduled for commercial release in Q3 2027 and the J900 in Q3 2028.10 Both dates are in the future and neither chip has been released. Our data contains no evidence about the pace of chip launches, so we draw no conclusion on it.
Capital and attention
Money and coverage continue to arrive. Latitude, a payments platform founded by an ex-Stripe crypto team member, raised a $35 million Series A on 10 September 2026 for stablecoin-to-local-currency payments.10 That is a payments deal, not an AI-agent deal, so it is context and not proof of the trend. CB Insights published a run of enterprise AI interviews on 22 and 24 September 2026, including Veridox's CTO Joey Clover, Binary World's Farrukh Furqan Elahi, Arlo's Uzair Hayat, Shepherd's Justin Levine and LARX's Tad Mielnicki.10 These interviews are the companies describing themselves. The investor side sounds similar. Emily Man of Primary said her firm was "immediately really excited" about Casap's founders "because of their backgrounds."11
Via News's analysis, and what to watch
Via News's analysis of this material suggests a pattern, not a proven conclusion. Vendors are selling agents on the promise of replacing labour in regulated work. Incumbents are buying governance tools at the same time as capability. The one source that puts numbers on the data-access gap has a weak record in our checks. Our data has no measure of user demand, revenue or error rates for these agents, so we cannot say whether adoption is running ahead of trust.
Three things would settle it. The first is whether Covecta's stated ambition to go global reaches continental Europe, where regulators and auditors will test the 'auditable' claim. The second is whether vendors publish error or audit results and not just market sizes. The third is whether the Manulife and Box style governance layers appear in more customer announcements.
