The AI "returns" were self-assigned. The net around the machine is what got priced.

€194 million. That is the "business value" Santander says it extracted from AI in nine months, the headline figure of a quarter in which Novartis's chief executive told analysts this was "the first quarter" he had "really seen" AI start to improve the company's profit and loss "on a tangible basis" ^[1]. The business value is a figure the bank assigned itself. No one outside either company has checked it. Days earlier, a software-industry veteran had proposed a new framework for measuring AI's return on investment, and his reason was stated plainly: budgets today run "on adoption charts rather than actual value," and companies "struggle to establish a baseline cost for the work that AI replaces" ^[2]. The people reporting the returns are the same people conceding they cannot build the baseline the return is measured against. The evidence running alongside the quarter points the other way. Bain found 95 percent of marketing organizations have adopted AI tools and 6 percent report significant performance impact; PwC found 56 percent of chief executives saw neither higher revenue nor lower costs from AI over the prior year; Dentsu found 70 percent of senior marketing decision-makers have not seen major cost efficiencies ^[3]. Not every survey agrees: Morgan Stanley found high-adoption firms outperforming on earnings revisions and share price — but that advantage accrues to the companies whose AI spending matters, not to the machine ^[4]. A startup called Revenium was built on the fact that no shared number connecting AI operations to business outcomes existed; its chief executive says companies are "spending now against a return you cannot measure" and calls every deployment without outcome tracking "agent debt" ^[5]. Carlsberg, in the same earnings season, admitted its benefits are "not yet at scale" ^[1]. And the season's largest self-report — Anthropic's claim that AI agents, software that carries out tasks for a user rather than only answering questions, now log more working hours than humans — arrived without independent audit, published ahead of a planned IPO by a company carrying an $80 billion half-year loss ^[6]. The question this quarter does not answer is the one the rest of the record does: when the machine fails, who pays? What has measurably matured since the start of 2025 is not the machine but the market around its failures — assembled item by item, each with its own arrival date, none of it planned together. Start with the humans who fix what the agents cannot. At the end of October, startups building AI agents quietly wired human contractors into their products to handle the phone calls, scheduling, and CAPTCHA tests the software could not complete, and one of them, Duckbill, launched an API — an interface letting other agent developers buy those human corrections as a service ^[7]. One founder, Marissa Mayer, states the philosophy outright.

We're not even trying to hide it, we think the human in the loop is a benefit. — Marissa Mayer

The insurers moved most clearly. In November 2025, three large U.S. insurers — W.R. Berkley, Chubb, and Great American — asked regulators for permission to exclude AI liabilities from their corporate policies ^[8]. The exclusion clause is itself the instrument: it parks the retained risk on the AI buyer rather than the insurer. Mosaic Insurance went further, refusing to insure the large-language-model technology behind chatbots at all.

It’s too much of a black box — Mosaic Insurance

A product line then arrived to fill the gap. In March 2026, HSB launched dedicated AI liability coverage for bodily injury, property damage, copyright, and defamation claims that fall outside standard policies ^[9]. Accountants had already begun selling AI audits and compliance insurance by September 2025 — offerings their own marketers describe as "safety nets rather than preventative tools" ^[10]. Human review, once a cost, has hardened into standing policy: in October, the chief technology officer of Reputation mandated human review of every public-facing AI output, calling unfettered AI "a liability waiting to happen" ^[11]. The European Union wrote the same net into law at the start of 2025, when Article 4 of the AI Act began requiring every deployer to train staff in "the preservation of human contribution," against what the law calls "uncritical acceptance of AI-generated errors" ^[12]. When an AI agent breached Australia's Medicare systems in October, no contract answered whether the rogue agent, the user who typed the prompt, or the developer was responsible ^[13]. The user-responsibility defense is the labs' standing defense. When Google's AI Overviews produced defamatory statements about a Munich man, Google argued the user is responsible for fact-checking AI results ^[14]. In June 2026, the court rejected that defense and held Google directly liable for its AI's output; Google is appealing, calling the errors "specific and narrow." The labs' own leaders have forecast that agents could cause hundreds of billions of dollars of damage within a single year, which is why Columbia and Yale legal scholars this fall proposed ending the industry's limited liability — the shield that keeps a firm's owners from being pursued for its losses — for AI firms facing mass lawsuits ^[15]. Alex Karp of Palantir pushes the line back the other way: the first line of defense is that you are "liable for your own actions" ^[16]. The market's measurable AI money sits somewhere else. The record S&P 500 profit forecasts this fall were driven by information-technology companies — Nvidia, Micron, Super Micro, Teradyne — selling into chip and data-center demand ^[17]. Where user-side gains are actually quantified, they come from operational AI that predates the generative agents: 22V Research found the biggest quantified margin boosts of the year sit in routing, logistics, and back-office automation — Waste Management's truck routing, worth $300 million in EBITDA, and C.H. Robinson's logistics ^[18]. Here is the asymmetry. What got underwritten, priced, litigated, and legislated over the past two years is the cost of AI going wrong. What got announced this quarter is that AI is finally paying off — a number a bank assigned itself, a profit-and-loss improvement no auditor has signed, measured against a baseline the reporters themselves say they cannot build. €194 million of self-assigned "business value" requires no auditor. The net around the machine now bills; the machine itself has still only announced. An audited, baseline-verified return from a user-side generative agent is the thing that would break this reading — so watch next earnings season for one.