A machine can draft faster than a junior analyst and polish cleaner than a tired associate at 11 p.m., but the real hazard is older than software: people begin to trust the output before they trust the process. That is how small errors become expensive ones. A new Macabacus report says the danger is already inside client-facing work in financial services, where 62% of respondents believe an AI-generated error reached a client or internal decision-maker in the past 12 months. The lesson is not that AI is failing. The lesson is that human institutions are still built like bridges without enough load testing.
Macabacus released its report, GenAI for Financial Services: Velocity and Verification, on Sept. 10, 2026. The company says the survey draws on its 75,000 users and conversations with hundreds of clients and prospects. That matters because it is not an abstract poll of casual technology users. It is a look at people whose work is meant to survive scrutiny. Yet even in that setting, 46% said an AI-generated error has probably landed in a deliverable without anyone catching it. That is not a minor hygiene issue. It is a sign that the modern office often mistakes acceleration for control.
The more revealing number may be the simplest one: 87% use AI daily or weekly to generate financial models and client presentations, but only 23% have comprehensive guardrails in place. In engineering, that would be like running a pressurized system at full output while checking only a fraction of the valves. In markets, it is the familiar temptation to assume liquidity will always be there until it vanishes. In both cases, the system looks stable until the stress arrives.
Macabacus also found a split in confidence by seniority. Among analysts and associates, 43% say AI has made them more confident in their work. Among VPs, directors, and MDs, only 29% say the same. Leadership is also 6 points more likely to say AI has made them less confident. That gap is more interesting than any single percentage. Juniors tend to experience AI as a force multiplier. Seniors tend to experience it as a liability multiplier, because they are the ones who must stand behind the final document.
This is where investor psychology should pause and look in the mirror. The first-order story is seductive: a tool that saves time, speeds drafting, and reduces repetitive labor. The second-order story is less glamorous: when the output becomes easier to produce, it also becomes easier to ship before it has been fully interrogated. History is crowded with systems that worked beautifully right up to the point where someone assumed the checks were optional. The trap is old. The costume is new.
The Macabacus figures come from a vendor report, so they should be treated as interested-party claims, not gospel. The company is also selling a product that tries to solve the problem it describes. It launched Deck Check AIR, an AI review tool built into PowerPoint, on the same day and cited the 62% figure as its rationale. That does not invalidate the warning. It does, however, mean the report is best read as a signal of anxiety inside the workflow, not as independent proof of a universal crisis.
Still, markets and institutions often reveal their weak points first through vendor behavior. When a company builds a product around review, verification, and workflow control, it is usually responding to a real pain, even if it has every commercial reason to frame the pain dramatically. In other words, a self-interested messenger can still deliver useful information. One should simply inspect the packaging before trusting the medicine.
Paul Ross, Macabacus’s chief marketing officer, put it plainly: “AI is now involved in the majority of models and documents that reach clients, with 87% of firms saying they use AI daily or weekly for this work.” He also said, “Deal teams should not slow down their use of AI. They need guardrails that let them move faster while maintaining accuracy and their clients’ trust. That is what Macabacus is built for.”
That framing is sensible, but the deeper point is harsher. Speed without verification is not efficiency. It is deferred error. In probability terms, a low-probability failure becomes more likely when repeated often enough across enough documents, decks, and decision points. A single mistake may be survivable. A culture that normalizes unverified output is different. It resembles a forest floor that looks healthy because nothing has burned recently. The fuel remains. The spark has simply not arrived.
The true danger in client-facing work is not a dramatic meltdown. It is the accumulation of small, plausibly deniable mistakes. Everyone gains from moving faster in the short run. Everyone loses if the organization’s trust quotient decays in the long run. That is a classic game theory problem: private incentives reward haste, while collective incentives reward caution. If no one wants to be the person slowing the room, then nobody is paid to protect the room.
That is why the 46% figure deserves attention. A report that an error probably landed in a deliverable without being caught is a confession of process weakness, not merely model weakness. People often talk about AI as though the model itself were the main risk. But in financial services, the model is only one layer. The larger vulnerability is the handoff between draft and delivery, where confidence, hierarchy, and deadline pressure decide whether review is real or ceremonial.
The broader lesson reaches beyond one vendor and one survey. Every institution that adopts AI is being asked the same question: do you want a faster pen, or do you want a safer process? The answer, of course, is that you want both. But in practice, organizations usually buy the pen first and postpone the process. That is the old human habit of admiring the visible machine while underfunding the invisible structure that keeps it honest.
Finance is merely the cleanest example because the work is measurable and the stakes are explicit. A wrong presentation, a mislabeled model, or a missed error can harm credibility immediately. Other sectors are more forgiving only in the sense that their failures are slower to surface. The law of institutions is unkind but simple: the longer an error can hide, the more expensive it becomes when discovered.
The paradox is that AI is both a productivity tool and a fragility amplifier. It can reduce friction, but it can also reduce friction in the wrong places. It makes drafting easy. It makes overconfidence easier. It shortens the distance between thought and output, which is useful only if the distance between output and verification does not disappear as well. In the ancient world, fortresses failed not because walls were absent, but because gates were left open on ordinary days.
That is the point investors and executives should absorb. The relevant question is not whether AI is being used. The report says it is, and heavily. The question is whether institutions are building enough friction into the right stage of the process to keep speed from turning into decay. The difference between resilience and fragility is often invisible during calm weather.
Macabacus has put a number on a feeling many desk workers already suspect: the workflow is outrunning the checkpoint. Its own figures suggest that adoption has moved faster than verification, and that the more senior the role, the more caution tends to appear. That should not be read as technophobia. It is closer to survival instinct. Senior people have seen enough systems fail to know that the first failure usually looks like convenience.
The market lesson is plain. When a process becomes easy enough to normalize, it also becomes easy to abuse by accident. That is why the strongest institutions do not merely buy better tools. They design better gates. The companies and teams that last are not the ones that move fastest in every hour. They are the ones that know exactly where speed must stop.
