AI Guardrails Lag as 62% See Errors Reach Clients
Most teams bought AI tools long before they wrote down who checks the output, and a report published September 10 puts a number on the consequence. Macabacus polled its own user base across finance and professional services. Of those people, 62% think at least one AI generated mistake has already landed in front of a customer over the past year, or in front of a senior colleague making a call on it.
That finding travels well beyond investment banking. If you send a proposal, a forecast, or a board update that an AI tool helped produce, then the same gap between speed and verification is sitting inside your company too, and AI guardrails are the only thing standing between a draft and a client.
What the Survey Actually Measured
The report, titled GenAI for Financial Services: Velocity and Verification, drew on responses from 75,000 platform users, plus what the company heard in several hundred customer and prospect discussions. Macabacus sells review tools, so it has a commercial interest in the answer.
The self reported nature of the data matters as well. Respondents were asked what they believe happened, not what an audit confirmed, so the real figure could sit above or below the headline.
The direction is still hard to dismiss. Four numbers carry the argument.
| Finding | Share of respondents |
|---|---|
| Reach for AI at least weekly on models and client decks | 87% |
| Run the full set of four checks described above | 23% |
| Think a flawed AI output got in front of a customer or a senior colleague | 62% |
| Suspect something slipped into finished work unnoticed | 46% |
The Gap Between 87% and 23%
Read the first two rows together, because that is the story. Nearly nine in ten people use these tools weekly on client facing work, and fewer than one in four sit inside a system built to catch what those tools get wrong.
The report set a fairly high bar for what counts. To qualify, a company needed all four of these running together: a named reviewer, a sanctioned tool list, a factual accuracy pass, and a check that the output matches the brand. Plenty of teams manage one or two, and very few manage the set.
A small team can close that gap faster than a large one. You do not need procurement approval to write down which tools are sanctioned and who signs off before anything leaves the building.
Seniority Splits on Whether AI Helps
The most interesting finding is not about errors at all. It is about who trusts the work.
At analyst and associate level, 43% report feeling steadier about the work they hand over. That figure drops to 29% among vice presidents, directors and managing directors, and the senior group was six points likelier to say the opposite.
Read that split as a warning about your own review queue. Junior staff ship faster and feel better about it, while the people who sign off carry more doubt, which usually means review becomes the bottleneck rather than the safeguard.
Why This Hits Small Companies Harder
A large firm can absorb one bad slide. It has compliance staff, professional indemnity cover, and a client relationship deep enough to survive an apology.
A ten person company usually has none of that. One fabricated figure in a pitch deck can end a deal that took nine months to build, which is why AI agent security and plain document review belong in the same conversation.
The build or buy question is tangled up in this too. Teams that use AI to ship their own internal tooling inherit the review burden along with the speed, and the build vs buy calculation should include the cost of checking the output.
A Review Layer You Can Set Up This Week
Start with a list of approved tools. Write down which AI products may touch client work, and name one person who can add to that list, because shadow tools are where unchecked output enters.
Next, separate generation from verification. Whoever prompted the model should not be the only person who checks it, since the person who wrote the prompt already believes the answer.
Then attack the highest risk output first. Numbers, citations, dates and names fail most often, so verify those against a primary document before anything else, and treat the free NIST AI Risk Management Framework as a starting outline rather than a compliance project. If you are still choosing which tools to standardize on, our guide to AI tools for small business covers where the current options actually help.
What to Watch Over the Next Two Quarters
Expect verification to become a selling point. Vendors are already pivoting from what their AI can produce toward what their AI can check, and buyers will start asking for that in procurement questionnaires.
Client contracts are the second thing to watch. Once one large customer asks for a written AI review policy, the request tends to spread through a sector within a year.
Get ahead of it. A one page policy you can email on request costs an afternoon and turns an awkward question into a credibility moment.
Common Questions About AI Review Policies
What counts as an AI guardrail?
The report counted a company as covered only when four things ran together: a named reviewer, a sanctioned tool list, a factual accuracy pass, and a brand consistency check.
How common are AI errors in client work?
Common enough that 62% of the finance and professional services people surveyed think one slipped through to a customer or a senior colleague inside twelve months.
Does a small team need a written AI policy?
Yes, and it can be short. A single page naming approved tools, a review owner, and the checks required before sending is enough to start.
Who should review AI output?
Someone other than the person who generated it. Independent review catches errors that the original author is primed to read past.
