New research reveals a majority of finance professionals believe flawed AI output has reached clients, with guardrails lagging far behind adoption
More than six in ten financial services professionals believe an AI-generated error has reached a client or internal decision-maker within the past year.
The figure comes from Macabacus, a New York-based Microsoft 365 productivity platform serving finance and professional services teams, which published its 2026 GenAI in Financial Services: Velocity and Verification report on September 10, 2026.
The company drew on a survey of its 75,000 users and analysis of conversations with hundreds of clients and prospects across investment banking, private equity, corporate finance, and advisory firms.
Of those who believe an error has escaped review, 46 per cent characterised it as a "probably, but no one noticed" situation - meaning mistakes may be circulating in presentations and financial models without detection. Only 38 per cent of respondents expressed confidence that no such error had occurred on their watch.
Speed has outpaced safeguards
Adoption of generative AI in financial services has accelerated sharply. According to the Macabacus report, 87 per cent of respondents use AI daily or weekly to produce financial models and client presentations.
However, only 23 per cent of firms have comprehensive guardrails in place - defined as a combination of approved tools, accuracy checks, brand compliance, and structured review workflows. A further 36 per cent of regular AI users work at firms with no guardrails whatsoever.
Errors in client-facing materials - whether in a financial plan, an investment proposal, or a portfolio summary - can carry regulatory implications under CIRO and OSC standards. As Wealth Professional has previously reported, getting AI accuracy right is now a compliance consideration as much as a workflow one, with a wrong figure in a client recommendation potentially triggering a suitability concern.
Paul Ross, chief marketing officer at Macabacus in New York, described the problem in direct terms: "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."
Junior confidence, senior doubt
The report surfaces a meaningful divide in how different levels of experience assess AI output reliability.
Among analysts and associates, 43 per cent said AI has made them more confident in their models and presentations. Among vice presidents, directors, and managing directors, that figure falls to 29 per cent and senior reviewers are six percentage points more likely to say AI has made them less confident in their work.
Junior staff are typically closest to the AI tools generating the output, while experienced planners and practice leaders are positioned to catch errors at the review stage. If the people producing AI-assisted deliverables feel more certain about them than the people checking them, the probability of errors passing through increases materially.
This divergence tracks with broader findings in the Canadian advisory landscape. Research cited by Wealth Professional on generative AI's transformation of Canadian wealth management has highlighted that while enthusiasm for GenAI is high, the approach among institutions remains deliberately measured - a gap the Macabacus report suggests has not yet closed at many firms.
What verification looks like in practice
When asked what controls would give them greater confidence in AI-generated content, respondents clustered around a few practical solutions.
Twenty-seven per cent called for a full audit trail showing precisely what AI had changed; another 27 per cent said mandatory human review at the handoff stage was the priority. A further 25 per cent wanted automated verification built into the tools they already use. Only five per cent said firm-approved tools alone would be sufficient.
Overall, 80 per cent of respondents said they wanted some combination of a technology layer, human review, or in-tool checks, a clear signal that the industry is not looking to eliminate AI from its workflows, but to build the verification infrastructure around it.
According to the report, 85 per cent of respondents currently spend 30 minutes or more checking AI-generated models and content before it reaches a client, overhead that better guardrails could redirect toward client relationships and higher-value planning work.
The full Macabacus report is available at macabacus.com.
