AI Agents for Financial Compliance: 86% Use AI. Only 38% Formally Test It.
ACA Group and NSCP's 2026 AI Benchmarking Survey finds 86% of financial firms use AI, but only 38% formally test its output and just 51% could quickly document their oversight. Where AI agents help compliance teams close the gap.

Compliance teams at investment advisers, hedge funds, private equity firms and broker-dealers have spent three years answering one question about AI: should we let people use it? That question is settled. The new one is harder: can you prove it works, and can you prove you checked?
The 2026 AI Benchmarking Survey from ACA Group and the National Society of Compliance Professionals puts numbers on the gap. 86% of firms now formally use AI, up 15 points from 2025 and 41 points from 2024. But only 38% have a formal program to test or validate what that AI produces.
Adoption Grew Fast. Assurance Didn't.
The survey was completed in September 2026 by 261 NSCP members and ACA clients. The most common respondent was a firm with 11 to 50 employees and $1 billion to $10 billion in regulatory assets under management, which is to say a small team carrying a large regulatory load.
The headline adoption numbers look mature:
- 89% of AI-adopting firms have a policy governing employee AI use, up from 70% in 2025.
- 66% have an AI governance committee or working group, up from 48%.
- 85% use enterprise or private generative AI tools, versus 18% using free public tools.
- 69% keep AI to internal use cases; 18% also use it in client-facing applications.
- Nearly one in four are experimenting with agentic AI.
Policies and committees are the paperwork layer. The survey's sharper finding is about what sits underneath them.
"Human Review" Is Doing a Lot of Work
Most firms rely on a person checking the AI's output. 60% run human review on an ongoing basis and another 26% do it ad hoc. That sounds reassuring until you read the next line: only 38% have a formal testing program behind it.
ACA's Aaron Pinnick put the risk plainly in the release: without clear evaluation criteria and the right expertise, "human review can become a procedural exercise rather than an effective control."
That is the difference between a reviewer glancing at an AI-drafted marketing piece and approving it, and a reviewer working against a defined test: what the output must contain, what it must never say, which source it has to match, and a record of the result.
The Documentation Gap Is the Exam Risk
Here is the number that should worry a chief compliance officer most. 81% of respondents agree their oversight is calibrated to the risk of the AI activity. Only 51% are confident they could produce documentation of that oversight on short notice.
ACA President Carlo di Florio framed it in exam terms: if an examiner asks how a firm reviewed AI-assisted decisions and outputs, only about half could promptly produce the evidence. And while the SEC has not adopted an AI-specific rule, he noted firms are still subject to existing obligations around fiduciary duty, recordkeeping, vendor oversight and data protection.
Other findings point the same way:
- No single technical control around AI tools is used by more than 36% of firms.
- Only 18% have an incident response plan that covers AI-specific incidents.
- 55% have enhanced due diligence on AI vendors, up from 43%.
- Top concerns remain information security (68%), data privacy (59%) and cybersecurity (55%).
Where Compliance Teams Already See Value
This is not a story about AI failing. 43% of firms use AI for at least one compliance activity, averaging about three tasks each. Marketing and advertising review is the most common: 41% of firms using AI for compliance apply it there, and 90% of those report improved efficiency.
That use case is a good model for everything else, because marketing review already has a rulebook. The AI checks a draft against known requirements, flags problems, and a human makes the call. The structure is what makes it testable.
What AI Agents Change for Compliance Teams
A chatbot answers a question and the conversation disappears. An AI agent works inside a defined workflow, with defined access, and every step it takes can be logged. For a compliance team trying to close the 81%-versus-51% gap, that difference matters more than raw capability.
Practical places an agent fits, each with a built-in paper trail:
- Marketing and advertising pre-review. The agent checks every draft against the firm's checklist, records which rules it tested and what it flagged, and routes it to a named reviewer. The review record exists whether or not anyone remembers to write it.
- Sampling and testing AI outputs. An agent can pull a weekly sample of AI-assisted work, re-check it against source documents, and log pass or fail rates. That is the beginning of the formal testing program 62% of firms don't have yet.
- Exam-ready evidence packs. When an examiner asks how AI outputs were reviewed, an agent can assemble the policy, the review logs, the test results and the vendor due diligence file in one place, instead of a scramble across inboxes.
- Vendor due diligence tracking. With 55% of firms expanding AI vendor reviews, an agent can keep questionnaires, renewal dates and open follow-ups current.
- Incident readiness. An agent can watch for AI tool outages or access anomalies and open a ticket against the firm's response plan, which helps the 82% without an AI-specific plan start building one.
How Workmate Fits
The survey's core gap is control and evidence, not enthusiasm. That's the problem Workmate is built around. Workmate builds custom AI agents around a firm's own workflows and business rules, and admins control exactly which apps and data each agent can reach. Credentials sit in a secure vault that can require human approval before an agent uses them, and each Workmate runs on its own separate computer with private file storage. Workmate is SOC 2 Type II audited.
For a compliance team, that means an agent that drafts the review, a human who signs it, and a record of both.
What This Means for Compliance Teams Going Forward
The ACA and NSCP data show the industry moving through three stages: whether to use AI, how to govern it, and now how to test and prove it. AI testing was the most common priority for compliance leaders over the next 24 months, and 24% are actively building formal testing programs.
The firms that get there first won't be the ones with the longest AI policy. They'll be the ones who can hand an examiner a clean record of what the AI did, what a human checked, and how they knew it was right.
Frequently Asked Questions
Does the SEC have a specific rule on AI use by advisers? According to ACA, the SEC has not adopted a rule specifically governing AI use, but firms remain subject to existing requirements on fiduciary duty, governance, recordkeeping, vendor oversight and data protection.
Isn't human review enough? It's a start. The survey found 86% of firms use some form of human review, but only 38% have formal testing. Review without defined criteria and documentation is hard to defend in an exam.
Who answered the survey? 261 NSCP members and ACA clients in September 2026, including compliance and risk professionals from asset managers, alternative investment advisers, private equity firms, hedge funds and broker-dealers.