AI Governance Watch: The Compliance Gaps Agentic Systems Are Opening Up
Over the past three weeks, a cluster of vendors has converged on the same uncomfortable question: what happens to compliance controls built for humans once an AI agent starts doing the work instead? Intapp published a run of posts on the topic, and Corlytics, Imprima and Arcesium each added a piece from a different angle. Read together, they sketch out where AI governance is actually contested right now.
The gap: controls built for people, not agents
In a post on the risks agentic AI creates for information barriers, Intapp identifies three specific weaknesses in information barrier controls once agentic AI is introduced:
- System-level enforcement. Access restrictions are typically configured within individual systems, such as practice management, document management or email. An AI agent can query multiple systems within a single operation, which can bypass restrictions that were never configured to operate jointly across systems.
- Audit trail gaps. A denied access request from a human user generates a log entry. Some AI systems process a restricted request without generating a corresponding record. This becomes a material issue the first time a firm needs to demonstrate, after the fact, that a barrier was enforced.
- Activation lag. There is often a delay between conflict detection and full barrier activation across all relevant systems. This delay is more consequential when AI can provision a new matter across multiple platforms within minutes rather than days.
In a related post on conflicts clearance, the same firm makes a related point: the bottleneck was never really the analyst, it was the manual process of assembling corporate trees and outside counsel guideline data. Its argument is that AI should do that assembly work while people keep the actual judgment calls, rather than firms chasing full autonomy with generic models that do not know the domain. That distinction, agents assemble, humans decide, is a reasonable line for a fund to hold a vendor to.
Corlytics took the more sceptical, industry-wide view in a post on AI’s effect on RegTech: AI will make it cheap to produce a plausible compliance answer, but producing an answer and maintaining a defensible compliance position are not the same thing. Its framing, that fluency is not provenance, is a useful test to apply to any vendor claiming an AI compliance win: can the system show its working, not just its output. Imprima’s post (cited from the vendor’s own recap in the news log; the source page itself returned an error when we tried to verify it directly) raises a related and more concrete risk: in due diligence document redaction, one missed name is a potential GDPR breach, so the model’s accuracy on the hard cases matters more than its accuracy on average. Arcesium’s contribution, in a post on private loan operations, sits a layer down, in private credit operations, where it argues that audit trails showing who set a data rule, who reviewed it and who approved an exception are what actually let a manager scale automation without losing control of the numbers.
What's still unresolved
None of this is settled. The vendors above are, unsurprisingly, confident their own products close these gaps; the harder question is whether any of them have been tested against a real audit or a real malpractice claim rather than a product demo. As Intapp notes, a majority of malpractice carriers now ask about AI use during intake, which suggests insurers are ahead of some firms’ internal policies. Regulatory pressure is also uneven by jurisdiction: the EU AI Act and Australia’s incoming privacy reforms get name-checked, but neither is a finished picture yet, and a control built for one regime will not automatically satisfy another.
Questions worth asking a vendor
| Question | Why it matters |
|---|---|
| Does the system log a silent access denial the same way it logs a granted one? | An unlogged denial means the barrier’s enforcement cannot be demonstrated retrospectively. |
| How much lag is there between a conflict being flagged and a screen going live everywhere? | Agentic workflows can outrun manual activation of controls. |
| Can the vendor show a real audit trail, not just a claimed one? | Fluency in an answer is not the same as documented provenance. |
| What is the error rate on the hardest cases, not the average case? | Redaction and conflicts work are judged on the misses, not the average. |
What to watch next
Expect more vendors to publish governance content as a way of differentiating from generic AI tools, so treat volume of posts as a signal of where the pressure is, not proof of who has actually solved it. Worth watching over the next quarter: whether any insurer or regulator publishes concrete guidance that names agentic AI specifically, and whether a firm using one of these tools discloses a near-miss publicly, since that is usually what moves a debate from marketing claims to real requirements.
For a fuller list of vendors building governance and compliance tooling for funds, see Qaike’s vendor directory.
This article is based on publicly available information and reflects Qaike’s own analysis and opinions. It is not intended to provide professional advice, endorsement, or a definitive assessment of any vendor or product. While every effort has been made to ensure accuracy, completeness and correctness cannot be guaranteed.
Created by Qaike – Powered by AI.