AI agents in finance demand accountability: governance must track decisions, not just outputs

The nature of AI governance is shifting as enterprises deploy systems capable of performing work autonomously rather than simply supporting human workers. This evolution carries particular weight for organizations like Workiva Inc., which manages reporting, audit and compliance processes where mistakes have serious ramifications. When AI agents operate in finance and other regulated domains, plausibility of output becomes insufficient. Organizations must demonstrate they can follow information trails, document approvals and account for every step an agent executes on their behalf, according to Krista Case, principal analyst at theCUBE Research.

We need to understand things like where did this information and where did these insights come from, who approved a particular action, and can you maybe trace what happened if an AI agent is taking an action on your behalf. When you think about finance and these other regulated processes, it's not really good enough for the action or the response to just look plausible. We have to really make sure that it can be substantiated.

Krista Case, principal analyst at theCUBE Research

Case shared her analysis during an exclusive broadcast on theCUBE, SiliconANGLE Media's livestreaming studio, speaking with host Alison Kosik at Workiva's Amplify event. The discussion centered on how enterprises can enable AI to handle high-stakes work while preserving trust, accountability and human control.

AI governance starts with reliable data

While AI systems can analyze information more rapidly and across broader operational areas, that expanded scope also exposes underlying data problems. Disconnected databases, misaligned terminology and ambiguous data stewardship have existed for years, predating the arrival of generative AI. Autonomous agents intensify these issues because they can convert poor-quality information into actions with unprecedented velocity and scope, Case explained.

I think it's commonly understood that our AI is only as good as the data that it's built on. What we talked about more specifically here at Workiva Amplify was the fact that if we have fragmented data stores, if we have inconsistent definitions and inconsistent ownership over data, these are not necessarily new problems that were created as a result of the enterprise adopting AI. They're problems that existed before.

Krista Case

The subsequent challenge involves establishing boundaries around agent autonomy: determining which decisions require human sign-off and which the system can handle independently. Mandating human approval for every action risks negating the efficiency gains that automation promises. This reality pushes organizations toward risk-based controls, where monitoring and exception-handling protocols become increasingly critical as AI deployments multiply, Case noted.

https://www.youtube.com/embed/8Kafp1RFE4Q?feature=oembed

What's interesting is that going back to the conversation around speed, if a human has to review and approve every action, then really the whole point or much of the value is moot. We need to kind of make decisions around when is AI assisting a human, when can it execute on its own, when does it need approval? I do think that some of these boundaries are still being defined, and I think they're going to evolve and change over time, especially as the business use cases evolve.

Krista Case