Artificial intelligence might be closing the book on traditional auditing, as AI agents in finance are turning manual spot checks into full-population assurance that scales with complexity without sacrificing oversight.
Across enterprise operations, the industry is moving away from task-level scripts and toward outcome-first agents that can compare contracts with billing, reason over exceptions and recommend actions while keeping humans in the loop. The payoff is scale: scrutinizing every invoice item against contractual logic and forecasts — not just a sample, according to Kari Mesick (pictured, right), business lead of treasury services at Deluxe Corp.
“I want to look at every [enterprise resource planning] invoice item number for every single client,” she said. “I want to go compare it all to what my projections had been before some work. That would’ve been impossible. Only with agentic would this have even been … a conceivable thing.”
Mesick and Satish Balasubramanian (left), vice president and head of architecture and shared services at Deluxe, spoke with theCUBE’s Rebecca Knight and Dave Vellante at UiPath Fusion, during an exclusive broadcast on theCUBE, SiliconANGLE Media’s livestreaming studio. They discussed how AI agents in finance move enterprises from script-based dependencies to agent-orchestrated outcomes. (* Disclosure below.)
AI agents in finance extend, not replace, humans in the loop
At Deluxe, Mesick partnered with Balasubramanian, working with UIPath partner qBotica Inc. as an accelerator, to pair a bots-first foundation with agents. In this model, robotic process automation, or RPA, handles fixed, rules-based extraction from long agreements while agents reconcile that data against live billing and contract logic. The combination shifts auditing from sampling to line-level reviews and surfaces discrepancies with suggested next steps. Crucially, the intent is augmentation — not replacement — with humans kept in the loop to make decisions at scale, according to Balasubramanian.
“The agents are not going to replace people,” he said. “It’s actually going to augment people, because Kari’s team was not able to go at scale; looking at all contracts, looking at all our customers for the price increase. That’s the benefit that the agents are giving right now.”
But trust is not easily won, which is why teams start by deploying bots and only then graduate to agents. In Deluxe’s model, bots handle the fixed pulls — and when they fail, they fail consistently and visibly, making root causes easier to fix, according to Mesick. Once that foundation is stable, agents are added that flag discrepancies and recommend next actions to a human approver.
“With the bot, we deploy that, and if it’s wrong, it’s very consistently wrong. So, you can fix things more easily,” she explained. “With the agent, I would think in the future it’s going to create a situation where someone will get a message saying, ‘Hey, we’re seeing this discrepancy, we recommend this action, should I proceed?’”
Governance and go-live for AI agents in finance
Governance sits at the center: before agents touch billing or contracts, the team classifies and protects sensitive fields and tightly scopes what a large language model can see. This includes information identified through a privacy impact assessment, or PIA, and protected confidential/controlled-access data, or PCA. Those categories drive the guardrails, according to Balasubramanian.
“With UiPath, you have the option of low-code/no-code, which helps you to make a prototype pretty soon,” he added. “I think it actually took time for us to ensure that we are not doing any PIA data leakages. Meaning, do the contracts have any PIA data or PCA data that we shouldn’t expose, or can we give it to LLMs? We had to make sure that our legal and compliance teams were aligned with what we were doing.”
Even with those mandates in mind, low-code tooling accelerates these prototypes from POC to deployment considerably. Deluxe’s use case is an example of how fast — while remaining compliant — these AI agents in finance can hit the ground, according to Balasubramanian
“Getting those approvals took some time, the typical governance processes,” he said. “The actual implementation I would say took four to six weeks. We were able to make the agents up and running in four to six weeks.”
Here’s the complete video interview, part of SiliconANGLE’s and theCUBE’s coverage of UiPath Fusion:
(* Disclosure: TheCUBE is a paid media partner for UiPath Fusion. Neither UiPath Inc., the sponsor of theCUBE’s event coverage, nor other sponsors have editorial control over content on theCUBE or SiliconANGLE.)
Photo: SiliconANGLE
AI agents in finance: Beyond RPA to full
