What stops an agent paying a $20,000 invoice in $499.99 chunks?

Hari Om Garg designs agentic systems for Fortune 500 companies at Wipro, where an agent might approve any invoice under $500 from a known vendor and send anything bigger to a person. Rules like that only hold if the agent can't learn its way around them, and the target is still a 15 to 25% return on investment in the agent's first year in production.

So the sharper question becomes:

How do you give an agent real authority without handing it a loophole?

In this episode of Agents After Dark, Matt Doughty sits down with Hari Om Garg, Managing Consultant at Wipro, to break down how he builds guardrails, confidence thresholds and human review into agentic workflows from the start.

Together they discuss:

• Why the $499.99 trick is a fraud detection problem, and how conditional access rules borrowed from cloud sign-in stop it
• How a confidence wrapper took one agent from 72% to around 90% accuracy, and why anything under 83% goes to a human instead of executing
• Why every engagement starts with a matrix of quick wins and high-impact use cases, each forecast to return 15 to 25% in its first year
• Why agents need their own identities, the way cloud platforms already give services managed identities
• Whether AI ROI is really headcount reduction, and why Hari measures what people do once the agent takes the repetitive work
• Why new use cases should come before rebuilding legacy RPA bots
• What a CEO under pressure should do in the next 30 days: cut the noise, spend two weeks finding real use cases, and ship quick wins inside three months

As agents move from answering questions to approving payments, the rules around them matter as much as the model. Hari treats guardrails as part of the architecture, designed with the use case rather than added after something goes wrong.

This conversation is for CIOs, CTOs and AI leaders who need to show a return on agents this year.

About Hari
Hari Om Garg is an AI solution architect and managing consultant at Wipro, where he designs agentic AI solutions for Fortune 500 companies. He has previously worked at Microsoft, HCL Technologies, LTIMindtree and PowerObjects, and writes for CIO.com on guardrails and intelligent automation for the enterprise.

About Prefactor
Prefactor makes AI agents better.

Prefactor learns how your agents are supposed to behave, evaluates what they actually do on every run, and shows your team how to improve them. Engineers, product managers and the subject matter experts who know what good looks like all work from the same picture, so the agent gets better with every release.

Learn more at prefactor.ai