Agents After Dark

Hari Om Garg on Agent Payment Guardrails

Hari Om Garg builds agentic systems for Fortune 500 companies at Wipro, where an agent might approve any invoice under $500 from a known vendor and route anything larger to a person.

Hari Om GargManaging Consultant, Wiprowith Matt Doughty
YouTubeSpotifyApple Podcasts

In this episode

  • The $499.99 invoice trick is a fraud detection problem solved with conditional access rules
  • A confidence wrapper lifted one agent's accuracy from 72% to around 90%
  • Anything scoring under 83% confidence routes to a human instead of auto-executing
  • Every engagement is scoped against a 15 to 25% first-year ROI target
Full transcriptclick to expand

Transcript coming soon.

From this episode

Frequently asked questions

What stops an agent from splitting a large invoice into smaller approved chunks?

Hari Om Garg treats it as a fraud detection problem rather than a simple rule enforcement issue. He applies conditional access logic borrowed from cloud sign-in systems, so the agent's pattern of behavior, not just a single transaction amount, triggers review.

What confidence threshold decides whether an agent acts or escalates to a human?

In the example Hari describes, a confidence wrapper raised one agent's accuracy from about 72% to around 90%. Anything the agent scores below 83% confidence gets routed to a human instead of being executed automatically.

What ROI does Wipro target for agentic use cases in year one?

Hari's engagements start by mapping a matrix of quick wins and high-impact use cases, with each one forecast to return 15 to 25% within its first year in production.

Do AI agents need their own identity systems?

Yes. Hari argues agents need managed identities the same way cloud platforms already assign managed identities to services, rather than inheriting a human employee's credentials.

What should a CEO under pressure to show AI results do first?

Hari's advice is to cut the noise, spend two weeks finding real use cases grounded in actual business processes, and ship quick wins inside three months rather than chasing a single large rebuild.

Stay ahead of the curve

No spam. Unsubscribe anytime. A resource by Prefactor.

Almost there — check your inbox to confirm your subscription.