#333 Adi Kuruganti: Why Your AI Pilot Is Failing and What It Takes to Reach Production

What It Covers

Adi Kuruganti, Chief AI and Developer Ops at Automation Anywhere, explains why most enterprise agentic AI pilots fail to reach production, how combining deterministic automation with agentic AI drives mission-critical outcomes, and what a realistic three-to-five-year path toward autonomous enterprise operations looks like.

Key Questions Answered

• •Pilot-to-Production Gap: Most enterprise agentic AI deployments stall because teams treat deployment as a technology problem rather than an outcomes problem. Identify two or three specific business outcomes first — operational productivity, regulatory compliance, or revenue impact — then build the use case around those targets. Having a line-of-business owner at the table alongside IT is a make-or-break factor.
• •80/20 Deterministic-to-Agentic Ratio: For mission-critical processes today, Automation Anywhere observes an 80% deterministic, 20% agentic split across its 1,500 live production deployments. Probabilistic agents chained together compound accuracy loss, so reserve agentic steps for unstructured content, exception handling, and complex decisioning — not routine, well-defined tasks that APIs or RPA bots handle reliably.
• •Human-in-the-Loop Is Non-Negotiable for Now: No current production deployment at Automation Anywhere runs agents autonomously end-to-end on mission-critical processes. The practical model is agents completing due diligence and surfacing recommendations — such as loan APR options based on customer risk profiles — while a human approves before the system executes transactions in financial or healthcare systems.
• •Context Scoping Beats Broad RAG: Feeding all available enterprise data into a single retrieval-augmented generation system actively degrades agent performance. Build targeted context graphs per workflow — an order management agent needs product catalog, order, and shipment data, not company-wide knowledge. Automation Anywhere's Process Reasoning Engine uses process metadata from over 400 million running processes to fine-tune context per use case.
• •Generative Recorder Raises RPA Resiliency by 60%: Traditional RPA bots fail when UI screens change because they mimic fixed user interactions. Automation Anywhere's Generative Recorder combines vision models with DOM structure analysis to detect and adapt to interface changes automatically, delivering a measured 60% improvement in bot resiliency — reducing maintenance overhead without requiring developers to rebuild automations from scratch.