Agentic AI: The End of Hallucination in Business Operations? Proven Use Cases from a A Salesforce VP
In this episode of Revolutionizing Business Operations with Tony Saldanha, we are joined by Ajit Kumar Kunji, Vice President and Head of Global Business Services at Salesforce, who shares his deep expertise in enterprise technology and business transformation.
Ajit, formerly a Senior Director of Enterprise Business Services at Walmart, brings over 20 years of leadership experience from the world’s biggest BPOs.
Key Topics Discussed:
The Disruptive Power of Agentic AI: Ajit and Tony dive into how Agentic AI is truly disruptive, going beyond traditional RPA or AI, which merely automate existing processes. Agentic AI is focused on totally reimagining work processes and connecting not just at the task level, but across the enterprise end-to-end and the company ecosystem.
The Journey to Agentic AI: Ajit describes his "aha moment" in August 2023 at Salesforce when looking at CRM software capabilities and the ability of Mulesoft to bring data from multiple applications into one custom object. His team began with eight proofs of concept (POCs) focused on real-time data, eliminating "swivel chair motion" (using Excels and macros), and using Generative AI for prediction building (Einstein) and strong workflow (omnichannel capabilities).
Eliminating Hallucination: After noticing a 7-8% hallucination rate with Einstein, Ajit's team shifted to Agentic AI, which is powered by the Atlas reasoning engine. They saw hallucinations "completely went off," and accuracy levels were closer to 100%. Ajit explains this is achieved by defining topics, setting guardrails, and exposing the data through Agent Force's agent interface, which uses 15-16 different LLMs to synthesize queries and refine responses.
Proven Agentic AI Use Cases: Ajit shares three deployed use cases for shared services:
Agentic AI Engagement Layer (Territory Operations): This layer handles request deflection by using three synchronized agents to perform real-time translation, determine the intent of the query (including detecting frustration), and validate data to ensure all required input is present. This filter has reduced the number of use cases that go through the process by 34%.
Quote Special Terms (Order-to-Cash): To accelerate the approval process for non-standard quote edits made by account executives, Salesforce uses Data Cloud to look at historical information and an agent to create a recommendation with strong guardrails and approval reasons. This process, which typically took 12–15 minutes, is now completed in 3–4 minutes.
Touchless Accounts Payable (AP): Facing the challenge of having to manually validate 100% of invoice data extracts from a popular AP application due to unknown field-level accuracy, Ajit’s team created a three-layer validation process.
This includes:
Intelligent Document Processing (IDP) from the Mulesoft kit, which achieves 96.5% out-of-the-box accuracy in testing.
Agent Force Multimodality, which acts as a second extraction layer.15
An agent that performs field-level validation and, upon a match, makes a decision for approval.
PO Matching using Snowflake and Data Cloud, based on category and description, to achieve a two-way match for a fully touchless flow.
The touchless payables solution is expected to be deployed in production in approximately three weeks, with the potential for 100% auto-approval once the initial human-in-the-loop hypercare phase is complete.
Agentic AI vs. RPA and LLMs: Ajit demystifies the differences:
RPA: Is purely basic, rule-based automation with no generative content ability.
LLMs/Generic AI: Are generative, using backend inference of training data but are missing the action component.
Agentic AI: Is not only generative but can also execute actions within defined guardrails. The agent's ability to orchestrate with enough autonomy to achieve a goal is the “magic” that replaces the hardcoding required in traditional automation.
Advice for GBS Leaders and Employees: Ajit urges GBS leaders to get ready to reset benchmarks as customer experience expectations rise to instant and immediate results. He anticipates a shift in focus: contact centers will move from solving problems to building relationships, and back-office functions will move from transactional solving to leakage avoidance and materially impacting organizational benefit. For employees, the advice is to embrace the technology, become an expert in working with Agentic AI, understand the end-to-end domain, and become "technofunctional".
The Clear Message: Agentic AI is not a future concept; it is available now, and early adopters are seeing significant benefits.
Don't forget to like, subscribe, and share this episode!