In a significant leap forward for automation, Air India has integrated Salesforce's Agentforce into its operations, enabling AI agents to handle customer service processes, including refund requests, name changes, and email inquiries. This innovative use of agentic AI, which can examine customer requests, verify information across different systems, follow prescribed procedures, and initiate actions, has reduced internal processing time from approximately 14 days to a remarkable four hours for refunds. The expanded deployment also covers automated email resolution and name-change requests.
The shift towards agentic AI marks a departure from traditional chatbots that merely provide information. Instead, these advanced systems can work towards specific objectives by planning steps, gathering information, using digital tools, and taking actions with a degree of independence. In the case of Air India, an AI agent can retrieve the booking, check the applicable policy, verify passenger details, initiate the authorized refund process, and update relevant records.
Agentic AI operates in a cycle of understanding requests, planning actions, using tools, examining results, and deciding the next step. For instance, when a procurement manager asks an AI agent to investigate a delayed raw-material shipment, the agent identifies the supplier, purchase order, expected delivery date, accesses approved company records, checks inventory levels, examines available logistics updates, identifies potential issues, contacts the supplier through an authorized channel, and prepares a revised delivery plan. However, placing an alternative order may still require human approval due to its potential financial implications.
Unlike a fixed sequence of automated instructions, an agent can adapt its next action based on information gathered during the task. The degree of flexibility depends on how the system is designed. Powering these AI agents requires more than just a language model. They need access to information, software tools, and instructions that determine what they can do. The four essential components are the AI model, tools and APIs, memory and context, and instructions and safeguards that establish the agent's responsibilities, permissions, and restrictions.