“We didn’t need more people. We needed smarter systems,” stated Shakti Goel, Chief Data Officer at Yatra, reflecting on the inefficiencies observed in the corporate travel division’s expense processing operations as of late 2023. The division was inundated with thousands of travel receipts from enterprise clients’ employees each month—PDF scans, email attachments, and photos—in over 100 languages, including Japanese, Greek, and Telugu. Each document required meticulous verification against corporate travel policies before reimbursement.
“It was repetitive, slow, and expensive,” Goel noted. “We were paying people to do something machines could do faster and more accurately.” The solution was an AI-powered receipt validation engine developed in-house, utilizing established large language models (LLMs) and seamlessly integrated with Yatra’s backend systems. The new system could process a receipt in under ten seconds, extract crucial data, validate it against policy rules, and provide a structured output ready for financial processing.
“The cost to run it was under ₹1,000 a month,” Goel explained, highlighting that the savings in salary costs were approximately ₹30 lakh annually.
This initial success underscored three critical insights: AI integration into live operations could be rapid, return on investment could be assessed within weeks, and implementation did not necessitate large budgets or extensive data science teams.
Building on this triumph, Yatra set out to develop a more extensive and ambitious initiative—a reusable Generative AI framework that could accommodate multiple intelligent “agents” across various operational and customer service workflows.
Goel emphasized a fundamental principle: “business alignment before technology investment.” He insisted that every AI initiative must begin with a clearly defined business case, measurable metrics, and appropriate stakeholder engagement. Instead of creating proprietary LLMs, Yatra’s framework leveraged established models like GPT-4 and Google’s Gemini. The organization’s edge lay in:
- Prompt Engineering: Crafting domain-specific instructions to ensure the AI comprehended context and adhered to corporate regulations.
- API Integration: Directly connecting AI to transaction systems to undertake actions—such as canceling flights or generating invoices—rather than merely providing information.
- Structured Outputs: Guaranteeing AI responses were returned in standardized formats like JSON for smooth integration with other systems.
Notably, the talent strategy fostered innovation. “We trained Python developers and even interns in-house,” Goel remarked. They did not require AI PhDs but needed the ability to understand processes, formulate effective prompts, and integrate APIs. Some of the production agents were reportedly developed by interns earning ₹20,000 a month.
Yatra formalized its approach through a 90-day playbook, a development cycle designed to advance an idea from conception to scale readiness or swiftly terminate it if it did not prove impactful.
The playbook’s first major test addressed a cumbersome customer process: flight cancellations. Previously, the cancellation procedure involved:
- Contacting an agent via phone or email.
- Allowing the agent time to verify booking details and passenger identity.
- Assessing compliance with corporate travel policies.
- Calculating cancellation fees and refund eligibility.
- Completing the cancellation in the booking system.
This process could be prolonged, particularly during peak periods.
Within the Generative AI framework, the team developed an AI-powered cancellation assistant capable of:
- Understanding natural language commands such as “Cancel my Mumbai–Hyderabad booking for August 5.”
- Real-time verification of corporate travel rules.
- Instant calculation of cancellation fees.
- Executing the cancellation independently.
“There are no menus, no hold times,” Goel asserted. “The bot understands your intent, applies the rules, and takes action.” The assistant rapidly iterated in a sandbox environment, with refinements made in response to misinterpretations or incomplete outputs. Safeguards were implemented to escalate ambiguous cases to human agents.
Within 60 days of deployment, the assistant was processing over 200 cancellations daily, with use extending from early clients like PwC and BCG to more than 25 large enterprises. Goel explained, “We measured not just usage, but whether the number of successful cancellations rose in proportion to bot sessions. That correlation demonstrated that we had successfully addressed the use case.”
Beyond just cancellations, the framework powered several additional functionalities, including:
- Email Automation: Managing over 1,000 corporate travel queries each day, decreasing average handling time from 30–40 minutes to under two minutes.
- Recommendation Engines: Merging user booking histories, corporate regulations, and live inventory to propose optimal itineraries.
- HR and Freight Bots: Automating internal service requests and logistics workflows.
- Invoice and Booking Assistants: Generating policy-compliant itineraries that include fare rules, baggage allowances, and cancellation penalties.
“Each new agent takes less time to build,” Goel noted. “We reuse the same infrastructure—only the prompts and API mappings change.”
The direct financial benefits have been significant:
- An annual savings of ₹30 lakh from the expense automation use case.
- Reduced hiring needs: Yatra previously required ten customer service personnel for certain workloads but now manages with only 2–3.
- Elastic scalability: Transitioning from a 2-core CPU to an 8-core, 64GB RAM setup on Google Cloud involved no additional hardware investment.
Goel emphasized that customer impact is equally paramount. “Enterprise clients now receive instant answers, 24/7, always aligned with their travel policies. This enhances their productivity and strengthens their trust in us.”
Risk management is an integral component of the framework. Early tests highlighted the potential for AI to act unpredictably; for example, cancelling a booking when the intention was simply to generate an invoice. Yatra now enforces:
- Strict domain scoping in prompts to avert off-topic actions.
- API verification layers before executing any irreversible tasks.
- Consistent structured output formats.
- Human escalation for unclear or multi-intent queries.
Looking ahead, the next objectives include:
- Expanding the cancellation bot to B2C clients.
- Deepening AI integration into analytics for demand forecasting and capacity planning.
- Developing workflows where different AI agents can transfer tasks among themselves.
“Don’t start with AI—start with a broken, costly process. Involve your business stakeholders early. Create a working MVP, test it in real-world scenarios, and within 90 days, you’ll discern if it’s worth scaling.”
This journey of AI adoption exemplifies the mission of the Making AI Work Summit & Awards, which promotes dialogue, collaboration, and innovation in AI, alongside its recognition platform—the ET AI Awards India—celebrating individuals and organizations making significant strides in the field.
