when NabangBusiness Travel and Expense Management Company was established to build virtual travel agents powered by artificial intelligence Called Avait faced the challenge of Please be sure The system did not hallucinate.
Even in the company's settings 1 AI generation error – Like Do you offer a refund that violates fare rules? It shows you're wrong Flight details – I did it This leads to customer dissatisfaction, financial loss, or regulatory penalties.
To overcome this, Naban built it Naban recognitiona platform with multiple layers of agents, AI supervisors, mindsets, inference tools, and rules-based and large-scale language model-driven assessment. These multiple backstops will reconfirm customer questions and ensure curitaCy.
Navan used this platform to drive AVAs. AVA processes thousands of customer queries every day. Two years, that There was nothing According to Navan, an incorrect upgrade or inconsistency between the cost and the statement. AVA handled tasks equivalent to the tasks that dozens of people first performed agent, It then expanded to work done by hundreds of workers.
“Our goal is to set new standards for business-ready AI,” said Navan Co-Founder and Chief Technology Officer. Ilan Twig He told Pymnts.
AVA is not just about AI chatbots. He is also an agent.
“The AVA actually takes action, whether it's a flight cancellation or change, a refund, a booking sheet or a class upgrade,” Twig says. “AVA can even understand that users get frustrated and that even if AVA can handle interactions on its own, it automatically forwards chats to human travel agents.”
Smarter AI Travel Assistant does More and more Especially important for business travel, to ensure customer satisfaction. I'll be resurrected. According to Global Business Travel Associationalmost half of travel buyers expect their company to take more business trip This year, 57% also saw an increase in travel expenses in 2025.
But the twig said Navan cognition can be used For use cases that go beyond travel.
“We've built something like an auto-scheduled personal mailing list so that users can request daily reports on a particular subject,” he said. “This mailing list is a 'conversation'. This means that users can not only read the news, but also continue the conversation natively for any part of the content they are interested in. ”
“We're experimenting with more applications. Honestly, it's just damaging the surface,” Twig added.
Read again: Navan IPO to test the appetites of investors on B2B Fintech platforms
How Navan cognition works
Twig said that the level of autonomy in Ava does not work without Navan cognition. Its guardrails include “rule-based and AI-driven supervisors, strict API verification, and filters that maintain protection of sensitive information.”
“The best part is that there is continuous context checking and automated intervention, which allows agents to actively catch and correct issues, ensuring that they are always accurate and policy-compliant,” Twig says.
A simple way to think about how Navan Cognition works is to compare it to the company's organizational chart, he said.
“It has a 'Reasoning Wisdom' module that specializes in leading an experienced team. A supervising node acting like a compliance department ensures that everything is in compliance with logic and business goals,” Twig says. Also, the searched Generation (RAG) “Behave like a manager, answer user questions, and enlist as administrator if you don't know the answer.”
The result is what Navan calls “critical hallucinations of zero.” Research paper From the company.
Naban is like that give Other companies access Navan recognition So they can Build your own “Non-Imaginable Hallucination” AI Agent.
“Cognition enables us, and ultimately, any business, to build trusted, professional AI agents that handle complex workflows behind the scenes, such as running a virtual travel assistant AVA,” says Twig. “It's empowering intelligence to support a seamless experience that customers already know.”
Navan recognition, he said, allows other companies to build similar AI systems tailored to their own workflows.
This platform relies on large language models, meaning it works with commercial or open source language models. It also offers one-click deployment, automated testing and an intuitive flow designer Let's forgive Teams build enterprise-ready AI workflows with no deep engineering resources.
Navan Cognition was designed for business users and product builders, not just for technical teams Quickly deploy AI-driven solutions. Its “no critical hallucination” ability makes it particularly relevant to companies in a highly regulated industry, Twig said.
“Like building and managing human teams, Navan Cognition provides tools to create AI agents, train domain expertise, monitor performance, and ensure that they operate within defined parameters,” according to the company. Blog post Written by Twig.
In one real-world test, Navan's team built and deployed a fully functional AI scheduling assistant within an hour using the Navan Cognition interface Post. I said. task I'll do it usually Need it Site reliability engineers and AWS expertise.
Other applications are coming, according to Twig. “I've just started,” he added.
Subscribe daily for all PYMNTS AI coverage AI Newsletter.
read more:
