Navan CTO on Building an Agentic AI Platform With ‘Zero Critical Hallucinations’

When Navan, a business travel and expense management company, set out to build its artificial intelligence-powered virtual travel agent called Ava, it faced the challenge of making sure the system didn’t hallucinate.

In corporate settings, even one AI-generated error — like offering a refund that violates fare rules or showing the wrong flight details — could lead to customer dissatisfaction, financial loss or regulatory penalties.

To overcome that, Navan built Navan Cognition, a platform with multiple layers of agents, AI supervisors, chain-of-thought and reasoning tools, as well as rule-based and large language model-driven assessments. These multiple backstops double-check the responses to a customer’s questions to ensure accuracy.

Navan used this platform to power Ava, which handles thousands of customer queries daily. For two years, it did not have any unauthorized upgrades or mismatches between costs and statements, according to Navan. Ava initially handled tasks equivalent to those done by dozens of human agents, and then later expanded to work done by hundreds of workers.

“Our goal is to set a new standard for business-ready AI,” Navan co-founder and Chief Technology Officer Ilan Twig told PYMNTS.

Ava is not just an AI chatbot; it’s also an agent.

“Ava actually takes action, whether that’s canceling or changing a flight, issuing refunds, booking seats or upgrading classes,” Twig said. “Ava can even understand when a user gets frustrated and transfers the chat automatically to a human travel agent even if Ava could handle the interaction by itself.”

Smarter AI travel assistants will increasingly be crucial to ensuring customer satisfaction, especially as business travel revives. According to the Global Business Travel Association, nearly half of travel buyers expect their companies to take more business trips this year, and 57% also see increased travel spending in 2025.

But Twig said Navan Cognition can be used for use cases beyond travel.

“We’ve built things like automatic scheduled personal mailing lists, allowing the user to request daily reports on any given subject,” he said. “This mailing list is ‘conversational,’ which means the user can not only read the news but also can natively continue the conversation over any part of the content that interests them.”

“We’re experimenting with even more applications, and honestly, we’re just scratching the surface,” Twig added.

Read also: Navan IPO to Test Investor Appetite for B2B FinTech Platforms

Twig said Ava’s level of autonomy wouldn’t work without Navan Cognition. Its guardrails include “rule-based and AI-driven supervisors, strict API validation and filters that keep confidential information protected.”

“The best part is that we have ongoing context checks and automated interventions, which allows our agents to catch and correct issues proactively, ensuring accurate and policy-compliant responses at all times,” Twig said.

A simple way to think about how Navan Cognition works is by comparing it to a company organization chart, he said.

“It has ‘reasoning wisdom’ modules that specialize like the experienced team leads of a company; supervisory nodes that act like the compliance department, ensuring that everything is in compliance with logic and business goals,” Twig said. It also has retrieval-augmented generation (RAG) “pipeline architecture that acts like a manager, answering user questions and enlisting management if they don’t know the answer.”

The result is what Navan calls “zero critical hallucinations,” according to a research paper from the company.

Navan is giving other companies access to Navan Cognition so they can build their own “zero critical hallucination” AI agents.

“Cognition makes it possible for us, and eventually, for any business, to build reliable, specialized AI agents that handle complex workflows behind the scenes, like running our virtual travel assistant Ava,” Twig said. “It’s about powering the intelligence that supports the seamless experience our customers already know.”

With Navan Cognition, other companies will be able to build similar AI systems tailored to their own workflows, he said.

The platform is large language model-agnostic, meaning it can work with any commercial or open-source language model. It also offers one-click deployment, automated testing, and an intuitive flow designer that lets teams build enterprise-ready AI workflows without deep engineering resources.

Navan Cognition was designed not just for technical teams but for business users and product builders looking to rapidly deploy AI-driven solutions. Its “zero critical hallucinations” capability makes it especially relevant for companies in highly regulated industries, Twig said.

“Just as you would build and manage a human team, Navan Cognition gives you the tools to create AI agents, train them with your domain expertise, monitor their performance, and help ensure they operate within defined parameters,” according to a company blog post written by Twig.

In one real-world test, Navan’s team built and deployed a fully functional AI scheduling assistant in under an hour using Navan Cognition’s interface, the post said. The task would normally require site reliability engineers and AWS expertise.

Other applications are coming, Twig said, adding: “We’re just getting started.”

For all PYMNTS AI coverage, subscribe to the daily AI Newsletter.

Read more:

• Business Travel Bounce in 2025 Demands Better Expense Management and Virtual Cards
• Brex and Navan Launch Integrated Business Travel and Payments Solution
• Citizens and Navan Partner on Travel Management Platform

The post Navan CTO on Building an Agentic AI Platform With ‘Zero Critical Hallucinations’ appeared first on PYMNTS.com.