How a travel and expense platform is breaking ground on a zero-hallucinations AI workforce

AI hallucinations are one of users’ biggest concerns when utilizing large language models (LLMs). And while many might expect front-runners like OpenAI and Anthropic to lead the way in addressing the issue, it’s a travel and expenses platform that may be breaking new ground.

On Wednesday, Navan revealed its new AI platform, Navan Cognition, which goes beyond single-purpose chatbots and basic AI tools to create an AI workforce capable of automating complex tasks.

“We never do cool technology for the sake of it,” says Ilan Twig, Navan’s cofounder and chief technology officer. “We use the best technologies to drive the best user experience.”

On June 20, the company confidentially filed for its initial public offering. Navan was last valued at $9.2 billion in 2022 after raising $304 million in equity and debt financing.

Navan’s Cognition-powered virtual travel agent, Ava, can book and reschedule flights and hotels, manage upgrades, process expenses, and provide 24/7 support through a conversational chatbot platform. But early in development, the company realized that for AI to be truly reliable, it must work unsupervised—and more importantly, with no critical hallucinations.

“A critical hallucination is when the user somehow, or the bot somehow, gets to the point where something undesirable happens,” Twig says.

A hallucination can impact both the user and the company, particularly in terms of travel, whether by “booking” a flight that doesn’t exist to satisfy a request, or offering a free upgrade the user isn’t entitled to.

With this in mind, Navan began using Cognition through Ava in 2023, ultimately finding that instead of using one generalist chatbot, a network of specialized agents working together produced more accurate and reliable results.

“We focused on a real-life problem, and we built the infrastructure to support that real-life problem,” Twig says.

Working as an organization

Inspired by the neural connections of the human brain, Navan Cognition deploys a modular multi-agent framework, with AI specialized in different areas, supervised for accuracy. In a way, Cognition works as a company org chart, breaking down AI into various departments with particular specialties, like booking flights or issuing refunds. Other “departments” serve as compliance for logic, and “managers” answer questions and liaise with others if a question is unknown.

Itamar Kahn, a neuroscience professor and principal investigator at Columbia University’s Zuckerman Institute says Navan identified the conditions that cause LLM hallucinations and developed solutions to eliminate them.

Kahn, a close friend of Twig, first heard about Cognition in its early stages three years ago, following the framework’s development. He also provided support for Twig’s recent white paper detailing how Navan Cognition works and what problems it aims to solve.

“I have a shared research interest with Ilan: high-order cognition. Essentially, any kind of system, artificial or biological, can react to changing circumstances in its environment, and respond to those efficiently,” Kahn says.

When a user asks Ava a question or assigns a task, Cognition routes it through several specialized agents to determine the best course of action. Meanwhile, a supervising AI agent checks responses for accuracy and credibility, acting as a safeguard.

Twig says he was inspired by the way supervisors in call centers review and learn from agent calls. “So I said, okay, I’m going to have a supervisor,” he tells Fast Company. “But instead of waiting for the end of the week and sampling two calls, I will actually do it for every response that the agent wants to send back to the user. It will first go to the supervisor to ensure that it doesn’t feel or smell like a hallucination.”

Navan’s departmental-like approach has proven effective, with Ava now handling around 8,000 chats daily, reportedly with zero critical hallucinations. The system’s lack of need for human oversight has also helped Navan scale without having to expand its travel support agent workforce.

Why is a travel platform at the forefront of AI innovation?

Innovation requires curiosity—a trait Twig has carried from childhood into his work at Navan.

At age 15, Twig became obsessed with light, building a virtual harp using mirrors, resistors, and the south-facing window of his childhood home in Israel. “I ended up having eight virtual beams of light connected to the computer. Whenever you disrupted any of the lights, it would generate a note. And it was the Do, Re, Mi, Fa, Sol . . .”

He has nurtured that same inventive spirit on his engineering team at Navan, encouraging fresh thinking on common problems. “We are curious. We are not afraid of making mistakes,” Twig says.

But it wasn’t just out-of-the-box thinking that led to Cognition’s development. Being a smaller company without a sole focus on AI forced Navan to innovate differently. “They’re not an AI company that is trying to solve the problem of large language models,” Kahn says. “They wanted to solve this problem for all of these customers. And I think this is why this system is working.”

With fewer resources than giants like OpenAI, Navan had to take a creative approach. “It’s a choice of architecture,” Kahn says. Rather than building another LLM to replace one with errors, Navan changed the inputs and outputs that inhibited the hallucinations.

With promising results, Navan is now preparing to scale its platform, making the Cognition framework available to other developers and companies to sign up for later this year.

“It is an amazing opportunity, because LLMs are new thing,” Twig says. “It opens the door to pretty much follow your imagination. And if you are persistent and curious, there is an opportunity to do something that no one else on the planet did.”

https://www.fastcompany.com/91353805/navan-cognition-zero-hallucinations-ai-workforce?partner=rss&utm_source=rss&utm_medium=feed&utm_campaign=rss+fastcompany&utm_content=rss