By 
[San Francisco, October 16, 2026] — Corporate travelers facing the stress of canceled flights and shifting itineraries may soon experience a more streamlined recovery process. Navan, a leader in business travel and expense management, announced on 16 October 2026 the deployment of NVIDIA Nemotron 3.5 Super VL to power its suite of AI travel agents. This strategic move is designed to increase the responsiveness and efficiency of digital assistants when managing intricate corporate travel requests.
The integration was completed with remarkable speed, as industry reports indicate Navan finalized the deployment within three weeks of gaining early access to the NVIDIA model. Rather than deploying the technology as a generic tool, the company heavily customized the model to align with the specific nuances of the travel industry.
Strategic Hybrid AI Architecture for Corporate Travel
Navan is implementing a sophisticated strategy that blends open and closed AI models to optimize its service delivery. By selecting specific models for distinct tasks and fine-tuning them with deep travel-sector expertise, the company can host critical systems on its own infrastructure. This architectural choice is intended to provide the company with tighter control over operational costs, system reliability, and the speed of responses provided to users.
For organizations that manage high volumes of employee transit, these technical optimizations could significantly reduce the friction associated with travel administration. While the deployment is active, it remains unclear if the updated model is currently available to the entire global user base or if it provides a universal guarantee for faster resolution of all booking conflicts.
Solving the Domino Effect of Flight Disruptions
The true utility of AI in the travel sector is tested most severely during irregular operations (IROPS). A single flight cancellation often triggers a cascade of logistical failures, requiring the traveler to secure a new flight, modify hotel stays, rearrange ground transportation, and communicate delays to business partners.
According to company statements, a high-functioning AI travel agent must move beyond merely notifying a passenger of a cancellation. The goal is a system that can autonomously identify viable alternatives, present these options clearly to the user, and execute the necessary transactional changes in real-time.
To achieve this, the platform utilizes a network of specialized AI agents dedicated to specific travel verticals:
- Flight Agents: Managing departures and rebookings.
- Hotel Agents: Handling lodging modifications.
- Car Agents: Coordinating ground transport.
- Train Agents: Managing rail itineraries.
A central coordinating agent acts as the "brain," routing user requests to the appropriate specialist while ensuring that data is shared across the entire trip itinerary. This interconnected approach aims to simplify the management of linked travel arrangements through a single digital interface.
Internal Benchmarks and the NVIDIA Nemotron 3.5 Performance
NVIDIA released the Nemotron 3.5 Super VL on 15 October 2026, specifically engineering the model for systems that require multi-step reasoning, the ability to utilize digital tools, and the processing of multimodal information.
Navan subjected the new model to a rigorous internal evaluation, comparing it against its existing self-hosted production systems and other market competitors. The testing framework utilized standardized prompts, diverse travel scenarios, and consistent evaluation criteria.
The results of these internal tests suggest a significant performance leap. Navan reported that the fine-tuned model either beat or matched competing systems across six of seven key metrics:
- Effective use of digital tools.
- Successful task completion.
- Adherence to specific constraints.
- Utilization of relevant data.
- Maintenance of coherent, useful conversations.
- Overall reasoning capabilities.
Furthermore, the company noted that this model achieved the lowest latency among all tested systems. It is important to note that these figures stem from internal company data rather than an independent third-party audit, meaning real-world customer satisfaction rates are still being monitored.
Technical Adaptation via Low-Rank Adaptation (LoRA)
General AI models often lack the specific knowledge required to navigate a particular company's travel policies or the technical backend of a booking platform. To bridge this gap, Navan employed NVIDIA’s NeMo AutoModel training tools and a technique known as low-rank adaptation (LoRA).
LoRA allows developers to customize specific parameters of a model without the prohibitive cost and time required to retrain the entire system from scratch. The technical specifications of the training process included:
- Training Data: Approximately 10 million tokens of supervised training information.
- Validation: Nearly 2,000 distinct validation tests.
- Focus: Text-based travel tasks and tool-driven execution.
The model architecture features approximately 120 billion total parameters. However, it utilizes a "mixture-of-experts" design, activating only about 12 billion parameters for each individual processing step. This design is specifically intended to maintain high-level reasoning capabilities while ensuring computational efficiency and rapid response times.
Impact on Corporate Travel Management and Privacy
The shift toward more capable AI assistance allows corporate travel managers to offload routine inquiries, freeing human professionals to handle high-stakes or highly complex travel crises. Additionally, the system helps employees adhere to approved travel options and corporate spending limits automatically.
While Navan indicates that the new model is already reducing the cost of customer conversations, the company has not released specific percentage savings or detailed how these efficiencies will translate into direct cost reductions for its corporate clients.
As AI takes a more central role in handling sensitive itineraries and personal data, privacy remains a primary concern. In March 2026, the US Government Accountability Office issued a report highlighting the inherent privacy risks of AI. Simultaneously, the Federal Trade Commission has reiterated that AI providers must strictly adhere to confidentiality commitments. While these are broad industry warnings and not specific to Navan, they underscore the regulatory environment in which these AI agents operate.
Why This Matters: The Shift to Action-Oriented Travel
For the business traveler, the value of this update is not found in the "fluency" of the AI's conversation, but in its ability to execute tasks. The industry is moving away from "Chatbots" (which provide information) and toward "AI Agents" (which perform actions).
From a logistical standpoint, this creates a future where a traveler doesn't have to spend two hours on hold with an airline after a cancellation. Instead, the AI agent identifies the next available flight, checks it against the company's travel policy, and re-books the ticket and the connecting car service before the traveler has even left the gate.
However, the success of this transition depends on three critical factors: the availability of real-time API data from airlines and hotels, the system's ability to handle payment transactions securely, and the availability of a "human escape hatch" when AI reasoning fails. If Navan can prove these results outside of internal testing, it could set a new standard for the corporate travel experience.
Slug: navan-nvidia-ai-travel-automation
