San Francisco-based Navan has deployed NVIDIA Nemotron 3.5 Super VL to power several AI travel agents handling customer conversations. Announced on October 16, 2026, the deployment targets more responsive assistance with business travel requests and changing itineraries.
The travel booking and expense management company adapted the model for specialised travel tasks. It says implementation took three weeks after receiving early access to the technology.
For business travellers, the potential benefit centres on managing journeys with less effort, particularly when flights or other arrangements change. However, Navan has not established that every customer will receive the model or experience faster problem resolution.
Disrupted Journeys Put Assistance to the Test
Flight cancellations can affect several parts of a business trip simultaneously. A replacement departure may require a different hotel night, revised ground transport and changes to scheduled meetings.
Navan’s approach uses specialised agents for flights, hotels, cars and trains. A coordinating agent directs requests to the relevant specialist while sharing information across the journey.
The intended result extends beyond explaining a disruption. Navan says its agents need to identify alternatives, present choices and help travellers complete necessary changes.
That capability could reduce the number of separate enquiries required during an interrupted trip. Nevertheless, successful assistance still depends on available inventory, current information and access to booking tools.
An AI system cannot create an available seat or hotel room where none exists. Its practical value depends on how effectively it handles the options available.
Internal Tests Support the Deployment Decision
NVIDIA released Nemotron 3.5 Super VL on October 15, 2026. The model supports tasks involving multistep reasoning, digital tools and different forms of information.
Before deployment, Navan compared its adapted version with an existing self-hosted production model and competing systems. Evaluations used consistent prompts, travel scenarios, tools and assessment criteria.
According to Navan, the fine-tuned model matched or exceeded competitors across six of seven measures. These covered areas including tool use, task completion, constraint compliance and relevant information handling.
The company also reported the lowest response latency within its comparison.
These results describe Navan’s internal testing. They do not establish independent industry leadership, improved customer satisfaction or a guaranteed outcome for individual travel requests.
Travel Expertise Shapes Model Training
Navan combined the NVIDIA model with training designed around its travel operations. General language capabilities alone do not explain how a platform manages reservations, company policies or support workflows.
The company used NVIDIA’s NeMo AutoModel tools and low-rank adaptation to customise the system. This approach adjusts selected model components without requiring complete retraining.
Its final training material contained approximately 10 million tokens, alongside nearly 2,000 validation tests. The evaluation focused on text-based travel tasks and tool use.
Navan also described a model architecture with approximately 120 billion total parameters, activating around 12 billion during each processing step. The design aims to balance reasoning capability with computational efficiency.
For passengers, those technical details matter chiefly through the service delivered. Clear options, accurate actions and dependable booking changes remain the meaningful measures.
Corporate Travel Teams Could Gain Support
More capable digital assistance could help corporate travel teams manage routine questions and straightforward itinerary changes. Human professionals could then focus on requests requiring more complex judgement.
Employees may also benefit when the system identifies options consistent with their company’s travel policy. However, the announcement does not demonstrate improved policy compliance across customer organisations.
Navan says the deployed model already supports customer conversations at a lower operating cost. It has not disclosed a precise saving or established corresponding reductions in corporate travel spending.
Lower technology costs do not automatically produce cheaper flights, hotel rates or service charges. Any financial benefit to customers would require separate evidence.
Reliability and Privacy Remain Central
Navan combines open and closed AI models, selecting systems for particular tasks. It fine-tunes them with travel expertise and hosts selected models on its own infrastructure.
The company says this provides greater control over speed, cost and reliability.
Meanwhile, travel assistance involves personal details, itineraries and business information. Companies adopting such systems need to understand how requests and traveller data move through their services.
Broader government attention to AI privacy reinforces the importance of clear commitments and accountability. It does not establish any specific misconduct by Navan.
Traveller Outcomes Will Determine Progress
The deployment advances Navan’s effort to provide assistance that can act on travel requests. Its internal results support that direction, while measurable passenger benefits remain to be demonstrated.
For business travellers, the decisive question is whether assistance delivers accurate choices and completed changes when plans unravel. Dependable service, clear communication and access to human help will shape confidence as these tools develop.
For more travel news like this, keep reading Global Travel Wire
