HRS chief product officer Martin Biermann talks:
- AI inference costs
- Model routing and context efficiency
- Data consolidation and cyber security
- Productivity gains
Artificial intelligence could cut the cost of completing a typical corporate travel task by 75 per cent within two years, according to HRS chief product officer Martin Biermann. The corporate lodging and meetings platform has been applying AI across its hotel sourcing processes for the past few years and last year introduced HRS Copilot, a client-facing large language model tool designed to support lodging strategy and decision-making. This will soon be rolled out across the company’s meetings offering.
Biermann has been vocal about the ‘hidden’ costs of building and operating AI infrastructure, particularly the computing power required for long, agentic workflows involving thousands of invoice lines, millions of traveller profiles or billions of rate impressions.
But just as quickly as the technology is evolving, Biermann says the conversation around cost is changing. Even after accounting for the necessary infrastructure and data security investments, he estimates HRS could reduce the overall cost of delivering savings in a corporate travel programme by about 75 per cent. The projection combines falling AI inference costs with internal headcount efficiency ratios and productivity gains.
He recently sat down with BTN executive editor Lauren Arena to run through the numbers and explain why he expects AI to reduce supplier costs across the corporate travel ecosystem. The following conversation has been edited for length and clarity.
Business Travel News: Where are the biggest costs associated with implementing AI in corporate travel management workflows?
Martin Biermann: The picture is changing. On one side, things are getting dramatically cheaper. On the other, adoption is stagnating even as AI capability increases. This shouldn’t be quantified as ‘how many jobs can we get rid of’. That doesn’t make sense. Instead, I’ve been asking: What's the cost of saving 1 per cent on a corporate travel programme on average? And I believe we can slash it by 75 per cent within the next two years.
The wider discussion around AI costs has focused on concerns that increasingly powerful frontier models from companies such as Anthropic, OpenAI and Meta will increase per-token inference costs. The price per million tokens has remained relatively stable, but these companies can effectively increase consumption by breaking data into smaller units, meaning users burn through more tokens.
At the same time, commercial pressure is increasing as Chinese models from Alibaba and Quinn try to undermine the market dominance of those big Western, particularly US-bound, models. Open-weight and open-source models can cost a fraction of leading frontier models. They are not suitable for every task and hosting them in data centres in Europe [as HRS does] adds cost, but they create significant downward pressure on pricing.
BTN: What could bring inference costs down further?
Biermann: Model routing is one major lever. Companies can direct work to a cheaper model or choose a model based on latency. They can also batch workloads at different times, particularly as providers introduce variable pricing.
Context efficiency is another significant factor. The more information you cache, compress or prune, and the more efficiently you structure data within a prompt, the lower the token consumption. Architecture matters too: structured data should be accessed through databases and APIs rather than forcing everything into costly vector databases.
Specialised inference hardware should also reduce costs, while energy availability and grid capacity will increasingly influence where and how AI workloads are run. Taking these factors together, we believe average inference costs could fall from a baseline of about $15 per million tokens to around $6. In a best-case scenario, they could reach $1 to $3.
BTN: How does that translate into corporate travel management?
Biermann: Across procurement, travel management and accounting, we estimate the fully loaded cost of an average human task at roughly $100. That could include data analysis, programme configuration, supplier communication or accounting work.
As AI takes over more of the RFP process, the human component could fall to about $20 per task. The AI component could settle at around $5, based on an average workflow using about one million tokens. That takes the total from $100 to approximately $25.
This is a prediction – and will vary according to the size of your travel programme, how diverse your travel use cases are, ecosystem complexity and your geographical footprint is. All of that has implications… But the 75 per cent reduction remains our central estimate.
BTN: Does that estimate still hold once cybersecurity and compliance costs are included?
Biermann: Cybersecurity is the biggest additional risk. Moving from human agents to AI agents means granting equivalent permissions, which requires guardrails, monitoring, logging, trained security operations and compliance with measures such as the EU Artificial Intelligence Act.
Companies also need to account for potential breach, insurance, containment and incident-response costs. Threat actors are using AI too, so investment in security needs to increase alongside investment in AI. Even so, the savings in inference and human effort mean the overall projection still holds.
BTN: In your view, which travel management tasks should remain in human hands?
Biermann: Humans should continue to set strategy, make complex decisions, manage escalations and own supplier relationships. They also need to determine how a programme reflects company culture, traveller experience and goals around compliance, safety and sustainability.
AI should handle the technical plumbing: spreadsheet work, email administration, data ingestion, analysis, RFP execution, booking-tool configuration and reconciliation.
Traditional hotel RFPs can take around five months. By the time a programme change is implemented, the market has moved on. Buyers should be able to review programmes weekly, make strategic decisions and have AI execute the changes.
BTN: Data consolidation is an ongoing challenge for both travel buyers and suppliers. How does the quality and structure of travel data affect the economics of AI?
Biermann: Data is the prerequisite. Corporate travel remains fragmented across online booking tools, GDSs, TMCs, mid-office systems, analytics, expense and payment providers. If information is incomplete, inaccurate or out of date, the model will reach the wrong conclusions regardless of how sophisticated it is.
Control of the data structure also affects cost. Where structured data is available, it is more efficient to let AI access databases and APIs. Reliance on unstructured third-party data requires more complex retrieval infrastructure, which carries both cost and quality implications.
BTN: HRS launched Copilot last year using Anthropic’s LLM alongside its own specialised model and data. How has the tool’s underlying architecture evolved since then?
Biermann: HRS Copilot now uses a mesh of specialised agents rather than a single model. Different agents analyse spend, develop sourcing strategies, support hotel negotiations, configure rate caps and sort-order logic, audit rate availability and compare booked rates with invoices.
We select commercial, open-source and open-weight models according to the task and host relevant models in Europe to protect data sovereignty. Clients do not see the underlying complexity. They interact with one orchestration agent that coordinates the specialist agents.
BTN: Are AI skills and recruitment becoming another major cost?
Biermann: The cost of AI-oriented engineering talent is beginning to come down following the democratisation of the tooling, and most corporate travel providers do not need to train tier-one LLMs themselves. The cloud providers have also made inference easier to consume.
Suppliers still need a mix of engineering, data, product and travel-industry expertise. At HRS, we’ve introduced AI training with a sort of driving licence model that certifies people at different levels of competence across the company – not just our product and engineering staff. This training extends to customer service and sales, and all the way up to finance. We focus on driving the right agentic technology into these departments, as well as training the people operating and building these capabilities.
The larger challenge is redesigning workflows and collaboration around the technology, rather than simply teaching people how to prompt a model. It's more so about, okay, how do you apply AI? How do you make it work for you? What does that mean for your interaction with other departments?
When you consider the adoption curve, the biggest barrier right now is us humans understanding how we're going to work tomorrow. It’s a total paradigm and mindset shift.
BTN: What has AI done to HRS's own development costs and productivity?
Biermann: Agentic coding has quadrupled the output of our engineering teams and reduced time to market by 50 per cent. That makes engineering investment more effective because we can address more customer problems faster. Importantly though, product direction and industry expertise still needs to come from people and from co-creation with customers.
BTN: So, should buyers expect AI to make corporate travel technology and management services cheaper?
Biermann: Yes. It wouldn’t be logical to adopt AI at scale unless it reduces prices. Inference is becoming cheaper, engineering is becoming more productive and fewer people are needed for repetitive servicing work. Cybersecurity expenditure could be a concern, but if AI is applied correctly, I don’t see a reason why the cost of software should increase. Nor should operational costs, because fewer people are needed in customer service. The same applies to roles combining servicing and AI – those costs come down too because the human agent’s specialisation level is reduced.
Buyers should ask all the entities they work with whether they have the data, technical expertise and operating model to leverage the AI capabilities available today and then operationalise them in a way that drives down costs. Moving forward, they should also be able to adapt and pivot as the relevant technologies are developed.