BMW purchasing chief Nicolai Martin is bringing AI into his leadership team. Agents such as “Mr. Flexibility” are intended to support him and his team in complex procurement decisions.BMW
BMW is taking agentic AI beyond routine automation with “Mr. Flexibility”, an assistant combining capacity, cost and production data for procurement decisions. The aim is to answer complex volume questions in minutes and support faster purchasing decisions.
Answering that question is less straightforward than it sounds. Additional volume has to be checked against production plans, available parts, technical approvals, supplier capacities and cost. The information sits across different functions and IT systems, so reaching a reliable answer can take weeks. BMW wants to compress that process with “Mr. Flexibility”, an AI assistant that combines the relevant information and shows much faster what extra volume could be built – and at what cost. The ambition goes beyond productivity: BMW wants AI to become an additional input into procurement decisions.
Why is BMW shifting from efficiency to effectiveness?
The scale of BMW’s purchasing organisation explains why the company is looking for a simpler way to navigate complexity. Its network includes around 2,600 direct suppliers and hundreds of thousands of companies further down the supply chain. Every day, roughly 36 million parts have to reach more than 30 production sites at the right time, in the right quantity and at the required quality. Around 70 per cent of a vehicle’s value creation is generated together with the supplier network.
For Nicolai Martin, BMW Board of Management member responsible for Purchasing and Supplier Network, this is where AI can have its biggest impact. He deliberately distinguishes between efficiency and effectiveness. “Am I only trying to make my employees more efficient in their routine work – or am I trying to use AI agents to manage the entire complexity of the supply chain in a way that creates real business value?” he asked during a demonstration of the assistant.
Martin puts the difference into financial terms. He describes the fixed cost base of the purchasing organisation as a figure in the mid-hundreds of millions of euros. Using AI to run that organisation more efficiently can reduce costs. But this sits against a purchasing volume of €80 billion. “Even small improvements in material costs, capacity or flexibility therefore have a disproportionately greater leverage,” he said.
BMW consequently does not want to measure the effect solely in hours saved. For its overall programme of agentic applications in purchasing, Martin is targeting a business impact in the hundreds of millions of euros next year. That includes better management of flexibility and material costs. The important shift, in his view, is from accelerating existing processes to enabling better decisions while there is still time to influence the outcome.
How does an AI assistant become part of management decisions?
That is why “Mr. Flexibility” is not intended to produce analyses somewhere in the background. Martin wants the assistant where the questions arise: in management meetings. Capacity, cost and scenario information should be available as close to real time as possible, allowing the leadership team to test alternatives during the discussion rather than taking a question away and returning later with an analysis.
This reflects how BMW is developing the technology internally. According to Martin, his management team has worked with AI every week for around 15 months. The process is deliberately not built around polished presentations. Instead, the agents are challenged directly: Why can this question not be answered? Which tool is missing? Where is the data foundation still insufficient?
Martin therefore dislikes describing the system as just another software application. “It isn’t an IT tool being rolled out; it is a new colleague. And the new colleague keeps learning.” Behind that metaphor is a staged target model. BMW distinguishes between an AI-assisted company, an agentic model in which data and systems are connected along processes, and eventually an AI-orchestrated organisation.
Hendrik Lang, Head of Purchasing and Supplier Network for High-Voltage Storage Systems, can now review capacity scenarios within minutes rather than weeks.BMW
How can BMW turn weeks of analysis into minutes?
The iX3 provides a practical example. If BMW wants to build additional vehicles, at least three information domains have to come together. Production provides the planned vehicle programme. Purchasing knows the available and contractually agreed capacities. Engineering holds the technical master data that determines which components may be used in which vehicle variant.
Even a single question about additional volume therefore requires data to be combined across organisational boundaries and assessed by specialists. “In the past, a request like this took several weeks before we received an answer. With Mr. Flexibility, we can answer this question within a few minutes,” says Hendrik Lang, who heads Purchasing and Supplier Network for high-voltage storage systems.
Why does battery capacity put Mr Flexibility to the test?
The battery is particularly useful for demonstrating the value of the assistant because several layers of complexity converge there. According to Lang, it is by far the most expensive component in a battery-electric vehicle, while additional capacity can only be added with long lead times. BMW says significant capacity increases require at least two years.
“Even seemingly simple alternatives create new dependencies,” he explains. Cells can only be produced in advance within a limited time window because of ageing requirements. Switching between regions can in turn affect the vehicle’s local content and therefore potential tariff costs.
This illustrates why time matters commercially. An increase in vehicle volume can initially appear technically feasible. If a full cost assessment then takes several weeks, BMW may discover too late that the extra units trigger a cost step elsewhere in the component base, after shifts and production schedules have already been adjusted. At that point, reversing the decision becomes difficult. A response within minutes therefore changes more than workload: it changes when the decision can be made. Scenarios can be compared before commitments become difficult to reverse.
Why does BMW keep calculations away from the LLM?
Because the answers can feed directly into management decisions, they cannot merely sound convincing. BMW therefore separates language processing from technical calculation. “The large language model does not calculate anything for us. We use deterministic tools everywhere,” explains Maik Leuthold, Product Owner of “Mr. Flexibility”.
The language model primarily acts as an orchestration layer. It interprets the user’s question, identifies the context and decides which tool is required. The LLM therefore provides the natural-language interface to existing IT systems rather than replacing their calculations. BMW is separating linguistic flexibility from computational reliability.
That also removes some technical complexity for users. “Our purchasers interact with the agent like they would with people,” Leuthold says. Questions can be phrased much as they would be to a colleague, without the user having to know every underlying source system.
Spotting bottlenecks at a glance: Maik Leuthold, Product Owner of “Mr. Flexibility”, explains how the agent works.BMW
BMW also wants to remain flexible in its choice of models. “We are provider- and model-agnostic,” Leuthold says. Language models can be selected according to performance, speed and cost, then replaced when necessary. The architecture therefore avoids tying the process to one model provider. But this makes a less glamorous issue even more important: the quality and freshness of the underlying data.
What makes data and governance critical for agentic AI?
An answer produced in three minutes is of little value if the capacity data behind it is three months old. BMW therefore makes data quality part of the operating process. The system shows how current stored information is. If data has not been confirmed for a longer period, its status changes and the responsible employee is prompted to review it. The system is also intended to identify and flag anomalies in data series.
Current data does not solve every problem, however, because not every employee is authorised to see every piece of information. Access through “Mr. Flexibility” follows the permissions of the individual user. The agent does not receive automatic company-wide information privileges merely because it can combine data from different functions.
A technical question consequently becomes an organisational one. Who is allowed to know what? Who assesses the recommendation? And who carries responsibility for the final decision? For Martin, that responsibility remains with people. “Mr. Flexibility” is therefore not an autonomous decision-maker at the table. It supplies answers, tests scenarios and prepares decisions.
