All casesNo. 066
BMW Group: An AI Agent Prepares Alphabet's Fleet Offers
An in-house agent automates about 90% of manual enquiry steps; people review its replies
- Case no.
- 066
- Primary source
- BMW Group news article (German original and English version)
- Source published
- 2026-05-22

1. Why this case
On 22 May 2026 the BMW Group published an article on how it uses agentic AI in day-to-day work. It gives two examples that are already running. The first is the offer process at Alphabet, the name under which BMW Group Financial Services bundles its European fleet business. There, an agent developed in-house reads fleet managers’ emails, moves the data into internal applications, starts the next steps, and drafts the reply. The second is the inventory of about 250,000 special tools that BMW manages, which is run through the suppliers that use them. There, an agent system sends the inventory orders, checks what suppliers send back, and closes the simple cases itself.
The case is useful for two reasons. First, the Alphabet example names where the manual work was (reading each email, pulling information from several systems) and puts a number on how much of it moved to the agent: about 90 percent of the manual work steps. Second, the two examples place the human check in different spots. At Alphabet, people review the agent’s draft replies and keep control over what they say. In the tool inventory, the agent approves unproblematic cases on its own, and people see only the cases that need expert judgment.
The article is BMW’s own account. Four months later, in its Capital Market Day release of 30 September 2026, CFO Walter Mertl called consistent use of agentic AI applications across all areas of the company a “game-changer,” and BMW set the goal of using AI across all core business processes, from development and purchasing through production and sales to aftersales. The two examples here are the concrete processes BMW pointed to before that announcement.
2. Problem
2.1 Fleet enquiries at Alphabet
Alphabet’s teams “support corporate customers with fleets of more than 50 vehicles.” BMW describes the enquiries in plain terms: “When it comes to ordering large numbers of vehicles for company fleets, enquiries are rarely ‘standard’.” And: “As the number of enquiries increases, so too does their complexity.”
The input is email. “Emails from fleet managers are often unstructured, requesting customised, multi-brand vehicle solutions” as well as details on mileage models or contracts. A single message can mix vehicle options from several brands, a question about annual mileage, and a contract detail.
The old way was manual. “Until recently, processing these enquiries required significant manual effort: each one was reviewed individually,” and relevant information was pulled together from multiple systems. Every enquiry needed a person to read it, work out what was being asked, and look up the answers in more than one internal system before a reply could be written.
2.2 Special-tool inventory in Purchasing
BMW’s Purchasing and Supplier Network manages “approximately 250,000 specialised tools worldwide,” ranging from casting moulds and models to templates. They are used in component production and in the maintenance of large machines at suppliers’ facilities, and the inventory requests go to those suppliers.
Those tools have to be inventoried. “Until recently, inventorying them was time-consuming, labour-intensive and complex.”
Confirmed facts: Alphabet = BMW Group Financial Services’ European fleet business; customers have fleets of more than 50 vehicles; enquiries rising in number and complexity; unstructured emails about multi-brand vehicle solutions, mileage models, contracts; before = each enquiry reviewed individually, information combined from several systems. About 250,000 special tools (casting moulds, models, templates) used in component production and machine maintenance at suppliers; inventory was time-consuming, labour-intensive, complex.
3. Solution
3.1 Alphabet: an in-house agent from email to next step
“A dedicated agent developed in-house now supports the process, largely through automation.” In the new workflow, the system “autonomously transfers data to internal applications” and “directly initiates the necessary steps.” Entering the enquiry data into internal applications is now the agent’s job, and the follow-up steps start without someone triggering them. Under the old way, staff reviewed each enquiry and combined information from several systems by hand.
The agent also writes the reply. The German version of the article says the staff check “die von der KI erstellten Antwortentwürfe,” the draft replies created by the AI. The English version calls them “the AI’s proposals.”
3.2 Alphabet: people keep the content decision
BMW states the division of labour directly. “Despite the high degree of automation, control over content remains firmly with Alphabet’s team members.” Staff “review the AI’s proposals, refine them as needed and intervene where necessary.” The article sums it up: “So, while the technology handles the preparation, decision-making remains firmly with humans.” The German subheading is shorter still: preparatory work by the AI, responsibility with people.
3.3 Purchasing: an agent system that runs the inventory round
For the special tools, “an agentic system supports Purchasing teams by automatically drafting inventory orders,” then “sending them to suppliers, reviewing responses and approving unproblematic cases.” People come in “only where specialist understanding and expertise are needed.” Here the agent does not stop at a draft. It sends the requests out, reads the replies, and closes the simple cases without a person approving each one.
3.4 AIconic as the base
The inventory system runs on the AIconic multi-agent system, which the German article calls the central AI application in the BMW Group’s Purchasing and Supplier Network. The English version describes what it does: “This interconnects various data sources, creates tasks, verifies output and provides transparency” across all process steps. So the platform does more than host the agents. It links the data, turns work into tasks, checks results, and shows the status of every step.
Confirmed facts: Alphabet agent developed in-house; transfers data into internal applications on its own and starts the necessary steps; drafts replies (German version), which staff review, adjust, and intervene on; content control and decisions stay with Alphabet staff. Purchasing agent system creates inventory orders, sends them to suppliers, checks responses, approves unproblematic cases; people intervene only where expertise is needed; built on the AIconic multi-agent system, which connects data sources, creates tasks, checks results, and shows status across steps.
4. Impact
All results come from the BMW Group article of 22 May 2026. No measurement period or baseline is given.
Alphabet: about 90 percent of manual steps. The English version says the agent’s work “replaces around 90 percent of previously manual tasks.” The German version: the agent “automatisiert so rund 90 Prozent der manuellen Arbeitsschritte,” about 90 percent of the manual work steps. The figure is a share of the manual steps in handling an enquiry. It is not a share of emails or enquiries handled without people, since staff still review the draft replies.
Alphabet: time moved to customers. BMW says the change is “significantly streamlining workflows and freeing capacity for direct customer engagement and support.”
Special-tool inventory: leaner, no figure. “It makes inventorying considerably leaner, more transparent and more efficient – as well as less labour-intensive.”
5. Insight
The 90 percent is preparation, not decision. The Alphabet agent took over moving the enquiry data into internal applications, starting the next steps, and drafting the reply, work that staff used to do by reviewing each email and combining information from several systems. The part that stayed with people is the content of what goes back to the customer. That is why a process can be about 90 percent automated by steps and still leave the reply in a person’s hands. Read the figure as a measure of preparation work removed, not of enquiries handled without staff.
The human check sits in a different place in each process. At Alphabet, the person reviews the agent’s output before it reaches a corporate customer. In the tool inventory, the agent approves the unproblematic cases itself and only the rest reach a person. The visible pattern is that the customer-facing offer keeps review on the output, while the internal check against supplier responses keeps review on the exceptions.
Purchasing is adding agents. The inventory system is built on AIconic, which already connects purchasing data sources and tracks every step. In October 2026, Dr. Web, citing Handelsblatt and ntv, reported four AI agents in BMW purchasing, including one that checks which suppliers can deliver more at short notice, and said purchasing still makes the decision. BMW’s own September release sets the goal of using AI across all core business processes.
Limits. This is BMW’s own article. It gives one result figure, the 90 percent, without defining how steps were counted. There are no enquiry volumes, no rollout date, no measurement period, and no figures for the tool inventory. No model or vendor is named. The English introduction calls the inventory “fully automated”; the section itself says people still handle cases that need expertise.
One-line takeaway: In a process built on unstructured customer emails, the agent’s share is the preparation: BMW’s Alphabet had an in-house agent automate about 90 percent of the manual work steps and kept people in charge of the reply.
6. Implementation hypotheses (not in the source)
6.1 From unstructured email to a structured enquiry record
Basis: Emails are “often unstructured, requesting customised, multi-brand vehicle solutions” and details on mileage models or contracts; before, information “was drawn together from multiple systems”; now the system “autonomously transfers data to internal applications” and “directly initiates the necessary steps.”
The agent most likely starts by turning each email into a fixed set of fields: which corporate customer is writing, what kind of request it is (new vehicles, a change to mileage, a contract question), which makes and models, how many vehicles, what annual mileage and term. Matching the sender to an existing customer account would let it pull that customer’s framework terms and current contracts from the systems staff used to search by hand. “Initiates the necessary steps” probably means creating the enquiry or offer case in the internal application and starting the offer calculation, and for multi-brand requests possibly requesting prices for non-BMW vehicles. When a field is missing or ambiguous, the safer design is a draft that asks the fleet manager for the missing detail, not a guessed value, since the draft goes to a person anyway.
6.2 One review queue before anything reaches the customer
Basis: Staff “review the AI’s proposals, refine them as needed and intervene where necessary”; the German version says they check the draft replies created by the AI; “control over content remains firmly with Alphabet’s team members.”
The draft reply is probably assembled from the data the agent collected (offer figures, contract terms) rather than written freely, and placed in the responsible account team’s work queue next to the original email and the record it created. The person edits and sends it. Because every reply passes this queue, the remaining manual share of the work is likely concentrated here: checking that the agent read the request correctly, adjusting commercial terms, and handling requests the agent could not structure. Logging how often and where staff change the drafts would be the natural way to see which request types the agent handles well.
6.3 Inventory rounds with auto-approval and an exception queue
Basis: The agent system creates inventory orders, sends them to suppliers, checks responses, and approves unproblematic cases; people step in only where expertise is needed; AIconic connects data sources, creates tasks, checks results, and shows status across all process steps.
An inventory round likely starts from the tool register: for each supplier site, the system lists the BMW-owned tools recorded there and sends the supplier an order to confirm them. When the answers come back, the agent compares each one with the register. A tool confirmed at the recorded location and in the expected state would count as unproblematic and be closed. A tool reported missing, moved, or damaged, or a supplier that does not answer, would become a task for a buyer in AIconic, with the original order and response attached. The status view across all steps would then show, per supplier and per tool, what is confirmed, what is waiting for an answer, and what is with a person.
7. Sources
- BMW Group — Wenn KI den Alltag leichter macht: Agentic AI bei der BMW Group.
https://www.bmwgroup.com/de/news/allgemein/2026/agentic-AI.html
2026-05-22 (accessed 2026-11-17) - BMW Group — How agentic AI is making life easier at the BMW Group. (English version of source 1)
https://www.bmwgroup.com/en/news/general/2026/agentic-AI.html
2026-05-22 (accessed 2026-11-17) - BMW Group press release — Agility, efficiency, AI-based innovations: BMW Group sets course for greater profitability and resilience
https://www.press.bmwgroup.com/global/article/detail/T0461229EN
2026-09-30 (accessed 2026-11-17) - Dr. Web (Michael Dobler) — BMW setzt im Einkauf vier KI-Agenten ein und will ab 2028 als KI-orchestriertes Unternehmen arbeiten
https://www.drweb.de/bmw-ki-agenten-einkauf-2028/
2026-10-13 (updated 2026-10-22; accessed 2026-11-17)