How Audi is using AI to rethink Order-to-Delivery
An operator remains in the cab – despite the emphasis on automation, Audi does not intend to translate major AI recommendations directly into action without human review.
Audi
Audi is applying AI across its Order-to-Delivery process instead of isolated use cases. By linking production, planning and logistics data, the OEM aims to cut manual work, improve forecasts and uncover previously hidden optimisation potential.
Order-to-Delivery is one of Audi’s broadest operational processes. It stretches from a customer order through production and final vehicle distribution, requiring planning, manufacturing, logistics and supply chain functions to work together. Unplanned bottlenecks can trigger extensive replanning, while new model launches create uncertainty because planners initially lack historical demand data.
A modern electric vehicle can involve thousands of suppliers and more than 5,000 part numbers that have to reach the production line in the right configuration and at the right time. Once a vehicle has been built and quality-checked, it still has to be distributed to the customer’s market and handed over. Audi’s response is to treat this chain as an integrated optimisation problem rather than a collection of isolated tasks.
Why is Audi treating O2D as one AI problem?
For Dieter Braun, Head of Supply Chain Management at Audi, the key distinction is between local AI improvements and transformation at process level. An isolated bottom-up application – for example in sequencing or weld-quality analysis – can generate significant savings in one activity. But even a local saving of around 50 per cent may have only a limited effect on the overall process.
Audi therefore wants to apply AI from the top down across Order to Delivery, a process that every customer vehicle passes through. The underlying idea is that embedding AI across an entire end-to-end process creates a larger lever than optimising individual activities separately. Better results can emerge when datasets that were previously used independently are combined and evaluated in context.
“If we apply the AI lever top-down to the Order-to-Delivery process that every Audi customer vehicle goes through, we achieve a very different benefit and a much greater savings potential,” Braun says. For that reason, O2D is one of the areas Audi is currently transforming with AI.
Audi also expects integrated data analysis to reveal planning options that were previously difficult to identify in routine operations.
Which AI methods does the process require?
There is no single model behind the programme. Depending on the process step, Audi uses machine learning, pattern recognition, dashboarding, operations research and large language models. Sebastian Reinisch, Head of Data Analytics, AI and IT Security at Audi in Neckarsulm, describes the challenge as using the full AI toolbox rather than forcing every problem into the same approach.
The rapid development of AI also changes which problems can realistically be addressed. Audi therefore revisits use cases to assess whether newer technologies can now solve tasks that were previously impractical.
Andreas Kühne, AI programme manager at Audi, says the company first analysed the O2D chain end to end and identified potential AI applications. At the same time, the team examined which information each use case requires and which datasets can be reused across several capabilities. The aim is to move from project-specific data preparation towards scalable data products.
Production alone contains several hundred systems that either generate or provide data, often for one narrowly defined purpose. When those datasets are reused elsewhere in Order to Delivery, their structure, quality and semantics may not fit the new task. Creating a shared data foundation across production and logistics therefore becomes part of the transformation itself rather than a preliminary IT exercise.
Kühne argues that this cannot be solved by IT alone. Business functions, IT teams and digitalisation specialists have to work together because the meaning and quality of the data are tied directly to operational processes.
How does the Cold Start Planner forecast demand?
One example is volume planning for new models. Before a launch, planners have no historical order pattern for the vehicle, yet they still need to estimate take rates for equipment and variants so that the correct parts can be ordered. Audi developed its scalable Cold Start Planner, or CSP, to support this task with production and planning data.
Historically, such estimates relied heavily on experience. The CSP adds signals from the first customer orders as well as other indicators around the launch. According to Annette Schermer, who is responsible for Order-to-Delivery within the supply-chain organisation, Audi can also take account of dealer launch events and responses on social media.
For a component such as a wheel, the system can estimate which variants are likely to be ordered in particular markets and models. This does not eliminate uncertainty, but it provides a broader evidence base for procurement and capacity planning where historical data are limited.
The same approach helps when planners need to evaluate alternatives. If assumptions change or a constraint emerges, the AI-supported process can identify options and assess their effects more quickly. Schermer says this reduces the manual effort involved in alternative planning and makes it easier to think in scenarios before a decision is taken.
Another application addresses the scheduling of customer orders into a production week. Such schedules have to balance requirements from manufacturing and suppliers. If a particular part number or supplier becomes constrained, planners need to identify which already scheduled vehicles are affected and what changes are possible.
AI can help analyse combinations that would be cumbersome to trace manually, particularly when relevant components sit several levels down within other assemblies and have to comply with market-specific configuration rules. Schermer says a stronger data foundation allows teams to spend less time assembling information and more time solving the actual problem.
Why does better AI expose data problems?
The programme also creates a feedback loop: using data more intensively for decisions can reveal where the underlying information is weak. Audi has encountered AI-generated optimisation proposals that remain implausible even after closer analysis. In some cases, Braun says, the cause turns out to be obsolete or inconsistent source data.
This means data quality becomes part of the AI programme itself. If a model is expected to influence planning decisions across multiple functions, errors and inconsistencies that might once have remained local become visible across the chain. Audi then has to correct the underlying data and establish processes that keep the information consistent.
Braun sees cross-departmental transparency as another prerequisite. Data that remain hidden inside one function cannot contribute to an end-to-end optimisation problem. The issue is not merely access: different departments may use the same everyday term while attaching different operational meanings to it in purchasing, sales, production or finance.
Removing those inconsistencies requires detailed communication as well as technical integration. For Audi, the task is therefore as much about shared definitions and organisational discipline as it is about models or algorithms. Braun summarises the principle simply: work on and with the data is a basic requirement for AI to function at all.
Where does Audi keep humans in the loop?
Despite the emphasis on automation, Audi does not intend to translate major AI recommendations directly into action without human review. Braun says the company’s position is that the final decision remains with a person. The capability to challenge an AI result critically therefore has to develop alongside the technology itself.
That principle matters particularly when a recommendation affects several stages of a customer order, production sequence or supply-chain decision. AI can accelerate analysis without removing human responsibility, which means users need enough process knowledge to recognise when a plausible output does not make operational sense.
Audi applies a similar logic to generative AI. Answers from chatbots and GenAI tools are accompanied by sources that link back to the original documents, giving employees a route to check the basis of a response rather than treating generated text as self-validating.
At the same time, Audi wants broader parts of the workforce to gain practical exposure to AI through tools such as Microsoft Copilot. Reinisch links this to internal AI networks and citizen-development approaches, in which employees with deep process knowledge can contribute more directly to transformation projects.
The broader lesson from Order-to-Delivery is that the highest AI leverage may not lie in one spectacular use case. Audi is instead testing whether combining specialised tools, reusable data products and human process expertise can improve decisions across an entire value stream. The challenge is both technical and organisational: systems, definitions, responsibilities and judgement all have to align before AI can optimise the process as a whole.
