At Rise of AI 2022 in Berlin, Claudia Pohlink, chief expert for data at Deutsche Bahn, argues that AI progress is bottlenecked by data, not models. Inspired by Andrew Ng's push for 'data-centric AI', she contrasts research, where nearly 99 percent of papers focus on improving models and only about 1 percent on improving data, with practice, where up to 60-80 percent of effort in AI use cases goes into data preparation, sourcing and access negotiation. She cites a Gartner estimate that poor data quality costs organizations around 13 million dollars a year on average, and notes demand for data engineers now outpaces demand for data scientists. Pohlink describes Deutsche Bahn's Data Intelligence Center, built three years earlier around a House of Data, a House of AI and a House of 3C (community, change, communication), and the company's 2020 board resolution making data accessible company-wide by default unless explicitly restricted. She walks through the group's roadmap of a central data catalog, data governance roles and a planned move toward data mesh architecture, and closes by warning against skipping data work in favor of 'sexy' model-centric AI projects.
Key takeaways
- Pohlink cites that about 99 percent of AI research papers improve models while only about 1 percent focus on improving data.
- She says up to 60-80 percent of real-world AI project effort goes into finding, accessing and preparing data.
- Deutsche Bahn's 2020 board resolution made data accessible by default company-wide, reversing the usual permission model.
- Deutsche Bahn organizes its data work into three units: House of Data, House of AI and House of 3C.
- She warns that skipping data governance for 'sexy' model work risks projects being set back, citing a Gartner prediction for 2024.
Detailed notes
Why research and practice diverge on data
Pohlink frames current AI progress around 'ABC': algorithms, big data and computational power, but says data management itself has not improved much. She cites a striking gap: nearly 99 percent of recently published AI papers focus on improving performance by changing the model, versus roughly 1 percent that address the data. In practice, she says, up to 60-80 percent of effort in AI use cases goes into data preparation, a figure she traces to a 2014 New York Times report, and preparation includes finding data sources across a company and getting permission to use them, often taking six to twelve months in larger organizations.
The cost of poor data and the case for data-centric AI
She cites a Gartner estimate that poor data quality costs organizations an average of about 13 million dollars a year, and notes that demand for data engineers has outpaced demand for data scientists two to one, or by other estimates four or five to one. Referencing Andrew Ng's landing.ai work, she defines data-centric AI as 'AI is code plus data' and the discipline of systematically engineering the data used to build a system, illustrating it with a train-image labeling example where structured rules produce one consistent bounding box instead of many inconsistent ones.
Deutsche Bahn's Data Intelligence Center
Pohlink describes Deutsche Bahn's Data Intelligence Center, established three years earlier at headquarters and organized into three units: the House of Data, which she leads and which makes data visible and available; the House of AI, led by colleague Thomas Thiele, which enables machine learning use cases with business units; and the House of 3C (community, change, communication), responsible for making adoption sustainable. The three houses meet monthly or quarterly to share AI project learnings and data-access problems.
Governance timeline and the open-data resolution
She traces Deutsche Bahn's data journey: the Data Intelligence Center's founding, a 2020 move to a decentralized chief digital officer organization with data officers in each business unit, and new roles such as data steward and data owner. She highlights a 2020 board resolution making data accessible across the whole company by default, so that anyone needing to restrict specific sensitive data must justify the restriction, rather than the reverse. This year's roadmap centers on a central data catalog (with a new underlying technology she says no other data catalog supplier has changed before), data governance rules, and architecture work including data mesh.
Q&A highlights
Asked for her favorite career learning, Pohlink says her advice is continuous learning, describing her own path from studying data science into AI and later into quantum computing work at Deutsche Telekom. Asked about synthetic data, she says Deutsche Telekom began exploring converting real telecommunications data, including location and timestamp fields, into synthetic data around 2015-16 to avoid data-protection issues while preserving useful information, and calls it a likely future direction for exchanging data between departments.
Quotes
The challenge of AI in practice does not lie in the models or model architectures but predominantly in the data.
Claudia Pohlink
Data is for everyone.
Claudia Pohlink
Questions answered in this talk
What is data-centric AI?
According to Claudia Pohlink at Rise of AI 2022, data-centric AI is the discipline of systematically engineering the data used to build an AI system, shifting focus from changing models and algorithms to improving data consistency and quality, a concept she attributes to Andrew Ng.
How much of AI project effort goes into data work rather than modeling?
Claudia Pohlink states at Rise of AI 2022 that in practice up to 60-80 percent of the effort in AI use cases is spent on data preparation, sourcing and access, not on building or tuning models.
How did Deutsche Bahn change its approach to data access?
Pohlink explains at Rise of AI 2022 that Deutsche Bahn's board adopted a 2020 resolution making data accessible across the whole company by default, so that data owners must justify restricting access to sensitive data rather than employees having to justify requesting it.