In brief
Becoming data-driven takes more than spending on big data and AI. Tanvi Singh argues that banks need a shared data platform built on a data mesh, in which each business area owns its data as a product. Her case is that central data teams become bottlenecks, and that vendor lock-in and technology-first projects waste money.
The central tension is between a long-term cultural and architectural shift and the near-term results sponsors demand. Singh's answer is to start from specific use cases, such as spotting unauthorized transactions, while building standards that scale. She also stresses planning for the full model lifecycle, including shutting models down.
Questions
What is a data mesh and why do banks want one?
Singh describes it as a distributed setup in which each business domain owns and publishes its own data as a product, instead of everything going through one central team. She says this avoids the bottlenecks and repeated work that central teams create.
Why do companies spend heavily on AI but still lack a data culture?
Singh cites a NewVantage Partners survey in which 99% of firms invested in big data and AI but under 25% reported a data culture. She says the problem lies in how the work is executed and in how data-literate the organization actually is.
How should a bank start without wasting money on technology?
Singh says to pick a specific problem, such as finding unauthorized transactions, and test it in a sandbox using near-real data. Governance and standards should be set up so they can be reused for the next use cases.
What happens to AI models after they go live?
Singh says they need MLOps, meaning operational upkeep so pipelines don't break, and eventually a planned shutdown. She notes that few organizations plan for shutdown, so costly pipelines keep running with no value.
Should a small company build a data mesh too?
Singh says no, not necessarily. For small and medium-sized firms she says a lakehouse from a single provider, which combines a data warehouse and data lake, can deliver much of the value.
Facts
Stated by the speakers. Each quote matches the transcript, which shows what was said, not that it is true.
Tanvi Singh spent nearly eight years at Credit Suisse, most recently as Managing Director of Digital Transformation and Product Labs.
“At the time of this recording, she was wrapping up nearly eight years serving with Credit Suisse, most recently as Managing Director of Digital Transformation and Product Labs.”
0:00:05 · Daniel Faggella · Guest background
Tanvi Singh is now group head of the ADA Foundations Hub at UBS.
“She is now the group head of the ADA Foundations Hub at UBS.”
0:00:05 · Daniel Faggella · Guest background
Tanvi Singh previously worked on analytics at General Electric and Monsanto.
“She previously worked with giants like General Electric and Monsanto, also on the analytics side of things.”
0:00:37 · Daniel Faggella · Guest background
Tanvi Singh earned a master's degree in software systems at the University of Zurich.
“And before that, received her master's in software systems at the University of Zurich.”
0:00:37 · Daniel Faggella · Education
NewVantage Partners published a report surveying about 500 mostly medium-sized and large organizations about data culture.
“So very recently, New Vantage partners pulled out a report for some 500 different companies, organizations, most of them medium-sized, big-sized, have participated.”
0:03:05 · Tanvi Singh · Data culture survey
In the NewVantage Partners survey, fewer than 25% of firms said they had a data culture, while 99% were investing in big data and AI, according to Singh.
“And the surprising part of the story was that about less than 25 % would say that they have data culture, while 99 % of them are investing in big data and AI.”
0:03:05 · Tanvi Singh · Data culture survey
Singh says firms ran separate data warehouse and data lake platforms, and that data warehouses have existed for decades.
“in our world that used to be data warehouse and the data lakes. So the data warehouse infrastructure was never a very interesting part of the conversation because it's been there for decades.”
0:04:45 · Tanvi Singh · Data warehouses
Singh defines a data platform as one that treats data as a product rather than as a service or an asset kept by individual teams.
“A platform per se is then the right definition of creating data as a product and not data as a service or data as an asset which people tend to keep to themselves”
0:05:41 · Tanvi Singh · Data products
Singh gives the example of a client data product that bundles the data, its metadata, compute, processes and APIs under one domain owner.
“So if you are responsible for the client division in an organization, then you're talking about a client as a data product where the data itself, the metadata supporting it, the compute of it,”
0:06:10 · Tanvi Singh · Data products
Singh says small and medium-sized organizations can adopt a lakehouse, which combines warehouse and data lake functions, instead of a distributed data mesh.
“For a small and medium-sized organization, it's pretty much okay to move into a lake house solution, which is a very technical way of solving the problem”
0:08:44 · Tanvi Singh · Lakehouse
Singh names Databricks and Snowflake as vendors that have tried to offer both warehouse and data lake capabilities.
“A lot of cloud vendors out there, Databricks I can name, Snowflake I can name, have been trying to address being a warehouse provider, trying to be a data lake provider”
0:08:44 · Tanvi Singh · Lakehouse vendors
Singh says European organizations face GDPR challenges when choosing vendors that are not based in Europe.
“Most of you who are in Europe would also have a lot of GDPR set of challenges to deal with.”
0:11:28 · Tanvi Singh · GDPR
Singh says some organizations invested mainly in Hadoop infrastructure and data ingestion, leaving sponsors unhappy because they saw no output.
“They think of Hadoop as the newer technology. So it put most of their investment into setting up the infrastructure, pulling in the data, which makes the sponsors really, really unhappy”
0:11:54 · Tanvi Singh · Hadoop
Singh predicts SQL-based tools for semi-structured data, comparable to those for structured data today.
“What I feel in the future would be coming in would be a very SQL-based set of solutions for semi-structured data as we see it for structured data today.”
0:10:37 · Tanvi Singh · Future technology
Singh says a financial crime use case such as finding unauthorized transactions should be tested on pre-production data rather than synthetic data.
“You're looking into the transactional data. The most important thing to do before you put in any of the regression model or think about using any of the deep learning model is to just have the pre-production data”
0:14:50 · Tanvi Singh · Financial crime
Singh describes MLOps as operationalizing machine learning so pipelines don't break once a model is in production.
“The third area of most of us in the AI world do not play a lot of focus on is the MLOps, which is being able to operationalize your ML.”
0:16:34 · Tanvi Singh · MLOps
Singh says very few organizations plan for the retirement phase of a model's lifecycle.
“Even more important to talk about is the phase four, which I've seen very few organizations even having on their roadmap to say is what happens when a model has gone through its lifecycle?”
0:16:59 · Tanvi Singh · Model lifecycle
Singh says she has seen organizations run costly cloud pipelines that produce no valuable output.
“So I've seen organizations running expensive pipelines on the cloud, which doesn't even have a valuable output.”
0:17:26 · Tanvi Singh · Cloud costs
Singh says the sandbox phase mainly involves subject matter experts and machine learning engineers, while later phases involve data engineers and operational teams.
“For Sandbox, you usually have a lot of very excited set of SMEs together with the machine learning engineers trying to experiment everything another way, right?”
0:19:00 · Tanvi Singh · Team roles
Singh argues that centralized data teams become bottlenecks and that decentralizing ownership is key to a data-driven culture.
“I think the decentralization of power, the centralized teams do not work. They just do not work because they kind of become a bottleneck.”
0:22:07 · Tanvi Singh · Data mesh
Singh predicts that over the next three to five years the data vendor landscape will simplify into more standard products.
“a lot of the data landscape or the technology landscape, we see that a lot of different vendors is going to simplify.”
0:24:35 · Tanvi Singh · Vendor landscape
Emerj offers an AI and financial services executive brief at emerj.com/fin1.
“download our AI and financial services executive brief. You can go to emerj.com slash fin1.”
0:25:37 · Daniel Faggella · Emerj resources