Data Flywheels: Building AI Advantage From Proprietary Data

COLLIDE Data + AI Conference 2024What are Data Flywheels? And how do you build one?Watch the full session · 28 min

At COLLIDE 2024 in Atlanta, Yali Sassoon, Co-Founder and CTO of Snowplow, argued that the durable advantage in AI comes from data flywheels: feedback loops in which an AI application generates the proprietary data that makes it better. In a session titled "What are Data Flywheels? And how do you build one?", sponsored by Snowplow, he explained how they work and listed five ingredients for building one. COLLIDE was held October 10-11, 2024, at the Sandy Springs Performing Arts Center.

What is a data flywheel?

Sassoon said he picked up the term at the Snowflake and Databricks summits earlier in 2024, where NVIDIA CEO Jensen Huang talked about it extensively. Huang's point about an AI customer service chatbot, as Sassoon recounted it, was that the value is not only automation: "you want to capture the engagement in the flywheel," because it creates more data.

Sassoon's example was Klarna, which he described as an early adopter of large language models. He said the company built a support agent that replaced the work of 700 human agents, speaks more than 35 languages and resolves issues in about two minutes versus eleven, with the same customer satisfaction and a quoted saving of $40 million a year. These figures were Sassoon's account of Klarna's reported results, not something he verified.

The flywheel part, he said, is that the agent is trained on proprietary data, such as documentation and past support chats. Every new conversation produces more of that data, which feeds back in and improves the agent, in a "positive feedback loop."

Why do flywheels create a competitive advantage?

Sassoon said flywheels apply wherever there is an agentic application or a machine learning process. He cited Huang's examples of chip design, bug databases, new products and supply chain. NVIDIA, he said, has more proprietary GPU design data than anyone and uses AI to design chips, so even a newcomer with the best AI would struggle without equivalent data.

He offered an analogy from military strategist John Boyd's OODA loop (observe, orient, decide, act). U.S. pilots flying F-86 Sabers in the Korean War, he said, outperformed technically superior Russian planes because better visibility and hydraulic controls let them cycle through the loop faster. A data flywheel does the same for a business, he argued.

What are the five ingredients of a data flywheel?

Sassoon said each could be a talk on its own. His list:

1. Proprietary data. "The only place left to differentiate in the AI stack is in the data," he said, because most companies use the same handful of models.
2. Real-time architecture. The faster data is computed on and acted on, the faster the loop turns and the harder it is for competitors to catch up. He said tools for real-time processing have improved quickly since Snowplow began in 2012.
3. Feature pipelines. Raw and unstructured data has to be turned into features, such as categorizing past emails by humor, voice, length and social proof to generate a personalized email.
4. Data quality and governance. Semantics, quality and permissions should travel with the data. "Data has to be actively manufactured," he said, so these controls are built into the process.
5. Capturing decisions. Organizations should record the actions and decisions of both humans and AI.

Why track AI decisions as well as user behavior?

Sassoon said the last item is the one most organizations get wrong. Web tracking assumed users moved through a fixed landscape, but personalization engines now change the experience.

He walked through a snowboard shop example. A visitor leaves the site, a model scores purchase propensity and picks likely boards, a service composes a personalized email, and the email is sent, opened and clicked. Each step should be tracked, he said, including the full text of the email, which is otherwise ephemeral, the facts about the customer behind the prediction and facts about the model. That lets people compare models and lets the AI learn from outcomes.

Does this work in practice?

Sassoon said Snowplow customers in retail and media have moved from a handful of algorithms to more than 100. In a media example he described, an organization began with an algorithm to measure the monetary value of content, then built one to measure placement impact, then moved on to paywall decisions and dynamic subscription pricing. He said it saw higher subscription rates and engagement within a year. These are his characterization of customer results.

Where can enterprise leaders continue the conversation?

COLLIDE 2027, the largest AI conference in the Southern U.S., will bring 1,000+ enterprise data and AI leaders to the Sandy Springs Performing Arts Center in Atlanta on Wednesday, October 13, 2027. Details are on the COLLIDE event page, and registration is coming soon.

Frequently asked questions

What is a data flywheel?

A positive feedback loop in which an AI application generates proprietary data that is fed back to improve the application.

How does Klarna's support agent illustrate a data flywheel?

Sassoon said it was trained on proprietary support data, and each conversation creates more data that improves the agent.

What do you need to build a data flywheel?

Sassoon listed proprietary data, real-time architecture, feature pipelines, data quality and governance, and capture of human and AI decisions.

Why track AI decisions?

He said personalization engines change the customer journey, so recording AI decisions lets teams analyze and improve them.

This article is based on remarks by Yali Sassoon, Co-Founder & CTO, Snowplow at COLLIDE Data + AI Conference 2024 on October 10, 2024. More coverage: COLLIDE AI Conference.

Filed under: AI Strategy & ROI, Data & AI Infrastructure, Observability & Data Quality