Data democratization has been a buzz phrase for years, but Chris Gifford, chief data officer at American Express, argues that it’s much more helpful to think about democratizing analytics. Making analytics more accessible enables employees, as well as AI agents, across the business to generate insights and act upon them within a governed framework.
“It’s not just capture the data, and then stack and rack the data,” he says. “It’s how do we accelerate analytics and then time-to-market insights into action.”
Gifford, who’s spent his entire career in the highly regulated financial services industry, says many people in businesses got the wrong idea about data democratization, thinking it meant they should be given the ability to pull data out of controlled environments and use the data to make decisions without IT general controls. That sort of thinking creates operational, data, and privacy risks.
“Stop pulling data out of our environments and I’ll give you all the controls,” he says. “I’ll set up rapid expansion of data elements and new data sources, and give you tools that allow you to do cutting-edge analytics, even. Your personal analytics agent will soon run analyses.”
Assessing uncertainties
Beyond the risks, Gifford says allowing employees access to raw data, and potentially the ability to move and store it outside the bounds of governance controls and muddying data lineage, creates inefficiencies that make it much more difficult to leverage gen AI and agentic AI.
“People are realizing the way we’ve managed, exposed, and distributed data might not be the best way to do it going forward, especially if we’re going to enable agents,” Gifford says. “If you’ve got multiple copies of data with different use case screenings, or filters on the data for different folks who’ve landed the data into your lakehouse architecture, an agent isn’t going to know that.”
As a result, democratizing analytics at American Express is largely focused on staging and deploying data for consumption, figuring out how to go beyond lakehouse architecture and make it easily consumable by humans, APIs, and agents.
Gifford’s team uses AI to crawl the company’s data, use cases, redundancy, obsolescence, and triviality, and then assess the golden sources of these data to point future agents toward.
“We’re also leaning more into having gen AI make recommendations on business metadata,” he says. “It can troll through all sorts of structured and unstructured data environments to help figure it out.”
Beyond staging and deployment, the company has begun piloting and testing “talk to my data” capabilities, which allow users to ask analytics questions using gen AI.
“Within a very confined environment with golden source, guardrails, and other governance controls applied, we’ve turned ‘talk to my data’ on and started experimenting with it,” Gifford says. “We’ve learned it’s not enough to take your data from current or historical stores into a modern platform, and then apply governance and controls to it.”
In early experiments, different testers would get different answers to the same questions, and hallucinations were an issue.
“Those kinds of things can be controlled better by adding metadata, ontology, semantic layers, knowledge graphs, master data management, and reference data management,” Gifford says. “Heavy investments into that are only going to repay in the speed, accuracy, and uptake of these AI capabilities in the analytics arena.”
Company recalibration
American Express takes a risk-based approach to data management, applying tiered security, governance, and compliance controls based on data sensitivity and organizational risk. As the company moves toward leveraging AI for its analytics, Gifford says a new data management layer that grades data’s readiness for gen AI and agentic AI is necessary.
“That layer needs to say the quality of the data and the governance controls on the data have to be at a minimum amount before we’re willing to unleash gen AI on it,” he says.