Kargo Keeps Ad Budgets Reaching Real People, Not Bots, With Snowflake CoCo

For the commuter on the subway, scrolling is a way to pass three stops. But for the brand paying to appear in that scroll, the budget gets used up by the second. The campaign deadline will not move, and neither will the question underneath it: Is that money reaching real people, in the right places?

Kargo exists to get the calculus right. It runs the marketplace behind many of these campaigns across connected television, social, retail media and the open web. Its offering includes building the creative and pairing it with predictive models in an approach it calls Creative Science. 

If something goes wrong, a different clock starts. Kargo’s data scientists and service desk engineers need to find the needle in the haystack buried in thousands of tables across 9 petabytes of data — a log feed that grows by a terabyte a day.

Kargo’s data and AI team set out to shorten that hunt. Developer productivity came first: response times, incident response protocols and bug-fix timelines. But other AI coding tools carried no context about the business, its data or its platforms. And an AI tool that writes a query but does not know what the data means only moves the problem somewhere else.

Kargo turned to Snowflake CoCo to close that gap and bring its own business context into its full AI development lifecycle, a different animal from traditional software. The first target was model deployment. Kargo had no formal machine learning engineering team, and the work sat distributed across multiple teams.

Early on, Kargo made the strategic decision to deliberately start Snowflake CoCo adoption from the bottom up. One session with Kargo’s Snowflake team was followed by a week of lunches devoting an hour a day, every day, to CoCo. Individual contributors worked real bugs with no VP in the room. From there, use spread to data engineering, analysts and then completely past the data organization to service line teams across a company of 600 global employees.

The shape of the work changed with it. A year before CoCo, good self-serve answers were hard to come by and teams leaned on their code editors. Now the interaction is almost entirely conversational, and Kargo extended read-only access to the correct tables to effectively everyone in the company. Onboarding went to CoCo instead of a training program. Outside engineering, people who had been analytics consumers rather than Snowflake users now work through CoCo in Snowsight.

Kargo’s investment in empowering its nontechnical employees with AI has already paid off. During one bot traffic incident, someone outside product and engineering used CoCo to build a model that flags unwanted traffic. About a week of refactoring later, it ran in Kargo’s live auction at submillisecond scale.

Kargo now also ships model versions at a rate it could not reach before. A large auction refactor the year before CoCo yielded two small model versions, while this year Kargo shipped three major ones in a matter of months, with dozens planned on an even smaller number of developers.

The mix of people in Snowflake changed too. More of Kargo’s site reliability and DevOps engineers are in the platform, keeping security and governance in place as usage spreads. They had rarely logged in before. Now they use it several times a week, if not daily. 

Next, Kargo plans to put CoCo deeper into its Snowflake Cortex AI inference work. In the meantime, whether it’s auctions, ad tags or production applications, Kargo’s around-the-clock service team opens the logs of any flow and finds the cause while the campaign is still running. For the commuter enjoying the ride, nothing changes. For the brand investing precious resources in a creative campaign, that is the whole difference.