“AI or die” isn’t hyperbole for Ramp’s data team
Last week, Ricky Meyers, Analytics Engineering Manager at Ramp, shared how Ramp's data team uses AI agents to maximize impact.
Ricky opened by quoting William Petrie, Ramp’s CFO, who says, “Product wants data to tell them what to do – more and faster.” In other words, data teams shouldn’t just report. They should guide: Incrementally, measurably, and daily.
AI is no longer optional for data teams that take this mission seriously. Ricky told us, “ To make great AI products, you need to use great AI products.” That’s why Hex and the Hex Notebook Agent are a key part of Ramp’s stack.
Watch the video above to see the entire event or catch up with the highlights below.
Ramp didn’t want its users worrying about which fund their business cards were pulling from every time they swiped. To provide a truly modern experience, they developed an auto-routing feature that uses context from previous transactions to automatically select the right funds.
It’s an ambitious feature, and Ramp’s data team turned to Hex to help them track its performance. Here are the ways that Hex's AI capabilities helped accelerate the reporting of the auto-routing feature.
With the Hex Notebook Agent, Ricky and his team can build analysis frameworks and visualize their results within minutes.
Ricky’s team starts by providing the Notebook Agent with background information on Ramp, descriptions of the business card’s specific functions and features, and context on the team’s data infrastructure. If the team is working on a particular feature, they can tag the data frame within a Hex Agent prompt to add feature-specific context.
The team then thinks through what it wants to track and what success means. The Notebook Agent works alongside them, planning the analysis and creating an entire roadmap from a blank page.
Within 90 seconds, the Notebook Agent can produce an analysis framework that compares metrics like accuracy rates and coverage rates. With a click, the Agent produces charts that visualize these metrics. The team can go back and forth as necessary, building dashboards rapidly, prompt by prompt.
Developing and improving new features isn’t an A-to-B process. Novel ideas need experimentation, exploration, and iteration – processes that the Hex Agent accelerates.
Ricky shared an example: He recently hypothesized that there might be a strong negative correlation between the number of different funds available and the accuracy rate of the auto-routing feature.
With the Hex Notebook Agent, Ricky could perform zero-shot chain-of-thought prompting to test his hypothesis without ever leaving Hex. “ The benefit of having something live in the notebook is that you're not jumping back and forth between tools,” Ricky says.
The agent thought it through and planned a step-by-step workflow, resulting in a plan and a Python script that executed the analysis. The time from idea to insight shrank to near immediacy.
For many data teams, capturing information and generating insight is the easy part; turning insight into action is when it gets hard. With the Hex Notebook Agent, this bottleneck disappears.
When Ricky has findings he wants to share, he can prompt the Hex Notebook Agent to synthesize insights and generate a report within minutes. To ensure everyone understands, he can prompt the Notebook Agent to assume the primary audience is product managers, for example, and focus the analysis on high-level results.
The report can include everything Ricky thinks his audience needs, including an executive summary, overview, insights, implications, and recommendations, which helps drive influence inside and outside the data team. “This is something I can literally just copy and paste to a PM or leadership in Slack,” Ricky says, “and tell them, ‘Hey, this is how things are going.’”
Key to Ramp’s success with AI is the semantic layer, and the effort the data team invests in ensuring that the most important information is well-documented.
Ricky’s team started by prioritizing the core 75%, the main jargon, lingo, and concepts that the company ran on. They reviewed the questions the data team had received over the years and identified the most common answers they could document.
Now, the Hex Notebook Agent starts every task already knowing:
Background information on Ramp, the company.
Context on product areas, such as corporate cards, reimbursement spend, and management automation.
Details on the team’s data infrastructure, including Snowflake, Marts and Agones metric schemas, and more.
Especially relevant, heavily used, or endorsed tables.
With a strong semantic layer, the Hex Notebook Agent can build analysis frameworks informed by company- and feature-level context, perform exploratory analysis without needing to repeatedly explain company concepts, and tailor findings to audiences the Agent is already familiar with.
“We supercharge Hex upfront before we even get into a notebook, so that we have the most success downstream,” Ricky explains.
Pro-tip from Caitlin Moorman, Analytics Engineer at Hex
"Our entire company had access to the Notebook Agent for about four months before we launched the public beta. To make the best use of it, we make sure that our docs are very high quality and very clear.
It also requires thinking through new systems: What is the right combination of metadata our docs files are creating within Snowflake? Rules files, semantic layers – what other kinds of contexts can you add? Can users drop a component into a project and work with the Agent to get additional scaffolding, deeper definitions, and starting points for analyzing this dataset?"
At the end of the session, viewers had the chance to ask questions. To see the rest of the questions and the answers in full, watch the live event above.
So far, Ramp’s data and engineering teams are the primary users of the Hex Notebook Agent.
The engineering team, which includes people who are technical but who are not experts in writing SQL, can use the Notebook Agent to build tracking quickly and entirely on their own.
Quasi-technical teams, which include people who can only write some basic queries, are learning that self-service is now possible. “ I see a lot of value here for quasi-technical people who can write a little bit of SQL but really don't know where to start for an analysis,” Ricky says. “They can just lean on the Hex Notebook Agent.”
Ricky says, “For every step that includes an LLM call, you need to have evals.” You can use a wide variety of tools or build a tool yourself, he adds, but they’re necessary.
Ricky and his team believe in governing by outcomes, and evals ensure that what should be happening actually is happening.
Ramp takes the words of Ravi Gupta, an investor in Ramp, as a strong influence: “AI or die.” That’s why Ricky says, “AI is now a baseline expectation across all areas at Ramp.”
Ramp’s data team knows, however, that the biggest results often come from small, consistent efforts made over time. “Our goal is to make Ramp, the business, and Ramp, the product, 1% better every day,” Ricky says. “AI and the Hex Notebook Agent really contribute to that goal.”