What does it actually take to put AI agents in charge of a company's go-to-market data — not just reporting on it, but acting on it?
Subscribe to AI That Works — a new episode drops every Friday. Follow the hosts too: Vaibhav (https://x.com/vaibcode)) and Dex (https://x.com/dexhorthy))
In this episode of AI That Works, Vaibhav and Dex are joined by John Kutay, Head of Growth Engineering at Rippling, for a hands-on look at GrowthOS — the internal AI system Rippling's growth engineering team built to give sales, marketing, and RevOps a single place to ask questions of their data and have AI agents act on the answers.
John walks through how Rippling got there: starting with deterministic, human-written SQL queries wired up as tools for an AI agent (retrieval as a tool), then moving into text-to-SQL, where the agent writes its own queries. That shift created real engineering problems — agent-generated SQL put unpredictable load on the data warehouse, so Rippling added SQLGlot, an open-source SQL parser, to validate and catch expensive or malformed queries before they run.
The conversation also covers Rippling's early testing of Jev, the fast "System One" decision model from TypeSafe AI that Vaibhav and Dex covered in a recent episode. John found it worked well for structured classification tasks, but underperformed on open-ended reasoning about whether a complex answer was actually correct — a concrete, warts-and-all data point on where a fast decision model is and isn't the right tool.
From there, the discussion turns to a broader industry shift: instead of a sprawl of specialized martech tools, companies like Rippling and Ramp are converging on one internal "super app" that a single team owns and continuously improves — the same pattern Meta's own Muse is chasing for consumer AI. John also shares how Rippling evaluates whether any of this is actually working: internal benchmarks running as part of CI, balanced against real user feedback from thumbs up/down ratings and Slack complaints.
GUEST BIO & LINKS
About John Kutay: Head of Growth Engineering at Rippling, where he leads GrowthOS, the company's AI-native go-to-market data and agent platform. He has over a decade of experience in enterprise data and AI pipelines, previously leading product and engineering at Striim, an enterprise data-streaming company.
linkedin.com/in/johnkutay | medium.com/@john.kutay
KEY TAKEAWAYS
• AI-native growth engineering turns dashboards into AI agents that can act on data, not just report on it.
• Retrieval-as-a-tool — deterministic, human-written SQL wired into an agent's context — was Rippling's safer first step before letting agents write their own SQL.
• AI-generated SQL needs real validation; Rippling added SQLGlot to catch expensive or malformed queries before an agent's query runs.
• Jev performed well on structured classification but struggled with open-ended reasoning about whether a complex answer was correct.
• Rippling checks AI output against both internal benchmarks and real user feedback (thumbs up/down, Slack complaints) — neither signal is trusted alone.
TIMESTAMPS
Episode Overview
Guest Intro: John Kutay of Rippling
Why Rippling Built an AI-Native Growth Engineering Team
From Dashboards to Prescriptive AI Decisions
Why "Chatting With Data" Is the New First-Class Analytics
Build vs. Buy: Should You Build Your Own Data Agent?
Whiteboarding Rippling's Old Data Pipeline
Rolling Out AI Agents on Top of an Existing Warehouse
Retrieval as a Tool vs. Text-to-SQL
How AI Agents Strained the Data Warehouse
Validating AI-Generated SQL With SQLGlot
Testing Jev as a Faster Decision Model
Where Jev Worked — and Where It Didn't
Building Rippling's "GTM Super App"
How Rippling Thinks About AI Hallucinations
RESOURCES / PEOPLE MENTIONED
• GrowthOS — Rippling's AI-native go-to-market platform, built with Databricks Genie as its conversational data layer. databricks.com/customers/rippling/genie
• SQLGlot — the open-source SQL parser Rippling uses to validate agent-generated queries. github.com/tobymao/sqlglot
• Jev — the TypeSafe AI "System One" decision model John tested against Rippling's workflows, covered in AI That Works Episode #75, "All About Jev."
• Meta's Muse — the consumer AI "super app" referenced in the conversation (misheard in the raw transcript as "Sync Muse").
TOPICS COVERED
GTM engineering
Growth engineering
AI agents
Go-to-market data
Data warehousing
Retrieval as a tool
Text-to-SQL
SQL validation
AI decision models
System One models
AI classification
AI benchmarks
AI evals
AI hallucinations
Internal AI platforms
Martech consolidation
RevOps
Enterprise data pipelines
AI-native software
Agent architecture
HASHTAGS
#AIThatWorks #GTMEngineering #AIAgents #GoToMarket #AIEngineering
Transcript
I've watched this one closely. Ask me anything about it and I'll point you to the exact moment.