AI Agents and the Fight for Customer Data
Fivetran CEO George Frazier joins a16z's Martin Casado to discuss the evolving data infrastructure landscape, the threat of SaaS vendors locking down API access in response to AI agents, and the Fivetran-dbt merger. The conversation covers why centralized data remains critical for AI agents, the overhyped 'SaaSpocalypse,' and Frazier's contrarian views on data gravity and Postgres.
Summary
The conversation opens with Fivetran CEO George Frazier explaining the company's core mission: replicating data from systems of record like Salesforce, NetSuite, and SAP into centralized data lakes. Originally built for business intelligence and reporting, Frazier argues this same data infrastructure is now essential for AI agents, which require broad business context to function effectively — analogizing data-starved agents to the pre-internet version of ChatGPT that couldn't answer questions beyond its training cutoff.
A significant portion of the discussion focuses on SaaS vendors responding to AI threats by locking down their APIs. Frazier highlights SAP's recent policy banning AI agent access except where explicitly approved, and notes Salesforce has become increasingly restrictive as well. Frazier argues this is bad for customers, historically misguided (drawing parallels to API debates in the 1990s), and ultimately self-defeating since customers will find workarounds. He recommends CIOs insist on data access rights in MSA contract language and points to opendatainfrastructure.com, a Fivetran-created benchmark scoring vendors on data openness.
Frazier dismisses the 'SaaSpocalypse' narrative — the idea that AI agents will replace SaaS entirely — arguing the real threat to incumbents is AI-native startups building better competitive products faster, not customers vibe-coding replacements. He notes that software spend is typically only 5-10% of headcount costs, making seat reduction an inefficient use of AI. He also argues that the agent-as-human-employee model (giving agents their own email, Slack presence, and identity) is valuable because it slots into existing workflows without requiring full system refactors.
Frazier offers two notable contrarian takes: first, that 'data gravity' — the idea that egress costs make centralizing data prohibitively expensive — is 'completely fake,' attributable to poorly designed pipelines that copy entire datasets rather than using change data capture. Second, he argues Postgres is fundamentally a bad database due to its age and technical debt, going so far as to say undergraduate database class projects produce better storage engines. He expresses genuine interest in building a new operational database backed by S3, similar to SQLite but for cloud-native AI workflows.
On the Fivetran-dbt merger, Frazier describes it as a natural fit since the two products have historically been used together — Fivetran ingests data and dbt transforms it into business models. He also argues dbt will be a major beneficiary of coding agents, as AI will write vast amounts of dbt models, and the SQL artifact serves as executable documentation of business logic. The conversation closes with Frazier identifying DIY connector-building via coding agents as his primary existential concern for Fivetran, while viewing AI's expanded demand for data infrastructure as the biggest opportunity.
Key Insights
- Frazier argues that AI agents require the same centralized data foundations built for business intelligence, meaning companies that already have modern data platforms (Snowflake, Databricks, BigQuery) don't need exotic new systems for AI context.
- Frazier contends that SaaS vendors locking down APIs is historically cyclical and self-defeating, drawing a direct parallel to identical debates in the 1990s that ultimately resolved toward greater openness — and predicts the same outcome this time.
- Frazier claims data gravity is 'completely fake,' arguing the myth arose from poorly engineered pipelines that copied entire datasets nightly rather than using change data capture, which makes actual data movement volumes far smaller than assumed.
- Frazier argues the bigger threat to SaaS incumbents is not agents replacing software but AI-native startups rapidly catching up to and surpassing established players by building better products faster with AI coding tools.
- Frazier observes that AI agents don't need granular user identities the way humans do — a single role or identity can do the work of hundreds of people — which undermines the 'more agents equals more seats' argument against the SaaSpocalypse thesis.
- Frazier states that both OpenAI and Anthropic are Fivetran customers whose internal data platforms look nearly identical to those of traditional enterprises, using the same Fivetran/dbt stack consultants deploy elsewhere.
- Frazier argues Postgres is fundamentally a poor database not because its creators lacked skill but because of accumulated technical debt from age, claiming undergraduate database course projects produce better storage engines than Postgres's heap storage engine.
- Frazier identifies Fivetran's primary existential risk as coding agents eventually becoming capable enough that customers simply vibe-code their own data connectors, and says Fivetran will pivot to providing the tooling for that use case even at short-term revenue cost.
Topics
Full transcript available for MurmurCast members
Sign Up to Access