AI Agents and the Fight for Customer Data

The central tension in this episode is between AI agents' need for consolidated business data and the growing willingness of SaaS vendors to restrict API access. SAP announced a policy banning all AI agent access except where specifically approved, though Fivetran CEO George Frazier noted that policy memo does not override existing customer contracts. Salesforce has historically been permissive but has more recently begun tightening access. Frazier characterized the data being locked down as context for agents rather than training data, and argued that very few companies train their own models, making vendor lockdowns harmful to customers without a coherent strategic justification.

Frazier's core argument is that AI agents require the same data foundations as traditional business intelligence, because data is always born in systems of record and cross-system questions require consolidation. He compared using AI without access to business data to using ChatGPT before it was connected to the internet. The agent identity problem compounds the access problem: agents need role-based identities rather than individual user identities, meaning a single agent could do the work of hundreds or thousands of human users, making it difficult for vendors to distinguish agentic access from traditional access. Andreessen Horowitz is reportedly considering onboarding AI agents like human employees, assigning document access and adding them to Slack teams, which Frazier argued means more software seats and more consumption, not less.

Frazier dismissed the idea that AI eliminates SaaS demand, pointing out that OpenAI and Anthropic are both Fivetran customers using it to replicate data from their own SaaS tools. Major AI labs including Andreessen Horowitz still use SaaS tools four years into the AI era. Software costs represent only five to ten percent of headcount spend at typical companies, and the more plausible AI use case is improving core business operations rather than shaving software spend from five percent to four and a half percent. Fivetran's own business data shows acceleration rather than slowdown, and Frazier said net dollar retention falling below one would be the data signal confirming a SaaS collapse thesis, and that signal has not appeared.

To push back on vendor lockdowns, Fivetran created opendatainfrastructure.com, which scores vendors on data access policies including egress charges, data completeness restrictions, and terms of use limitations on a one to three scale. Model contract language guaranteeing data access is available on the site and recommended for inclusion in master service agreements, though Frazier said this is only worth pursuing for large contracts in the five hundred thousand to one million dollar range. He argued vendors restrict access simply because customers do not push back, and that requesting data access language in an MSA sends a meaningful signal even when the clause is not ultimately granted.

Frazier called data gravity, defined as egress charges making it prohibitively expensive to move large business data sets, essentially fake. Fivetran serves seven thousand significant customers yet moves surprisingly small data volumes because change data capture replicates only incremental changes. He attributed the perception of data gravity to naive pipelines that copied entire data sets nightly, causing read amplification that exaggerated apparent volumes.

Fivetran's merger with dbt Labs and acquisitions of Census, Tobiko Data, and SQL Mesh beginning in 2025 represent a strategic shift for a company that previously described itself as non-acquisitive. Frazier framed the dbt merger as obvious because Fivetran moves data into a central location while dbt organizes it into models encoding business rules, and the two products have historically been deployed together. Coding agents are already writing large numbers of dbt models, making dbt a direct beneficiary of the AI coding wave. Frazier acknowledged that coding agents capable of reliably writing connectors would threaten Fivetran's core replication business, but said the company intends to provide the DIY connector-building tools customers would use in that scenario even at short-term revenue cost.

On agent interfaces, Frazier predicted that within five years the majority of agent usage will go through the same APIs and interfaces humans use rather than novel agent-specific interfaces. He acknowledged MCP servers solve practical problems around authentication, authorization, and discoverability, and that tooling on the consuming side has grown around MCP causing it to gain traction in practice. He also argued the right data foundation for AI is simply whatever a company already has, whether Snowflake, Databricks, BigQuery, or an Iceberg data lake, and that AI is currently creating more demand for infrastructure rather than commoditizing it, with the highest abstraction layers most exposed to displacement by agents.