Snowflake’s AI Bots inherit each employee’s data permissions

The Bots sit on top of two products Snowflake already sells. CoCo is its coding agent for data teams. CoWork is a personal work agent for knowledge workers.

The detail that deserves a data platform owner's attention is where the Bots get their authority. If an agent works under the permissions of the person it acts for, a company's existing role grants decide what a tireless piece of software can reach.

Agents that borrow an employee's access

Inheriting permissions is a sensible default because it reuses controls companies already audit. What changes is how often those controls get exercised. Access granted to a person who rarely uses it becomes access a background agent can use on every scheduled run. An agent that inherits a user's exact permissions is only as well-scoped as that user's last access review.

One official definition of revenue

Snowflake's own release names the failure it wants to prevent: an agent that doesn't know how a company defines revenue can give a confident wrong answer that someone then has to catch. A companion announcement the same day, also distributed through Business Wire, goes after exactly that.

The problem Snowflake describes will be familiar to anyone who has sat through a metrics meeting.

Joe Tobey, head of data product engineering at dentsu, said in the announcement that once agents become the way people query data, definitions can't stay in a static data dictionary; the ontology gives each one an owner and links it to the data that implements it.

Inherited permissions settle what an agent may touch. The ontology settles what the numbers it touches mean.

Novo's rollout at 900 terabytes

Loïc Giraud, vice president of data and AI foundation at pharmaceutical company Novo, inherited a Snowflake setup when he joined in 2024, used by affiliates in 80 countries for sales, marketing, finance and procurement. His first move was organizational: he folded the Snowflake teams and technology into one global function, then integrated more datasets before bringing in CoCo.

Novo now holds more than 900TB in Snowflake, and around 25,000 people across its affiliates use the platform directly or indirectly, through a front end, tabular output or a chat engine. Novo is also exploring CoCo, which its data engineers use to build new connections to services. Giraud said the company has created some agents that build data pipelines, an approach he said can reduce the number of engineers it needs focused on producing products and apps.

The order of operations is the transferable part. Platform first, richer data second, agents third, with Giraud insisting that guardrails and controls be in place before end users build their own solutions.

Novo also marks where a single platform's boundary stops. Giraud said data processed in Snowflake can be pushed out to ChatGPT, Snowflake services or anything else. Even if a Bot's inherited permissions govern work inside Snowflake, agents running elsewhere need controls of their own, which is the territory Cortex AI Gateway is pitched to cover and the claim buyers with mixed agent estates will want to test first.

Claudeforce takes the other road

The designs point in different directions. Claudeforce, from Salesforce and Anthropic, brings a system of record's data and rules into the assistant where the seller already works. Both promise governed action. For a CIO whose customer briefings draw on CRM records and warehouse data at once, the sharper question is which platform's permissions decide what the agent sees when the two systems disagree.