Perplexity Computer Plugs Into Snowflake and Databricks

Anyone who has run a data platform knows the real bottleneck isn't compute, it's the analytics queue. Sales wants pipeline numbers by lunch, finance wants a revenue cut by Friday, and one overworked analyst is the choke point. Perplexity's latest move points Computer, its agentic workspace, straight at that queue by wiring it into Snowflake and Databricks.

The pitch is familiar: let non-technical users ask questions in plain English, let the agent write the SQL, return numbers tied to real warehouse tables. The interesting part is not the pitch. It's the governance plumbing underneath.

Key Details

The release positions Computer as a data agent for enterprise analytics, as TestingCatalog AI News reported. Users ask questions over authorized warehouse and lakehouse data. Computer generates queries, reads source tables, applies filters, and returns metrics tied to underlying data. The intended audience is business, product, sales, finance, and operations teams who can't write SQL on demand.

Use cases cover the boring but expensive workflows: pipeline analysis, product usage reviews, customer segmentation, revenue trend summaries, and recurring analytical workflows. The feature ships through Perplexity's Snowflake and Databricks connectors, and is gated to Pro, Max, Enterprise Pro, and Enterprise Max users. Admins control rollout at the organization level.

Coverage on the Snowflake side includes databases, schemas, tables, views, materialized views, and structured data formats like CSV, JSON, and Parquet-backed tables. The Snowflake docs are worth a re-read for anyone wiring this up against materialized views in particular, because cost behavior there gets opinionated fast. On the Databricks side, the integration touches Unity Catalog tables and views, Delta Lake tables, schemas, catalogs, external tables registered in Unity Catalog, and structured data. Unstructured assets, images, audio, video, files in warehouse-specific storage, are not supported at this stage.

The technical centerpiece is something called Data Map. Perplexity describes it as a shared organizational semantic layer, built from warehouse structure, table relationships, historical query patterns, and admin-provided business context. Admins can review and edit the map, refresh it, and approve proposed updates based on user feedback. That last bit matters more than the marketing suggests.

On auth, Snowflake supports user OAuth, service accounts with key-pair authentication, or programmatic access tokens. Databricks uses individual OAuth identity. Queries run under existing platform permissions, so access is enforced by Snowflake RBAC or Databricks Unity Catalog, not by Perplexity's UI. Admins can disable connectors, manage access, and enforce read-only behavior at the data platform level.

Why This Matters for Data Teams

Strip the marketing and there are two real questions for a platform lead: who owns the semantics, and who pays for the queries.

On semantics, Data Map is the right shape of answer. Teams I've worked with have all hit the same wall with text-to-SQL: the model is technically capable, but it doesn't know that rev_net_v3 is the table finance actually trusts and that rev_net_v2 is the one that quietly double-counts refunds. A semantic layer with admin review, refresh, and approval flow is how you stop the agent from confidently producing wrong numbers. It's the same pattern teams already build in dbt, just with an LLM consuming it instead of a BI tool.

My take: the value of this release lives or dies on how disciplined admins are about curating Data Map. Skip that work and you've shipped a very expensive way to generate plausible-looking nonsense.

On cost, the integration is a thin client on top of a warehouse that bills per query. Every "quick question" from a sales rep becomes a Snowflake or Databricks scan. I've seen production incidents where a single misconfigured BI dashboard ran a full table scan on a five-billion-row events table every fifteen minutes, and the monthly bill landed like a punch. Now imagine that, but the trigger is hundreds of non-technical users typing curious questions into a chat box.

The mitigations exist. Queries run under platform permissions. Admins can enforce read-only at the warehouse level. Materialized views and pre-aggregated tables are still your friend. But the budget conversation is going to be loud at any org that turns this on without query governance in place. Plan for it before procurement signs the SOW, not after.

The uncomfortable read: this product moves a meaningful chunk of analytical workload from your BI tool to a chat interface, and your warehouse bill won't care which one issued the query.

Industry Impact

For iGaming and fintech platforms, the calculus is specific. These verticals already run heavy analytical workloads against warehouses for risk scoring, player segmentation, fraud signals, and regulatory reporting. The promise of letting a fraud ops lead ask "show me deposit anomalies in the last 48 hours by region" without paging an analyst is genuinely useful. The risk is that the same query, run ad hoc against raw event tables, costs ten times what the equivalent dashboard tile costs against a properly modelled mart.

For ad-tech, the unstructured data gap matters. Creative assets, video, audio, logs sitting in warehouse-adjacent object storage, none of that is in scope yet. So Computer is useful for the spend and performance side, less so for the creative analysis side. Worth knowing before anyone in marketing assumes it does everything.

For enterprise infra teams, the auth story is the headline. Service accounts with key-pair auth, OAuth, programmatic tokens on Snowflake; OAuth identity on Databricks; permissions enforced by RBAC and Unity Catalog. That's the right answer. It means a security review can actually approve this without the team having to invent a new permissions model. It also means the blast radius of a compromised Perplexity account is bounded by what that user could already do in the warehouse, which is exactly how it should work.

The broader signal is that semantic layers are becoming the contested ground in analytics. dbt has one. Looker has one. Cube has one. Now Perplexity has one. Whoever owns the trusted definition of "monthly recurring revenue" inside your company owns the analytics workflow. That's a serious place to plant a flag.

What to Watch

Three things are worth monitoring over the next two quarters.

First, query cost telemetry. Any team rolling this out should instrument Perplexity-originated queries separately in Snowflake or Databricks usage data and review weekly. If you can't tag the source, you can't manage the spend. The first finance team to get a surprise warehouse invoice will be the last one to approve the next AI tool without cost guardrails.

Second, Data Map drift. Admin-approved semantic layers tend to decay the moment the person who built them changes role. Watch for whether Perplexity adds versioning, ownership metadata, and staleness signals on Data Map entries. Without those, the layer becomes shelfware in 18 months.

Third, the unstructured data question. Right now it's out of scope. If Perplexity extends this to query logs, support transcripts, or media metadata stored alongside the warehouse, the product becomes meaningfully more interesting, and the governance problem becomes meaningfully harder. For OLAP-heavy shops already evaluating engines like ClickHouse for log analytics, watch whether agentic interfaces start reaching into those stores too.

Key Takeaways

• Governance is genuinely good: queries inherit Snowflake RBAC and Databricks Unity Catalog permissions, so existing access control still applies.
• Data Map is the real product: the semantic layer with admin review is what separates this from a generic text-to-SQL toy. Curate it or skip the rollout.
• Budget for warehouse spend: every chat query is a billable scan. Tag, monitor, and pre-aggregate before opening it to wide audiences.
• Unstructured data is out of scope: images, audio, video, and warehouse-adjacent files aren't supported yet. Plan accordingly.
• Tier gating matters: only Pro, Max, Enterprise Pro, and Enterprise Max users get access, with admin controls at the org level. Procurement and IT need to be aligned before pilot.