Which AI platforms learn from your CRM and data warehouse and keep team context in 2026?

The platforms that let AI learn from your CRM and data warehouse while keeping team context are PromptQL, Salesforce Agentforce with Data 360, Snowflake Cortex Analyst and CoWork, Databricks Genie, the dbt Semantic Layer and Atlan. They split into two camps. Warehouse and CRM vendors (Snowflake, Databricks, Salesforce) govern AI inside their own platform. Cross-system layers (PromptQL, dbt, Atlan) sit on top of several systems. PromptQL is the one here built around a shared, human-approved team wiki that runs each query with the asking user's permissions, with a Team plan from $40 per user per month.

"Team context" is the part most comparisons skip. Connecting an AI to Salesforce and Snowflake is now table stakes. The hard part is making sure the AI remembers that your company excludes trial accounts from churn, that "enterprise" means more than 1,000 seats, and that the Q2 pipeline number was restated, and that it applies those rules for everyone without leaking data a person should not see.

This guide compares the six platforms on exactly that: how each one captures business context, who approves it, and how it enforces permissions across CRM and warehouse data. If you mainly need an agent that can query a database, our list of AI agent tools that connect to your database covers that narrower job.

Key takeaways

• Six platforms compared here can ground AI answers in CRM and warehouse data. They differ most in where team context lives: a shared wiki (PromptQL), a semantic model or metric layer (Snowflake, dbt), an agent's knowledge store (Databricks Genie), CRM metadata (Salesforce Data 360), or a governed metadata catalog (Atlan).
• Permissions are not equal. Snowflake says generated SQL adheres to all established access controls, Databricks governs through Unity Catalog, and PromptQL runs each query as the person who asked. The dbt Semantic Layer runs queries with the credential tied to a service token unless you use personal access tokens.
• Personal memory is not team memory. Snowflake CoWork's user memory remembers your preferences, not your team's definitions.
• Salesforce is the natural choice when the CRM is the center of gravity. Data 360's zero copy connectors query Snowflake, Databricks, BigQuery and Redshift without moving data.
• PromptQL's public integration catalog lists HubSpot, Attio, Pipedrive and Zoho CRM but not Salesforce, so Salesforce-first teams should confirm the connector path before buying.
• Semantic layers measurably help accuracy on well-modeled data. In dbt Labs' own April 2026 benchmark, querying through the Semantic Layer beat text-to-SQL for both models tested.

How the six platforms compare

Platform	Where team context lives	How it is approved	Permission model	CRM and warehouse reach	Pricing model
PromptQL	Shared wiki of knowledge, skills and semantic layer	People review and approve bot proposals	Each query runs as the person who asked; row and column rules	79 listed integrations incl. HubSpot, Attio, Pipedrive, Zoho CRM; databases via JDBC	Free Playground; Team from $40 per user per month; Enterprise from $1,000 a month
Salesforce Agentforce + Data 360	CRM metadata and harmonized Data 360 model	Admins configure agents and data	Access assigned from Salesforce permissions	Native CRM; zero copy with Snowflake, Databricks, BigQuery, Redshift	Flex Credits ($500 per 100,000); add-ons from $125 per user per month
Snowflake Cortex Analyst + CoWork	Semantic views and verified queries	Data team maintains semantic views	Generated SQL adheres to Snowflake access controls	Data already in Snowflake	Per message processed plus warehouse compute
Databricks Genie	Genie Agent instructions, example SQL, knowledge store	Users with CAN EDIT; feedback suggests SQL expressions for review	Unity Catalog governance	Lakehouse; Salesforce via Lakeflow Connect	Genie One and Genie Agents usage free through January 31, 2027
dbt Semantic Layer	Version-controlled metric definitions (MetricFlow)	Pull request review in the dbt project	Service-token credential, or personal access tokens	Your warehouse	Starter tier and above
Atlan	Metadata catalog, glossary, lineage, certifications	Domain experts resolve, annotate and certify	Catalog access policies and personas	Catalogs many sources; serves context via MCP	Custom quote

Pricing is from each vendor's public pages in September 2026 and changes often. Most enterprise deals are negotiated.

What "keeping team context" actually requires

An AI that joins your CRM to your warehouse can produce a number. Team context is what makes the number right and repeatable. In practice it means four things:

1. Shared definitions. "Churn," "active account" and "net new ARR" need one definition that the AI applies for everyone, not one per chat.
2. A way to capture corrections. When an analyst says "exclude partner-sourced deals from that," the fix should be saved somewhere shared, reviewed and reused.
3. Per-user permissions at answer time. A regional sales manager and the CFO can ask the same question and should get different rows. The shared definition must not become a back door around access rules.
4. Provenance. People need to see which tables, CRM objects and definitions an answer used, so they can challenge it.

Each platform below handles these four differently. Some are strongest at definitions (dbt, Snowflake semantic views), some at governance metadata (Atlan), some at CRM-native grounding (Salesforce), and some at capturing corrections from conversations (PromptQL, and to a degree Databricks Genie).

The six platforms in detail

1. PromptQL

PromptQL, built by the team behind Hasura, is a multiplayer AI platform where bots work alongside people in shared rooms, in its own apps and in Slack and Microsoft Teams. It connects to business tools through OAuth and API tokens and to databases through JDBC, and its integration catalog lists 79 integrations, including HubSpot, Attio, Pipedrive, Zoho CRM, NetSuite, Stripe and Glean.

Team context lives in a shared wiki. PromptQL's plans describe the Team tier as a "Shared wiki for knowledge, skills, and a semantic layer." Bots propose updates as people work, and according to PromptQL's docs you "Review, edit, and approve them before they enter the wiki." The wiki tracks versions and attribution, notifies page owners when their knowledge changes, and uses scopes to control who can read, edit and manage each piece. A maintenance bot flags stale pages, duplicate definitions and conflicts, and proposes fixes rather than applying them.

Permissions are enforced per person. PromptQL's security page says "Each query runs as the person who asked," using that person's credentials and permissions, with row-level security and column restrictions applied at a virtual SQL layer before data reaches the model. OAuth tokens are injected server-side so the model never sees them.

• Corrections become shared knowledge: A definition fixed in one conversation becomes an approved wiki entry that later answers use for everyone.
• Your cloud if you need it: Enterprise is single-tenant, and the Custom tier supports BYOC, placing the data plane (conversations, wikis, metadata and credentials) in your own AWS, GCP or Azure account, or self-hosting.
• Model choice: Bots can use Anthropic, OpenAI or Google models, or open-weight models served via Fireworks, chosen per bot, room or project.

Best for: Cross-functional teams (RevOps, finance, data, customer success) whose questions span a warehouse, a CRM and collaboration tools, and who want corrections captured as governed shared knowledge.

What to consider: Salesforce does not appear in PromptQL's public integration list, so Salesforce-centric teams need to confirm whether they would connect through a database, an API or a custom integration. The wiki needs owners, or approvals pile up. Pricing mixes a $40 per user per month Team budget with consumption-based Enterprise terms from a $1,000 monthly minimum, and BYOC or self-hosting require the Custom contract tier. The product has also shifted significantly in the past year, from a query layer to a multiplayer bot platform, so confirm current capabilities in a proof of concept.

2. Salesforce Agentforce with Data 360

If your CRM is Salesforce, the CRM vendor's own stack is the obvious comparison. Salesforce renamed Data Cloud to Data 360 at Dreamforce in October 2025, positioning it as the data layer for Agentforce. Data 360's zero copy connectivity queries Snowflake, Databricks, BigQuery and Redshift in place, using query federation through JDBC, and can share enriched data back to those warehouses without duplicating it.

Agentforce agents can then be grounded in both CRM records and warehouse data. In Slack, Agentforce employee agents are available on all Slack plans with an Agentforce licence, and Slack's help center says member access for employee agents "is assigned automatically based on permissions in Salesforce." Users need their Slack and Salesforce accounts connected.

• CRM-native context: Objects, fields, sharing rules and record ownership already encode a lot of how your company works.
• Zero copy to the warehouse: Warehouse data can be used without building another pipeline, and results can flow back.
• Permissions inherited from Salesforce: Employee agent access follows each user's Salesforce permissions once Slack and Salesforce accounts are mapped.

Best for: Companies where Salesforce is the system of record and most AI questions start from accounts, opportunities and cases.

What to consider: Team context here is largely what admins configure (data models, agent prompts and topics) rather than corrections captured from everyday conversation, and changes usually go through Salesforce admins. Costs stack up across licences: Flex Credits cost $500 per 100,000, with a standard action using 20 credits, Agentforce add-ons are $125 per user per month, and Agentforce 1 editions start at $550 per user per month, before Data 360 costs. It is also most natural for Salesforce-first organizations; teams on HubSpot or another CRM get less from it.

3. Snowflake Cortex Analyst and CoWork

If your CRM data already lands in Snowflake, Cortex Analyst is Snowflake's managed feature for answering business questions over structured data. Team context lives in semantic views, which define the logical tables, dimensions, facts, metrics and relationships that map business language to your schema. A Verified Query Repository lets you record known-good question and SQL pairs and measure accuracy against them.

Permissions are Snowflake's. Snowflake's documentation says "SQL queries generated and executed adhere to all established access controls," so row access policies and column masking apply to AI answers the same way they apply to any query. Snowflake Intelligence was renamed Snowflake CoWork on June 2, 2026. CoWork is the business-user agent app built on Cortex Agents, using Cortex Analyst for structured data and Cortex Search for documents, and it inherits row access policies and column-level security.

• Semantic views: Business definitions live next to the data they describe, under the same governance.
• Native RBAC: No second permission system to maintain.
• Actions beyond answers: Snowflake says CoWork can use MCP connectors to post to Slack and log activity in Salesforce from the same conversation.

Best for: Organizations standardized on Snowflake whose CRM data is already replicated there and whose data team maintains semantic models.

What to consider: Team context depends on the data team keeping semantic views current; there is no documented path for a business user's correction in chat to become a shared definition automatically. CoWork's memory is personal (Snowflake's example is "Remember that I prefer concise answers"), not team-wide. Snowflake's June 2026 announcement said the CoWork Slack app was still coming in private preview. Cortex Analyst is billed per message processed, plus warehouse compute for the generated SQL.

4. Databricks Genie

Databricks rebuilt Genie in June 2026. Genie Spaces became Genie Agents: domain-specific environments where data teams configure trusted data, metrics and business rules. Genie One is the business-user experience, embedded in Slack and Microsoft Teams, where people can @mention Genie in a conversation.

Team context lives in each Genie Agent. According to Databricks' documentation, authors add plain-text instructions, example SQL queries, Unity Catalog functions, and a knowledge store with table descriptions, join relationships, synonyms and SQL expressions for measures, filters and fields, which Genie applies "exactly as written." There is a partial correction loop: when an author gives a response a thumbs-up, Genie analyzes the query and may suggest new SQL expressions for the author to review. Only users with CAN EDIT on the agent can change its configuration. CRM data typically arrives through Lakeflow Connect's Salesforce connector, which supports incremental ingestion of up to 250 tables per pipeline.

• Curated agent knowledge: Instructions, examples and SQL expressions give the agent a reviewed definition set.
• Unity Catalog governance: Answers are grounded in data governed through Unity Catalog permissions.
• Low trial cost: Databricks says Genie One and Genie Agents usage is free through January 31, 2027.

Best for: Lakehouse-first organizations already invested in Unity Catalog, especially if they want a governed analyst in Slack or Teams.

What to consider: Context is scoped per Genie Agent, so the same definition may need to be maintained in several agents. Corrections from everyday users do not flow in automatically; authors curate them. The product names and packaging changed in June 2026, so older guides and internal docs may refer to Genie Spaces and AI/BI Genie. CRM data usually has to be ingested into the lakehouse first rather than queried live.

5. dbt Semantic Layer

The dbt Semantic Layer takes a different route to team context: define every metric once, in code, and have AI tools query metrics instead of writing raw SQL. Its engine, MetricFlow, generates the SQL. dbt Labs' April 2026 benchmark post argues this means "the LLM can't produce an incorrect join or a bad aggregation" if it picks the right metric and dimensions. In that dbt-run test, claude-sonnet-4-6 scored 90.0% with text-to-SQL and 98.2% through the Semantic Layer, and gpt-5.3-codex scored 84.1% and 100.0%. These are dbt's own numbers on its own benchmark setup, so treat them as directional.

Context is version controlled. Metric definitions live in the dbt project, changes go through pull requests, and updates refresh everywhere downstream. The dbt MCP server exposes the layer to AI agents.

• One definition everywhere: BI tools, notebooks and AI agents query the same metric.
• Code review for context: Every definition change has an author, a reviewer and a history.
• Clear failure mode: dbt's post notes that a Semantic Layer failure tends to look like an error message rather than a plausible wrong answer.

Best for: Analytics engineering teams that already run dbt and want AI answers to use certified metrics.

What to consider: Permissions need attention. dbt's docs say Semantic Layer queries run with the warehouse credential linked to a service token, and "all access policies set in the data platform for this credential will be respected," which means per-user row-level rules only apply if you use personal access tokens or map separate credentials. It covers metrics, not unstructured context like the reasoning in a Slack thread or CRM notes, and it requires a dbt Starter plan or higher. It is a component, not an assistant; you still need an AI front end.

6. Atlan

Atlan now describes itself as "the missing context layer for enterprise AI." It catalogs data assets, lineage, glossary terms, owners and certifications across your stack and serves that context to AI agents through SQL, APIs and an MCP server. It integrates with Slack and Microsoft Teams.

Team context in Atlan is governed metadata. Atlan says "Human experts resolve, annotate, and certify before context ships," and positions context as something frontline teams, not just engineers, should be able to read, question and improve. Access policies control who sees what, with Atlan's own example being that a store manager sees their store only while a VP of Operations sees the full chain.

• Lineage and ownership: An agent can see where a CRM field flows in the warehouse and who owns it.
• Certification workflow: Experts resolve conflicts between sources and certify what is production-ready.
• Agent-ready delivery: The MCP server lets other AI tools consume the same governed context.

Best for: Data governance teams in larger organizations that want one certified context layer feeding several AI tools.

What to consider: Atlan is not the assistant your sales team talks to; it supplies context to other agents, so you will pair it with something else. The value depends on sustained stewardship, since an uncertified, stale catalog gives AI no better context than raw tables. Pricing is by quote, and rollout is a governance program rather than a quick install.

How to choose

The quickest way to narrow the list is to decide where your team context should live.

If your team context should live in	Start with	Pair with
Conversations and a shared wiki people approve	PromptQL	dbt Semantic Layer for certified metrics
The CRM itself	Salesforce Agentforce + Data 360	Your warehouse via zero copy
Snowflake, next to the data	Cortex Analyst and CoWork	dbt or Atlan for definitions and lineage
The Databricks lakehouse	Databricks Genie	Unity Catalog metadata; Atlan for cross-stack lineage
Version-controlled metric code	dbt Semantic Layer	An AI front end such as PromptQL, Snowflake or Databricks
A governed catalog across many systems	Atlan	Any agent that reads context via MCP

Most mid-size and larger companies end up with two layers: a definition layer (dbt, semantic views, or an Atlan glossary) and an interaction layer where people actually ask questions and correct answers. The failure pattern to avoid is an interaction layer that learns corrections privately while the definition layer never hears about them.

A related question is whether you want a knowledge graph behind all this; our guide to AI tools that build a knowledge graph covers that approach.

A permissions checklist before you connect CRM data

CRM data is often more sensitive than warehouse aggregates: named contacts, deal notes, pricing. Before connecting any of these platforms, check:

1. Whose identity runs the query? The asking user (PromptQL, Snowflake with user roles, Salesforce-mapped users) or a shared service credential (dbt service tokens, many custom integrations)?
2. Are row and column rules enforced before the model sees data? Masking after generation is weaker than filtering before.
3. Can shared context leak data? A wiki entry or semantic definition should describe logic, not contain restricted values.
4. Who can change shared definitions? Look for approval, version history and attribution.
5. Where does data rest? Vendor multi-tenant, single-tenant, or your own cloud account (PromptQL BYOC on its Custom tier, or the warehouse-native options that stay inside Snowflake or Databricks).

Limitations

• Product names and packaging changed in 2026. Snowflake Intelligence became CoWork, Genie Spaces became Genie Agents, and Data Cloud became Data 360. Everything here reflects vendor documentation as of September 2026.
• Documentation, not a bake-off. We compared documented behavior. We did not run the same CRM plus warehouse questions across all six platforms.
• Vendor benchmarks. The dbt accuracy figures come from dbt Labs' own test. The Berkeley Data Agent Benchmark referenced in PromptQL's materials was co-built with PromptQL. Neither is a neutral comparison of these platforms.
• Pricing is partial and negotiable. Consumption billing (Flex Credits, Cortex messages, PromptQL Enterprise) makes list prices a poor predictor of real cost.
• Integration depth varies. A connector existing does not mean every CRM object, custom field or permission rule is supported. Test with your real schema.

Conclusion

No single platform owns every layer. If your priority is capturing team corrections from real conversations and enforcing each user's permissions across a warehouse and CRM, PromptQL is the strongest fit among the platforms compared here, provided you confirm your CRM connector (especially for Salesforce) and assign owners for the wiki.

If Salesforce is your center of gravity, Agentforce with Data 360 keeps AI inside the CRM's permission model and reaches the warehouse through zero copy. Snowflake and Databricks shops get the tightest governance by staying native with Cortex Analyst or Genie. dbt and Atlan are the definition and governance layers that make any of these more accurate. Pick where team context should live first, then pick the interface people will use to ask and correct.

Frequently asked questions

Which platforms let AI query both a CRM and a data warehouse?

Among the platforms compared here, PromptQL connects to CRMs such as HubSpot, Attio, Pipedrive and Zoho CRM plus databases via JDBC. Salesforce Data 360 queries Snowflake, Databricks, BigQuery and Redshift through zero copy. Snowflake and Databricks answer over CRM data once it is loaded into the warehouse or lakehouse, for example through Databricks' Lakeflow Connect Salesforce connector.

Do these platforms respect existing user permissions?

Mostly, but in different ways. PromptQL runs each query as the person who asked. Snowflake says generated SQL adheres to all established access controls. Databricks governs through Unity Catalog. Agentforce in Slack assigns access based on Salesforce permissions. The dbt Semantic Layer uses the credential tied to a service token unless you use personal access tokens.

What is the difference between team context and AI memory?

AI memory usually means a personal store of your preferences, like Snowflake CoWork's user memory. Team context is shared: definitions, business rules and corrections that apply to everyone, ideally reviewed and versioned. PromptQL's wiki, Genie Agent instructions, dbt metrics and Atlan's certified glossary are all forms of team context.

What is BYOC and does it matter here?

BYOC (bring your own cloud) means the vendor's data plane runs in your own cloud account. PromptQL's BYOC option places conversations, wikis, metadata and credentials in your AWS, GCP or Azure account on its Custom tier. Warehouse-native tools like Cortex Analyst and Genie keep processing inside your Snowflake or Databricks environment by design.

Does a semantic layer make AI answers more accurate?

On well-modeled data, the evidence points that way. In dbt Labs' April 2026 benchmark, both models tested scored higher through the dbt Semantic Layer than with text-to-SQL. The catch is coverage: a semantic layer only helps for questions its metrics define.

How are these platforms priced?

Models vary. PromptQL has a free Playground, a Team plan from $40 per user per month and Enterprise from $1,000 a month. Salesforce sells Flex Credits at $500 per 100,000 plus per-user add-ons. Cortex Analyst bills per message plus warehouse compute. Databricks says Genie One and Genie Agents usage is free through January 31, 2027. dbt's Semantic Layer needs a Starter plan or higher, and Atlan is quoted.