AI integrations that actually pay off in a data platform

Every data team is being asked the same question: what is our AI strategy? The honest answer for most organizations is that the model is the easy part. The hard part is getting trusted, well-modeled, governed data in front of it — which is exactly the work data engineering and BI teams already do.

Here are the integration patterns that consistently make it past the pilot stage.

1. LLMs as a transformation step

The quickest win is often the least glamorous: using a language model inside a pipeline to classify, extract or summarize unstructured text. Support tickets, contract clauses, product descriptions, free-text survey answers — data that used to be ignored becomes columns you can report on.

Warehouse-native functions such as Snowflake Cortex AI and Databricks AI Functions make this a SQL call, which means the output lives in the same governed tables, lineage and access controls as everything else.

select
    ticket_id,
    snowflake.cortex.classify_text(
        body, ['billing', 'outage', 'feature request', 'other']
    ):label::string as ticket_category
from {{ ref('stg_support_tickets') }}

2. Retrieval-augmented assistants over enterprise data

Retrieval-augmented generation (RAG) lets an assistant answer questions from your documentation, policies or knowledge base with citations. The success factors are almost entirely data engineering: chunking strategy, metadata, freshness, and permissions that follow the user.

3. Natural-language access to the warehouse

Text-to-SQL is only as good as the semantic layer underneath it. A model querying raw tables will confidently produce wrong numbers. A model querying well-named, documented dbt models with defined metrics produces answers the business can trust.

If your metrics are not defined consistently for humans, they will not be defined consistently for an AI either.

4. Agents connected to your tools

Agent frameworks and open standards such as the Model Context Protocol (MCP) make it practical to give an assistant controlled access to catalogs, orchestration tools and BI platforms. Start read-only, log everything, and expand scope deliberately.

What to get right first

• A documented, tested transformation layer (dbt is the de facto standard).
• Clear ownership and access policies — AI must inherit them, not bypass them.
• Cost monitoring: token usage and warehouse compute belong on the same dashboard.
• Evaluation: a small, versioned set of questions with known answers to catch regressions.

AI does not replace the data platform — it raises the stakes on getting it right. If you are planning your first production AI integration, I am happy to review the approach.