A new kind of BI that connects straight to the data warehouse and writes live SQL. What will it solve, and where will it fall short?
The old way vs the new way
Anyone who works with data knows this: a large part of our time goes not into deep analysis but into ad-hoc requests from the business, such as "Could you put together a quick report comparing last week's numbers with the previous month?"
Anthropic's newly announced Claude Dashboards feature is a step that could change that habit at its root. Until now, people uploaded a CSV to an LLM and used it for static, one-off charts. The new feature is no longer a visualization tool next to a chat window; it has become a live business intelligence (BI) layer that connects directly to the data warehouse.
I spent some time with it to see how it works in practice, what it means for security and where its limits are. Let's go through the technical details, the data security risks and the architectural consequences from a data professional's point of view.
More Than Charts: Live Orchestration on the Data Warehouse
What separates Claude Dashboards from classic chat tools is that it connects directly to modern data warehouses such as Google Cloud BigQuery, Snowflake, Databricks and Redshift.
Looking at the mechanism behind it, one architectural detail stands out: push-down compute, meaning the computation is pushed to the data warehouse.
When you connect it to a warehouse like BigQuery, Claude does not pull the data into its own memory and draw it. If you grant the permissions, it can create new tables inside your warehouse, clean empty columns in the raw data, and run the ETL that turns messy text fields into numeric columns you can average, all as SQL executed on the warehouse itself.
It also addresses the black-box problem that managers and data teams fear most in AI output, through transparency:
- Click any KPI card or chart on the dashboard and the Sources panel shows the live SQL query that produced that number, along with every transformation, line by line.
- When needed, you can edit that SQL directly in the panel and save it.
- The Refresh button at the top re-runs the query against the warehouse and pulls the latest numbers.
One-Shot Cloning: From a Screenshot to a Live Dashboard
The most concrete answer to the question "Is the era of building charts over?" shows up in migrations between BI tools and in rapid prototyping.
Normally, moving from an old reporting system or from Power BI, Tableau or Metabase to another environment takes days: understanding the data model, rewriting the metrics, rebuilding the layout. With this new approach you can take just a screenshot of an existing dashboard, give it to Claude and say "Build an exact copy of this with the same colors, metrics and visuals."
The steps it follows are impressive:
- It analyses the image and works out the visual components, the color palette and the layout in a single pass (one-shot).
- It goes to the connected warehouse and writes, from scratch, the GROUP BY, filter and ROUND queries needed to reproduce every metric in the image.
- Within seconds it renders a live dashboard whose metrics match the original report exactly.
In short, aligning boxes by drag and drop, matching color codes and building simple summary charts from scratch now comes down to uploading a single screenshot.
The Other Side of the Coin: Security, Privacy Law and the Right Architecture
Connecting an AI tool straight to BigQuery or Snowflake, the heart of the company, sounds great. But once we get to data security and data governance, our red lines come into play.
Handing Claude the keys to your main database is a serious risk in a corporate setting. So how should the integration be done, and what should you watch for?
1. Isolate and Mask Personal Data (PII)
In finance, insurance, healthcare or e-commerce, customer tables contain personal data (PII) such as national ID numbers, names, phone numbers, home addresses or account details. Under KVKK (Turkey's data protection law) and GDPR, exposing these raw tables directly to a third-party LLM connector is a serious breach. The layer Claude connects to must not contain person-level sensitive data; where needed, these fields should be masked (hashed or anonymised) in the pipeline.
2. Not Raw Data, but Pre-Computed Summary Tables (Data Marts)
Connecting Claude to raw transaction tables with millions of rows causes two big problems:
- Hallucination and miscalculation: The AI cannot know your company's specific business rules (for example, how cancelled orders are subtracted from revenue, or how you define an active customer). When it calculates from raw data on its own, the chance of a wrong result is high.
- Cloud cost (FinOps) explosion: If every click on Refresh sends a SELECT over terabytes of raw data, you will face a huge cloud bill at the end of the month.
What does the right architecture look like? Instead of letting the AI into the kitchen (the raw data), we should only hand it the serving plate (the mart layer). Using dbt or similar tools, clean the data in advance and build tables that are calculated according to business rules, anonymised and aggregated. Claude should connect only to these safe, pre-computed summary tables.
3. Read-Only Access and an Isolated Environment
If you give Claude write access to the warehouse, it may create new tables and split data on its own initiative to make visualization easier. An AI creating tables at will in a production warehouse is architectural chaos. Connections should therefore always go through a restricted read-only service account that can see only specific schemas.
Why It Can't Replace Power BI, Tableau and Metabase Yet (Technical Limits)
Beyond security and architecture, there are important gaps that show it cannot immediately replace traditional BI tools at enterprise scale:
- No version control (Git): Being able to undo changes is vital when you build a data product. Right now only the history of the underlying SQL is kept; when you change the visual design, the new design overwrites the old one, and there is no Revert button in the interface. To track versions, you have to export the code and move it to your own Git/GitHub manually.
- Sharing barrier: When I wanted to share a live dashboard with a teammate, they needed access to the same BigQuery table and their own Claude connector to see the numbers. To share reports quickly with external stakeholders or clients, you still have to export to third-party tools such as Amplitude, Grafana, Hex or Mixpanel (or Tableau and Looker, which are coming soon).
- Reproducibility and consistency: On small, clean data sets AI can work wonders. But in an environment with hundreds of tables and complex business rules, you still need strong BI governance so that every metric gives consistent, reproducible results in every department.
A New Era for Data Analysts: Where Are We Heading?
The rise of tools like Claude Dashboards sends a clear message: the front-end layer (visualization and BI) is quickly becoming a commodity and being automated.
If your only value as a data analyst is running SELECT queries and turning them into nice-looking charts, you have to accept that this part of the job is changing. On the other side of the coin, a far more valuable space is opening up for data professionals:
No matter how good our tools get, AI cannot work miracles on its own unless the engineering behind it is designed properly.
What makes technology powerful is not just the tools, but the knowledge and engineering approach that let us use them well.