Anthropic published a writeup on how they use Claude for self-service data analytics. It is worth reading, partly because it is not really a "look what Claude can do" post.
At least that was not my takeaway.
The interesting bit is how much has to sit around the model. Definitions, curated context, evals, provenance, routing, correction loops. The stuff that gets cut from the demo because it does not sparkle.
The model is the part people notice. The data model is the thing making the answer possible.
Self-service still has a service
Most companies say they want self-service analytics.
Usually this means fewer Slack drive-bys to the data team, fewer one-off dashboard requests, fewer "can you pull this for me real quick" messages that are never actually quick.
Fair. Nobody wants the data team to be a human query endpoint.
But self-service does not remove the service. It moves it.
Instead of asking an analyst to write the query, users depend on the definitions, joins, examples, permissions, and caveats the data team already encoded. If those are thin, stale, or contradictory, the chat interface just makes the confusion faster.
The model should not guess the business
The obvious failure mode for AI analytics is wrong SQL.
The worse failure mode is right-looking SQL built on the wrong business meaning.
What does "active customer" mean? Which revenue number counts? Are trials included? Is churn based on invoice status, account state, product usage, or something else? Which timestamp is the real event time? Which table has the current truth and which one is a convenient trap?
These are not syntax questions. They are product questions.
If the model has to reverse-engineer the answers from table names and a pile of old queries, the system is already doing the risky part in the wrong place. A better model might guess more often correctly. It is still guessing.
The semantic layer becomes the UI
In older BI systems, the semantic layer often felt like cleanup work.
Make dashboards agree. Define metrics once. Stop five teams from shipping five versions of "pipeline."
Still useful. Also a little thankless.
With AI analytics, that layer moves closer to the center of the product. It is not just there so dashboards render the same number. It is there so the model can understand what the company means when someone asks a question.
That includes metrics and joins, sure. It also includes the words people actually use, the questions they tend to ask, the tables everyone should avoid, the caveats that never made it into the dashboard, and the permission rules nobody wants Claude to improvise.
Raw warehouse access is not enough. A warehouse schema is not a business interface. It is an implementation detail with years of decisions fossilized into it.
Analysts become product owners
I do not buy the version of this story where analysts vanish.
The job gets weirder and probably more important.
Someone still has to decide what counts as the canonical metric. Someone has to notice that sales and product use the same word to mean different things. Someone has to look at a bad answer and say, "the model did what we told it to do, and what we told it to do was wrong."
That is not dashboard work. That is product work.
The product is the organization's understanding of its own data, expressed in a way software can use.
Users experience that model every day. They just experience it through a chat box, dashboard, report, export, or agent.
The unglamorous version
You do not need Anthropic's exact stack to learn from it. Most companies could start with one messy domain and do the annoying work:
- pick one domain
- define the ten metrics people fight about
- document the joins and caveats
- collect real questions from users
- write evals for those questions
- require answers to show where they came from
- review failures weekly
This will not look impressive in a launch video. It will make the system less embarrassing.
The model can make analytics feel conversational. Great. But the conversation is only useful if the system knows what the words mean.
So yes, Claude doing analytics is interesting.
But the real lesson is less glamorous: self-service analytics works only when the data model is good enough to be treated like the product.