Claude Can Write SQL. <sarcasm>BFD</sarcasm> But Can It Understand Revenue?
Codifying institutional tacit knowledge can be converted into agent-readable operating procedures.
(And, YES! It is a BFD that Claude CAN write SQL! No sarcasm. But we’re ppast that being impressive. AGI has been here for over a year.)
This is a MUST READ article from Anthropic on how they’re using Claude to automate business analytics and data science.
But the most important point is not that Claude can write SQL. That part’s easy <cough>. The hard part is knowing which table to use, which definition of “active user” the business actually means, which dashboard is canonical, which metric is stale, which filter everyone forgets, and whether the answer is safe enough to forward to leadership.
That’s the real work.
Anthropic’s post makes something very clear: the future of analytics is not about giving everyone a chatbot connected to Snowflake. That’s how you manufacture confident nonsense at enterprise scale.
The future is a governed analytics harness: semantic layers, canonical datasets, agent skills, evals, CI, provenance footers, adversarial review, and human-owned definitions. In other words... today… the model is not the moat. The harness is. (For now.)
Anthropic says their internal data science team now automates 95% of business analytics queries using Claude, with roughly 95% aggregate accuracy, and around 99% accuracy in some domains.
Nice. Very nice.
But the real insight is this: when business analytics fails, it’s usually not because the AI can’t write SQL, R, or Python. It’s because the AI doesn’t understand the business.
Because “revenue” is not always revenue. “Active user” is not always active user. “Retention” is not always retention. Every company has a swamp of definitions, exceptions, stale dashboards, undocumented assumptions, legacy tables, and institutional scar tissue.
Welcome to enterprise data.
Please wipe your feet.
That’s why this is not really a post about SQL. It’s a post about analytical trust. And that means the analyst’s job is changing.
The analyst is no longer just the person pulling numbers or whispering to the data science team. The analyst becomes the architect of the system that makes AI-generated analysis trustworthy.
That means the new skill stack is not just SQL. It’s metric governance, semantic layer design, agent instruction design, evaluation engineering, and provenance design. Can you define the metric? Can you route the question through the canonical data model? Can you encode the judgment of a senior analyst into reusable agent skills? Can you test the agent’s answers before executives rely on them?
That is the work now.
The analyst does not disappear. The analyst gets promoted into the higher level architecture.
(Side note: This is an example of the effects of “Disruptive Innovation”.)
The Skills Are the Interesting Part
Anthropic’s skills are where this gets really interesting. Their pattern has three parts: knowledge skills, runbook-style skills, and reference docs written for LLM retrieval.
Knowledge skills route the agent to the right domain references, semantic layer, tables, joins, exclusions, and gotchas. This is the map that tells Claude where the reliable knowledge lives.
Runbook-style skills encode the workflow of a senior analyst: clarify the question, find the right source, use the semantic layer first, run the query, review assumptions, validate the result, and produce the final answer. This is not just knowledge. This is procedure.
Reference docs written for LLM retrieval turn messy human knowledge into agent-readable operating instructions: scope, exclusions, required filters, “when to use / when not to use,” routing triggers, and known traps.
This is the part I find most fascinating. We are watching institutional tacit knowledge get converted into agent-readable operating procedures.
The stuff that used to live in someone’s head now has to become explicit: “Don’t use that dashboard.” “That table is deprecated.” “Finance uses a different version of revenue.” “Ask Sarah before sending that number to leadership.”
All of that messy human knowledge now becomes part of the harness.
That is the real transformation.
We’re able to codify the tacit knowledge in an organization!
A Few Caveats
The 95% automation / 95% accuracy claim is impressive, but still under-specified. How complex were the queries? What counts as accuracy? How many answers needed human correction? And how much of this depends on Anthropic being Anthropic?
The jump from 21% without skills to 95%+ with skills is also striking. But that proves the point: the gains are not just coming from the model. They are coming from the operating environment around the model: the skills, documentation, governance, CI, semantic layer, and maintenance discipline.
Many companies do not have that level of canonical data hygiene.
And this quietly undercuts a lot of “just connect AI to your database” vendor hype. Access alone is not enough. Your enterprise chatbot does not become smart because you connected it to a swamp. It becomes dangerous.
The Big Takeaway
Is this going to be generalizable? Maybe. Highly likely.
Is this a state machine? It feels like it.
Is this a decision tree? It can be.
Is this deterministic? Yes, similar to an expert system.
Is this fragile. Yeah… everything is fragile. (But…???)
Is this gonna burn lots of tokens? Yes. By design.
Is this better than fine tuning a model? Fine tuning models is a waste.
Anthropic’s article is not really about Claude writing SQL. It is about the new enterprise analytics architecture: governed data models + semantic layers + agent skills + evals + provenance + adversarial review + human ownership.
That is the new stack.
(Well… for now. I give it a few days/weeks and something else will replace it.)
The analyst’s job is not going away. It is moving up the value chain from pulling numbers to designing trust, and from answering questions to governing how questions get answered.
So yes...
Claude can write SQL.
Big freakin’ deal.
The real question is:
Can it understand revenue?