This week, Anthropic published a detailed account of how they built self-service data analytics internally using Claude. It’s an unusually candid piece — and the candor is what makes it valuable. Not because it showcases what Claude can do, but because it documents what it actually took, and where the limits are.
We read it carefully. Here’s our perspective.
1. The problem is real
Anthropic didn’t publish this piece to market their product. They published it because connecting AI to real business data turned out to be genuinely hard — hard enough that their own engineering team had to build a significant internal infrastructure project just to make it work for themselves.
That’s worth pausing on. The company that builds the most capable AI models in the world needed months of specialized engineering work to answer their own internal business questions reliably. Not because Claude isn’t powerful. But because the problem isn’t a model problem. It’s a data problem.
If the leading AI lab in the world needs to invest at this level to solve this for themselves, that tells you something fundamental about the nature of the challenge.
2. LLMs alone can’t solve it — because data is not software
This is not a new insight for us. In 2021, I read the original Transformers paper and tested GPT-2. The models were impressive — but it was immediately clear that the hard problem wasn’t the model. It was the data underneath it. That became the founding premise of Teramot, and it’s what I’ve said in every pitch, every talk, and every architecture decision since.
In June 2026, Anthropic published it as a section title in their own blog. That’s not a coincidence. It’s a confirmation.
Code is deterministic and self-describing. You can read a function and understand what it does. Data is different. It carries business context, organizational history, ambiguity, and constant change that accumulates over years inside a company. A model — no matter how capable — cannot infer what “revenue” means in your specific context, which entities your team refers to as “the Q2 launch,” or why two departments calculate the same metric differently.
Anthropic tested this assumption directly. When they tried to bootstrap their semantic layer using LLMs — having the model auto-generate metric definitions from raw tables and query logs — the output was “net-negative on our evals versus a smaller, human-curated layer.” The model produced plausible-looking definitions that encoded the very ambiguities they were trying to eliminate.
You can’t prompt your way out of a data problem.
3. Their solution required a world-class team — and they’re still not satisfied
What Anthropic built is technically impressive. Governed canonical datasets, a human-curated semantic layer, a business context graph with indexed documents and organizational structure, colocated artifacts across layers with CI checks that protect cross-layer integrity, and a provenance footer on every response showing where the data came from and how fresh it is.
That’s exactly the right architecture. The issue is what it took to build and maintain it.
It required a specialized team working over multiple months. Every incremental change to the metadata triggers evaluation runs that take an hour each and cost significant resources. Skills need to be reviewed and updated every month as business logic and data models evolve.
And after all of that investment, they report 95% accuracy in production responses — and are candid that they don’t yet have a robust solution for what they call “silent failures”: the cases where the answer is wrong, looks plausible, and gets used without anyone noticing. One answer in twenty reaching a senior leader undetected isn’t a production system. It’s a very well-built prototype.
4. Every critical layer requires humans in the loop
Perhaps the most telling detail in the piece: after all the engineering investment, human oversight remains structurally necessary at every critical layer.
The semantic layer cannot be automated — as their own experiments showed, model-generated definitions are worse than smaller human-curated ones. The business context layer, which they describe as “the layer most teams skip,” requires continuous human curation to stay current with how the business actually works. Every response destined for leadership requires explicit human sign-off before it’s trusted.
Their conclusion is direct: generate documentation with AI, but have a human own the definition.
That’s the right answer for Anthropic. They have the team to make it work. But it raises the obvious question: what does a company do if it doesn’t have that team?
For any organization that wants to replicate what Anthropic built, the honest answer is that they need to hire and maintain a dedicated group of specialized data and AI engineers — not to build the product, but just to keep the AI data layer running. That’s not a product strategy. That’s a cost center. And it’s one most companies can’t sustain.
What this means — and how Teramot approaches it differently
The architecture Anthropic describes is not wrong in itself. The problem is the operating model it requires.
The reason their approach depends on ongoing human maintenance is structural: they built an AI layer on top of existing data infrastructure, and compensated for the lack of a structured foundation with skills, reference documentation, and human curation. When the underlying data changes — which it does, constantly — the text-based knowledge layer drifts, and humans have to realign it.
Teramot starts one layer below. Before anything reaches an AI agent, we deploy a full data lakehouse — Bronze, Silver, and Gold layers — where every transformation is deterministic SQL and all metadata lives in a structured relational database, not in text files. Once that foundation is in place, the AI doesn’t infer — it queries. The result is the same every time.
Does this require human iteration to build? Yes — and we think that’s honest. Business knowledge doesn’t emerge automatically. The difference is that with Teramot, that iteration is fast — typically under an hour to deploy a first functional version, and five minutes for each new iteration — and stable over time. You don’t rebuild it every month. You build it once, and it holds. Because the knowledge is structured and relational, changes to the data model propagate correctly instead of silently breaking a text file somewhere downstream.
After three years of R&D and production deployments across companies in Latin America, we launched in February 2026 the version that makes this fully self-serve: a company can go from zero to a governed, AI-ready data lakehouse by signing up on our website and entering a credit card. No onboarding from our side. No specialized engineers on theirs.
What Anthropic described is the right problem, solved the right way given their constraints. What we’re building is the same outcome — AI that actually works on real business data — designed for the companies that don’t have Anthropic’s team.