Anthropic says Claude now automates 95% of its business analytics queries, with about 95% accuracy overall.
The figures came as the company outlined how it has built an internal self-service analytics system around its large language model, shifting routine work from data science teams to automated query handling. That, it says, has allowed staff to spend more time on causal modelling, forecasting and machine learning.
Anthropic presented the move as a response to a longstanding problem in business intelligence. Traditional self-service analytics has often relied either on making data models more accessible to non-technical staff through broader, less normalised tables, or on creating more ringfenced environments with curated dashboards and metrics.
As organisations grow, Anthropic argues, both approaches can create problems. Wide, denormalised tables can lead to overlapping views and inconsistent definitions, while ringfenced environments can miss less common business questions and contribute to a growing number of dashboards and duplicated metrics.
It also argued that simply connecting a language model directly to a data warehouse can create a false sense of confidence. In its view, the main challenge is not writing SQL but ensuring a user's question is matched to the correct, current entities in the underlying data model.
Anthropic identified three main causes of inaccurate answers from analytics agents: ambiguity between business concepts and the data entities that represent them, stale data or definitions, and retrieval failures in which relevant information exists but is not found by the model. It says these issues account for most of the errors it sees.
Data model focus
To address those problems, Anthropic says it has built what it calls an "agentic data stack". The stack is designed to reduce ambiguity through stronger data foundations and sources of truth, keep models and definitions current through maintenance and validation, and improve the model's ability to find and use the right information through a layer of structured skills.
At the base of that approach is a more tightly governed data warehouse. Standard data engineering practices remain central, including dimensional modelling, earlier testing in the development process, and checks on data freshness and completeness.
What changes, Anthropic says, is the nature of the user. Instead of trained data scientists working directly with the warehouse, the end user is often an automated system acting on behalf of staff with limited understanding of the underlying infrastructure. That means the output must be reliable without depending on the user to verify the logic behind it.
One of the company's main steps has been to create a smaller number of canonical datasets that serve as single sources of truth. The aim is to reduce the number of plausible tables, columns or metric definitions that might answer a question such as revenue or weekly active users.
Those governed models are backed by tooling, continuous integration checks and internal mandates, intended to steer teams toward canonical definitions rather than near-duplicates. Nearly all data code, semantic definitions, documentation and dashboard definitions are held in a single repository, with checks designed to catch changes that would break downstream uses.
Sources of truth
Beyond the warehouse itself, Anthropic says it relies on several reference layers to guide Claude's responses. The first is a semantic layer containing compiled metric and dimension definitions. When a question maps cleanly to one of those metrics, the system can return a number consistent with other reporting surfaces inside the company.
Its agents are structurally required to consult that semantic layer first. Anthropic added that an attempt to generate metric definitions automatically from raw tables and query logs did not work well because it reproduced underlying ambiguities instead of removing them.
Another layer is lineage data and the transformation graph, which helps the model identify which upstream models support a concept, which assets have been deprecated, and which datasets share the same grain. That allows the system to make a governed choice even when a metric has not been predefined in the semantic layer.
The company also discussed its use of historical SQL from dashboards, notebooks and earlier analyses. Although that might seem a rich source of precedents, giving the model direct retrieval access to large volumes of old queries improved accuracy by less than one percentage point, Anthropic said. It found it more effective to distil that material into structured domain documents and repeatable analysis patterns.
A further component is business context. Anthropic says analytics systems can answer the literal wording of a request without understanding the internal references, team definitions or organisational context behind the question. To address that, Claude draws on a company knowledge graph that includes indexed documents, roadmaps, decision logs and organisational information.
Human oversight still plays a central role, the company said. While Claude is used to draft documentation, suggest metric descriptions and flag gaps in coverage, ownership and curation remain with employees rather than the model itself.
That distinction underpins Anthropic's broader view of self-service analytics. "Therefore we recommend generating the documentation with Claude, but having a human own the definition," it said.