Conversational BI has huge potential for organizations to make data more broadly accessible & shared. But giving AI tools like Copilot, Claude, or Genie access to all your data is not enough & Anthropic just proved it on their own data. In this opinionated article, we share what they have learned & how we feel it teaches us something about how organisations should invest in Data & AI.
Conversational BI or “AI for BI”
We have been trying to crack the self-service BI topic for years: with tools like Qlik & Power BI business teams could work & play with the data themselves. AI tools like Claude now change this again drastically as: with a connection simple set-up or a download+upload, Claude can analyze your data directly & answer your business questions: “top sales”, “anomalies in the order book”, “finance bookings of last 3 months related to Supplier X on product Y”, etc.
Good side reads: Conversational BI in Fabric | element61 & Conversational BI in Databricks | element61
However, once connected, it’s often not that great… Putting an AI agent on top doesn't magically make our problems disappear. Also, AI tools struggle with our (!) inconsistent definitions, data that’s not up-to-date, undocumented data, multiple versions of the truth and platforms that we can’t fully trust. In reality: AI exposes the problems we know that are there and AI raises the bar for how well we need to build, govern and document our Data Foundations.
What Anthropic learned from their internal Conversational BI projects
Anthropic gave Claude access to thousands of SQL queries, dashboards, notebooks and other analytical artefacts to build its internal AI-powered analytics capability. Accuracy did not exceed 21%.
What Anthropic noticed was that information was not missing: for roughly 80% of the questions answered incorrectly, the information required to answer correctly was available somewhere in the context but not in the data itself nor the data platform. After investing in governed sources of truth, canonical datasets, clear business definitions, semantic context, documentation and reusable skills, accuracy rose to more than 95%.
Access to data is not the same as understanding data.
Just giving an AI tool access to your BI ecosystem won’t allow you to guarantee that AI can answer questions truly correctly. Your data needs more context…
Bart Van Der Vurst - Partner element61
Why this matters for your data platform
For years, a Modern Data Platform has focused on bringing data together and making it available for BI, Analytics and Data Science. AI adds another requirement: it does not just need access to your data; it needs to understand what that data means.
These are the questions an AI agent cannot answer from raw access alone:
- What exactly is an active customer?
- Which revenue definition should it use?
- Which dataset is the trusted source?
- How are customer, product and order related?
- Who owns that definition and what happens when it changes?
Data Platform and Data Governance: two sides of the same coin
A modern Data Platform and Data Governance are increasingly two sides of the same coin. A platform can technically bring all your data together, transform it and make it available at scale but that doesn't automatically make the data understandable or trustworthy. You still need to know what a metric means, which source is authoritative, who owns it, whether the data is of sufficient quality and how it should be used. That context has always mattered for BI, but it becomes even more critical when an AI agent is consuming the data autonomously.
The opposite is equally true. Data Governance without a strong Data Platform risks becoming documentation on the side. A beautiful catalogue or business glossary has limited value if those definitions aren't connected to the actual datasets, transformations, lineage and quality controls underneath.
The real opportunity is to bring both together: a Data Foundation where the platform provides the trusted data and governance provides its meaning, ownership and context: making data not only technically available, but genuinely understandable to both humans and AI.
Each governance building block answers a question AI needs answered before it can reason reliably.
| Building block | What it can help provide to the AI |
|---|---|
| Data Platform |
Trusted data, technically available |
| Data Catalog |
Discoverability: what exists and where |
| Business Glossary | The language of the organisation |
| Data Lineage | Where the information comes from |
| Data Quality | Whether it can be trusted |
| Master & Reference Data | Consistent meaning of customer, product, supplier across systems |
| Semantic Layers & Ontologies | Machine-readable definitions, so humans and AI agents reason over the same business concepts |
In an AI-first organisation, context data (metadata) becomes part of the product
Documenting data, defining business concepts and maintaining a catalogue have sometimes been treated as administrative work around the “real” data program. That view is becoming outdated: the next consumer of a Data Platform may not be a person reading a Power BI dashboard. It may be an AI agent selecting a source, interpreting a metric and taking action based on the result. In that world, metadata is not an appendix: it is part of the operational product and potentially the most important thing to have.
What this means for your Data & AI roadmap
This isn’t revolutionary, but things you might have considered “nice to have” like documentation or enriched metadata might suddenly become crucial to succeed with AI: what used to be good practice now decides whether an agent returns a right or a wrong answer.
- Treat metadata as production infrastructure. Definitions, lineage, ownership and quality signals need the same maintenance discipline as pipelines and code: versioned, tested, monitored and owned by someone with a name. In practice this means metadata earns a place in your release process instead of living in a side document that gets refreshed once a year. When a definition breaks, that should be as visible as a failing pipeline.
- Create governed paths to trusted answers. An agent that treats every table, dashboard and notebook as equally valid will sooner or later pick the wrong one. Guide it towards canonical datasets, approved metrics and reusable analytical skills, and make everything else explicitly secondary. The point is not to restrict curiosity, but to make sure the shortest path is also the most reliable one.
- Make business meaning machine-readable. Semantic models and ontologies translate organisational language like an active customer, net revenue or churn, into context that AI can apply consistently. A glossary in a document helps people; the same glossary exposed through a semantic layer helps people and agents at once, and keeps them reasoning over identical concepts.
- Design for change. Definitions, schemas and policies evolve, and every change ripples through the answers your agents produce. Embed maintenance and validation in the operating model: who approves a change, how it is versioned, how consumers are informed, human and machine alike. Governance that quietly assumes stability decays as soon as the business moves.
- Measure answer quality, not only technical success. A query that runs is not necessarily a correct answer. Evaluation has to test business interpretation, provenance and consistency: does the agent choose the authoritative source, apply the agreed definition and return the same answer twice? A recurring benchmark of real business questions is the most honest progress indicator you can give yourself.
Taken together, these five priorities describe a shift in ambition: from making data available to making data understandable. That is the bar AI has raised, and it is where the return on your Data Foundation is now decided. The good news: tools like Microsoft Fabric & Databricks have templates for this & it’s just up to you (and us together) to implement these together!
In summary: AI raises the bar
AI does not reduce the need for a strong Data Foundation. It raises the bar for what a good Data Foundation needs to be.
The next consumer of your Data Platform may not be a human looking at a Power BI dashboard: it may be an AI agent deciding which data to use, interpreting it, and acting on it. The quality of your documentation is now directly related to the quality of your AI.
This is also why AI Governance and Data Governance are converging: reliable AI requires more than controlling models. It requires trusted data, clear ownership, shared definitions and a semantic foundation underneath.
Want to continue reading
Dive into the Data Governance content on our website: Data Governance | element61
Talk to us: element61 helps organisations build the Modern Data Platform, Data Governance, Data Quality, Master & Reference Data and AI Governance foundations that make AI reliable.
This perspective was inspired by Anthropic’s article “How Anthropic Enables Self-Service Data Analytics with Claude” (June 2026), Available here.
FAQ
Conversational BI, or "AI for BI", lets business users ask questions of their data in natural language instead of building reports. With a simple connection or a data upload, tools like Microsoft Copilot, Claude or Databricks Genie can analyse data directly and answer questions such as "top sales", "anomalies in the order book" or "finance bookings related to Supplier X". element61 implements Conversational BI on Microsoft Fabric and Databricks.
AI does not reduce the need for a strong Data Foundation; it raises the bar for what a good one must be. AI agents expose the problems organisations already know exist: inconsistent definitions, outdated data, undocumented datasets and multiple versions of the truth. Where a human analyst compensates with tacit knowledge, an agent simply returns a wrong answer.
Anthropic gave Claude access to thousands of SQL queries, dashboards and notebooks to build an internal AI-powered analytics capability. Accuracy did not exceed 21%. For roughly 80% of incorrectly answered questions, the required information existed somewhere in the context but not in the data or platform itself. After investing in governed sources of truth, canonical datasets, business definitions, documentation and reusable skills, accuracy rose above 95%.
Access to data is not the same as understanding data. An AI agent with raw access cannot tell what an "active customer" is, which revenue definition applies, which dataset is the trusted source, how customer, product and order relate, or who owns a definition when it changes. Without that semantic context, an agent will confidently select the wrong source and produce a plausible but incorrect answer.
Data Foundation is the combination of a modern Data Platform and Data Governance: the platform provides trusted data at scale, while governance provides meaning, ownership and context. Together they make data not only technically available, but genuinely understandable to both humans and AI agents. Neither half delivers reliable AI on its own.
They are two sides of the same coin. A platform can bring data together, transform it and serve it at scale, but that does not automatically make it understandable or trustworthy. Conversely, Data Governance without a strong platform becomes documentation on the side: a catalogue or glossary has limited value if its definitions are not connected to the actual datasets, transformations, lineage and quality controls underneath.
Each building block answers a question AI needs answered: the Data Platform provides trusted, technically available data; the Data Catalog provides discoverability; the Business Glossary provides the organisation's language; Data Lineage shows where information comes from; Data Quality shows whether it can be trusted; Master & Reference Data gives consistent meaning across systems; and Semantic Layers & Ontologies make definitions machine-readable.
Documenting data and maintaining a catalogue were often treated as administrative work around the "real" data programme. That view is outdated: the next consumer of your Data Platform may not be a person reading a Power BI dashboard, but an AI agent selecting a source, interpreting a metric and acting on the result. In that world metadata is operational product, not an appendix.
We recommend five priorities: treat metadata as production infrastructure; create governed paths to trusted answers; make business meaning machine-readable through semantic models and ontologies; design for change with clear approval and versioning; and measure answer quality, not only technical success. Together they mark a shift from making data available to making data understandable.
A query that runs is not necessarily a correct answer. Evaluation must test business interpretation, provenance and consistency: does the agent choose the authoritative source, apply the agreed definition, and return the same answer twice? A recurring benchmark of real business questions is the most honest progress indicator an organisation can give itself.
Reliable AI requires more than controlling models. It requires trusted data, clear ownership, shared definitions and a semantic foundation underneath, because the quality of your documentation is now directly related to the quality of your AI. element61's approach brings Data Governance, Data Quality, Master & Reference Data and AI Governance together into one operating model.
Start by benchmarking real business questions against your current platform to see where answers break down, then prioritise canonical datasets, definitions and semantic layers. Tools such as Microsoft Fabric and Databricks already offer templates for this. element61 helps organisations build the Modern Data Platform, Data Governance, Data Quality, Master & Reference Data and AI Governance foundations that make AI reliable, get in touch.
