Self-service analytics powered by AI promises to revolutionize the way decision makers across every enterprise gain a deep understanding of the state of their business. Until now, answering questions like “How is our customer churn developing in the last quarter?” or “Why did our production volume decrease?” required complex chains of custom Excel reports, dashboards and data aggregation pipelines on top of data warehouses. And every one of those elements in the analytical tool chain required careful, brittle synchronization between business and technical users to ensure every calculation and visualization is aligned with how the business defines “customer churn” or “production volume”.

With AI, self-service analytics has the chance to become truly self-serve for every user across the enterprise. Just ask any question about any metric or KPI in natural language and the AI will provide a reliable metric analysis and even recommendations on how to improve the metric further.

Fulfilling this vision is not easy though. Anthropic finds that by simply connecting their data sources with the latest AI models, the answer accuracy is only 21% - which is completely unacceptable for reliable adoption across the enterprise.

Every business rolling out self-service analytics with AI is currently facing this same challenge. And when trying to increase answer accuracy a list of common questions arises:

1. Why can’t you just connect your AI to your data sources?
2. Why can’t you just add some Slack, Google Drive or SharePoint connectors?
3. Why can’t AI build up the context itself from your databases?
4. Why can’t you just write some skill files?
5. Can you create your semantic layer once and be done with it?
6. Do you need to buy new data platforms or AI tools?
7. Ok, it seems you do need a semantic layer. How do you add that to your existing AI and data landscape?

Below we will answer each of those questions, leaning heavily on the important findings shared by Anthropic in their article. Every answer will get us closer to the understanding of how every enterprise can reach the 95% answer accuracy Anthropic was able to achieve for their AI-powered self-service analytics.

Why can’t you just connect your AI to your data sources?

Anthropic’s article clearly shows what happens when you connect your AI agents with your databases without any extra scaffolding: a 21% answer accuracy.

The answer to this question is that databases often just don’t contain the necessary context. A database column might be called “revenue_final” inside a table “finance_monthly_reporting”. Nowhere does it say how this revenue number treats returns or trial customers. And nowhere does it say whether this is the right table to pull monthly reporting numbers from in the first place. You will need to provide clear definitions for both to your AI agents to ensure correct answers to the question “How did revenue trend in the last quarter?”.

Duodata allows you to cleanly capture exactly those two important context elements - what does each metric mean and which is the canonical source system, table and column or field. We call this the Metrics Ontology.

Why can’t you just add some Slack, Google Drive or SharePoint connectors?

Your intuition is right: Slack, Google Drive and SharePoint do contain important context for your AI agents. The trouble is that this context is not curated. It may contain outdated or conflicting information.

So when leveraging your existing work output as context source, a two-step process is best: let AI comb through all your content across Slack, cloud storage etc. and create candidate metric semantics for your semantic layer. Then a human will need to review those drafts, add or correct definitions and sign them off.

Duodata is the right place where your AI agents can create proposed metric definitions and your business users can easily review, edit and approve them before the metric becomes part of your Metrics Ontology.

Why can’t AI build up the context itself from your databases?

Many data platforms start to offer the feature of automatically creating semantic layers from the data platform schemas. Anthropic tested that approach and it actually reduced analytics query answer accuracy!

The reason for that is that - as outlined in the question about querying databases with AI directly above - the data schemas don’t contain the necessary business meaning to enable the correct understanding of business metrics.

Whether sales commissions are included or excluded in “gross profit” is something that needs to be written down by the very humans who took the business decision of whether that is the case or not. This can’t be inferred from a database schema.

Duodata is supporting such an approach. Business users create and approve the definition of every metric in your Metrics Ontology. Duodata agents then project those definitions tied to the right table and column into your data platforms to ensure your AI can talk to your data.

Why can’t you just write some skill files?

Anthropic shows how important it is to write good skill files. But likely not in the way you think.

Early analytics agent pilots tend to use skill files as a hotch-potch of duct tape, business semantics and data normalization in order to make the agents answer a set of pre-canned analytics queries reliably. This works for a quick pilot, but not for production deployment.

As you can see in the Anthropic article, a good skill file consists of a small set of clear instructions around which methods and context sources to use for every query. Every context source itself must live outside of your skill file. First, you want to separate your skill files, which evolve with model changes, from your business context, that holds true no matter which model generation is working with your data. In addition, your business users need a clean user interface to create and maintain their metric meaning definitions together with governance and approval workflows instead of fighting with thousands of lines of complex free text.

Duodata is the right place to create and maintain this model- and platform-agnostic semantic layer. Your agents can then easily leverage the Metrics Ontology either via MCP or via semantic layers that Duodata projects across your data platforms and source systems.

Can you create your semantic layer once and be done with?

No, and Anthropic's numbers show how fast your semantic layer decays. Their answer accuracy dropped from around 95% at launch to around 65% within a month. Data sources, schemas and business definitions change constantly, and every change turns a correct definition into a subtly wrong one. Anthropic’s fix was to make the semantic layer a living asset: definitions live next to the data models they describe, a change to one triggers a review of the other, and scheduled agents scan for wrong answers and propose corrections that a human signs off.

This is exactly what Duodata is built for. Business users can easily make updates to existing metric definitions or add new ones to the Metrics Ontology. And every metric definition is linked to its source table and column, allowing Duodata agents to automatically detect when data schemas start to diverge from the metric definitions.

Do you need to buy new data platforms or AI tools?

No. As Anthropic shows in their article, they are leveraging their existing Claude AI as the interface and reasoning layer. And they are accessing their existing database and data platform landscape. What they are inserting in between those is a thin but well-thought-out set of instructions and most crucially a human-curated semantic layer.

We at Duodata believe that you should use your existing AI tooling and your existing data sources - connected with a Metrics Ontology as the semantic layer - to enable self-serve analytics. Your users will not need to learn yet another new tool. And your IT department will thank you for not introducing yet another element into the already complex IT landscape.

Ok, it seems you do need a semantic layer. How do you add that to your existing AI and data landscape?

Anthropic shows how incredibly simple the integration of a human-curated semantic layer is into AI-enabled analytics workflows. All it takes is a set of simple instructions for the AI model inside the skill file to always consult the semantic layer first before answering any query from the existing data sources. No big migration or engineering required.

Duodata offers an MCP interface that is easy to use for every agent via exactly such a simple instruction set in your skill file. Duodata also maintains semantic layers across your data platforms - in that case you don’t even need a single change in your skill file, since your AI agents will access your data platforms via the Duodata-created semantic layer.

How to get started with AI-powered self-service analytics

In summary, we at Duodata are thrilled about the prospect of AI agents becoming fully capable coworkers in the enterprise, making every employee and business more productive.

In order to achieve that goal, the first step is to give every agent a deep and correct understanding of every business metric and where to find the right data source for each metric.

This is what Duodata delivers. A powerful, easy-to-use interface for business stakeholders to curate a Metrics Ontology and for IT to link it to the right source. A MCP interface to ground every agent in the correct business understanding. And agents that maintain the semantic layers across all data platforms.

If you want to bring your self-service analytics into the AI age, start creating your own Metrics Ontology on https://app.duodata.ai/ or reach out at contact@duodata.ai.