What Do We Mean By Agentic Analytics

This article is one of a series looking at using ClickHouse for AI. Articles in the series include "Why ClickHouse Is The Best Data Platform For Agentic AI", "What Do We Mean By Agentic Analytics", "Using Open Source and Open Weight LLMs With ClickHouse" and "Self Hosting An Open Source Model With ClickHouse".

Agentic analytics involves using AI to ask questions about data. It can be thought of as a better alternative to business intelligence, where instead of using dashboards and reports, we allow users to ask questions about their data conversationally and in natural language.

This is a very appealing proposition for businesses and is generating considerable interest. For decades, businesses have been trying to deliver self service analytics but never really achieved this at scale. Now, the promise that people can literally ask whatever they want and perform all kinds of arbitrary investigations on demand and without any specific data skills. Anyone in the business can start with one question, follow up, change direction, and pursue a train of thought without anyone relying on anyone to build a new report. This exploratory quality is powerful because it means the questions you can ask are no longer constrained by what someone anticipated and built in advance.

More Than Text-To-SQL

Early versions of this approach were essentially "text to SQL", where we asked a question, translated it to SQL, ran that SQL against the database and returned results.

Today, we have moved from this simple translation towards agents which do much more. The agent plans, decomposes a question into multiple steps, runs intermediate queries, inspects the results, and decides what to do next in a loop. This allows the agent to answer more complex questions that don't map neatly onto a single query. This dramatically improves the outcomes that can be achieved with agentic analytics.

A Strategy Partner

A point I often make to my clients is to think of agentic analytics as more than just an agentic querying tool. We can use the same interaction to perform what if analysis (what if sales decrease by 10%), or to use it as a strategy development and business partner (which should I open a new store) based one what it finds in the database. This is a very powerful use case for agentic analytics which people haven't quite taken advantage of yet.

The Need For A Semantic Layer

Knowing that a table is called orders with a column called amount is not the same as understanding what "revenue" means in your business, for instance whether it should be net of refunds, or which currency conversions apply. A lot of the hard work in agentic analytics is bridging this gap between the raw schema and the business meaning on top of it. This is achieved using tools such as the semantic layer, skills, documentation, fine tuning and good database design practices. We need to give the agent the tools to be successful and return accurate answers which accurately reflect the users intent.

Democratising Access To Data

Perhaps the biggest shift is who gets to ask questions. Historically, getting an answer out of the data often meant queuing behind a data team or knowing SQL yourself. Agentic analytics lowers that barrier dramatically, letting anyone in the business interrogate data in plain language. Done well, this is genuinely transformative for how organisations make decisions and moving towards a data driven culture.

Accuracy And Trust

It is relatively easy to connect an LLM to a database and expose agentic analytics. However, the key challenge is in building accurate, trustworthy and governed responses. A confident, well-formatted answer that is subtly wrong is arguably more dangerous than no answer at all, because people will act on it. An agent that can check its own work, show the SQL it ran, explain its assumptions, and let a user drill into the underlying numbers is far more valuable than a black box.

Because you are giving a much wider group of people the ability to run queries, you also have to think carefully about safety and control. That means enforcing access permissions, limiting the cost and scope of queries an agent can run, keeping an audit trail of what was asked and answered, and preventing sensitive data from leaking.

Why ClickHouse Is A Good Fit For Agentic Analytics

In a previous article I explained how ClickHouse is fast, concurrent and scalable to the needs of agentic analytic workloads. These needs are particularly important for the agentic analytics use case where one innocuous question could lead to tens of concurrent queries, some of which might be unoptimised or heavy on the database.

If a business finally achieve their aim of moving to a data driven culture internally, and expose analytical agents to their end customers, we could be in a situation where thousands of people are issuing thousands of questions which turn into millions of queries. It is critically important that the data platform at the heart of the solution can stand up to a workload like this at an acceptable price/performance.

In the next article in this series, I look at how we can pair ClickHouse with open source and open weight models served from third party inference providers such as Fireworks.ai and Baseten.