What is agentic analytics?

It was inevitable – the alliterative allure of “agentic analytics” was just too good, and it is on its way to becoming the latest big buzzword in data. Unfortunately, most of the posts I’ve seen purporting to explain this feel anodyne, thin, or are just pitching a product. ^[1]

Despite seeming like empty jargon, however, there is a core idea and shift in how AI agents can help us accomplish analytics tasks, and it’s worth really understanding what this means and where it’s going to take us.

So, I want to share a bit of how we talk about this internally at Hex, but completely separated from Hex itself. This isn’t a sales pitch – just sharing some mental models everyone can use as they think about where this is going.

The three horsemen of agentic analytics:

We think about this along three dimensions of change: depth of insight, breadth of reach, and speed of insight. These are things everyone has wanted for a long time, and have long been promised by data tools – but with AI we can finally get them.

Depth

This is the most important change: deeper, more useful output beyond single node insights (and the corresponding fade of legacy dashboarding).

Up until now, most data tools have been built around “single layer” assets – basically dashboards with a bunch of tiles, each showing one metric. More in-depth workflows (i.e., via a notebook) were harder and less accessible, and therefore rarer – meaning most analytics just describe “what happened”.

Today, though, it’s way easier to go way deeper. The fundamental property of “agentic” AI is that it can work step-by-step, reflecting on what came before, and planning what to do next. This is perfect for more in-depth analysis workflows – driving to more detailed insights and causal understanding.

As an example, an agentic system can run a first query on what happened, look at the first few rows, add a chart, look at it, decide to run another query, join the results together, do an exploratory query to look at trends, and then decide to summarize the results to the user. It can reach for tools – like schema inspection, semantic search, or world knowledge – reasoning about what to do with them along the way.

This capability of being able to go deeper and get way better answers is going to quickly eclipse the dashboards of yore– and ultimately become the predominant way of interacting with data.

Breadth

But it’s not just current “data people” who will be able to go deeper – way more people will have access to analytics than ever before.

Think of all the decisions made every day in your organization. Think about how many of them were rooted in real, insightful analysis… yeah, not that many, right?

But now, anyone can ask a natural-language question and get a first draft. And what happens when the friction to get data answers basically crashes by 100x?

First, you are going to get a lot more questions answered! The breadth of who can drive insights will go up way, way higher – the dream of self-serve will finally be fulfilled.

But second, you’re going to have the data team’s role shift toward specialist analysis, charting new territory, and curating the context so they and everyone else can answer data questions. The time spent on “QQs” will stop, and the time spent on more thoughtful, focused work will spike. It’s a good thing for everyone!

And agentic systems will be able to reach users in more places – not just the “BI tool”. Imagine being able to @-tag an agent in Slack, or initiate a request from another system via an API. These agentic systems can become “infrastructure for insight”, using their context and tools to answer questions for people who aren’t interacting directly with the software itself.

Speed

In the old world, everything is slow.

If you’re doing insights with “self-serve”, you’re clicking through charts, finding data, and waiting for things to load. And if you’re sending something to the data team – well, they’re probably going to do a deeper and better job, but you’re going to be waiting for it.

In fact, this is the single biggest problem we hear about when we start engaging with a customer – the data team is moving too slow (and they know it!)

I actually think this is worse than people give it credit for initially – it’s not just that the answers you get back are slow, but it’s that people stop asking because they don’t want to wait. Every day many (most? almost all?) data questions that could be answered aren’t – in part because it’d just take too long.

In the agentic era, however, this will change.

Well, in one way, things may actually feel… slower? Remember the “depth” we talked about above? That takes time: it’s a lot of tokens through the AI agent, and because the insights are more in-depth, they can take longer to run.

But in reality, everything will actually be much faster! The full path to get to that insight will happen much more quickly than it’d be for a human to click and type it all, and you’re probably avoiding a lot of back-and-forth on follow-up questions. We’ll also see more adoption of “async” patterns, where you’ll give an agentic system a task, let it chug, and it’ll let you know when it’s done, freeing you up to do other stuff.

Even better – as an agentic system comes to learn more about users and teams, it can surface things proactively, basically moving time-to-insight negative.

Agentic analysis is real, and it’s starting to work

2025 is the year that AI is starting to change data, and we’re going to start feeling it along all three of the dimensions above. It’s going to be exciting, confusing, and fun – if teams approach it with open minds. ^[2]

^[1] or patently written by an AI, which I guess you could argue demonstrates true commitment to the bit!

^[2] This is usually the part where I’d pitch you on why Hex is the best solution for this, but I’ll save that for another post...