The four levels of agentic analytics, and why most products stop at two

By Meet Patel · 2026-09-06 · updated 2026-09-12

Summary

Agentic analytics has four levels. Level 1 is query: natural language to SQL. Level 2 is investigation: decomposing a question and following a trail, still human-initiated. Level 3 is initiation: deciding what to investigate without being asked, which requires a model of what matters. Level 4 is disposition: carrying findings to an owned decision and reviewing the outcome. Most products stop at level 2, and the barrier to level 3 is organisational, not a limit of model capability.

Key Metrics & Takeaways

4 levels
query, investigation, initiation, disposition
Level 2
where most products stop
134 investigations
run in the first ten weeks of my own product

Every product in this space now says agentic, and the word has stopped carrying information. So here is a ladder, with the rungs named, so that a claim can be checked.

Four levels. Each one requires something the level below it did not.

Level 1 — Query

You ask a question in English. The system writes the SQL, runs it, returns the answer.

This was hard and is now largely solved. The remaining problems are semantic rather than linguistic: the model needs to know that active customer means one thing in your company and something else in mine, and that knowledge has to come from somewhere other than the model.

Nothing here is agentic. It is a very good translator, and calling it an agent is where the word started to rot.

Level 2 — Investigation

You ask why revenue fell. The system does not run one query. It decomposes: which segment, which stage, new business or renewal, which region, is it volume or price. It runs a dozen queries, discards eight, follows the two that look live, and returns a chain of reasoning rather than a number.

This is genuinely different and genuinely useful. It compresses the thing a good analyst does over an afternoon into a few minutes, and it does not get bored on the fourth branch.

It is also where most products stop, and stopping here is a defensible commercial choice. It demos beautifully, the failure modes are visible, and a human is in the loop at the start of every run so nobody has to think hard about trust.

But look at what has not changed. Nothing happens until someone types. The system has no view about what deserves looking at. Its judgment is fully downstream of somebody's curiosity, and curiosity is not evenly distributed across a week or across a team.

Level 3 — Initiation

Level 3 is when the system decides, on its own, that something is worth investigating.

Nobody asked. Nothing broke. A pattern that usually holds stopped holding, and the system judged the deviation material enough to spend compute on and, eventually, attention on.

This is the wall. And the reason is not that models are not good enough — they are, and have been for a while.

The reason is that initiation requires a model of what matters to this company. Materiality is not a property of data. Fifteen per cent movement in a segment is an emergency in one company and Tuesday in another, and the difference is not in the numbers, it is in strategy, margin structure and what the team is currently trying to prove.

Most companies have never written any of that down. It lives in the head of whoever has been there longest. So the honest sequence for building level 3 is that you have to extract the company's own judgment first, and that is slow, political, and nothing like an engineering problem.

The second obstacle is trust in the negative case. A level 2 system is judged on the answers it gives. A level 3 system must also be judged on the weeks it says nothing, and to trust that silence you need to know what it looked at and rejected. Which means the record of the rejections is a product surface, not a log file. I did not understand that at first and built it as a log file.

Level 4 — Disposition

Level 4 carries the finding to a decision that has an owner and a deadline, and then goes back later to see whether the decision worked.

Not a notification. A disposition: this is what changed, this is what we think it means, this is our confidence, here is who has to decide, here is by when, and here is what we will check afterwards to find out whether we were right.

Almost nothing does the last part, because it means keeping a public score. If the system recommended eleven things last quarter and four of them were acted on and two of those helped, that is an uncomfortable table to render inside your own product.

It is also the only mechanism by which any of this improves. Without an outcome loop you have a system that generates recommendations at a constant quality forever, and confidence that grows for no reason.

What the ladder is actually measuring

Read the four levels again and notice what changes between them.

Level 1 to 2 is a capability jump — the system can do more. Level 2 to 3 is a responsibility jump — the system now decides what is worth your attention. Level 3 to 4 is an accountability jump — the system's judgment now has a score attached to it.

Only the first is about the technology. Which is why the products all cluster at 2: it is the last rung you can climb without asking a company to change how it works.

My own numbers here are small and I would rather publish them than imply otherwise: one hundred and thirty-four investigations in the first ten weeks. Not a proof of anything. But the pattern in them was consistent and it was not what I expected — the failures were almost never bad analysis. They were investigations that were perfectly well executed on a question that did not matter that week.

That is a level 3 problem, and no amount of improvement at level 2 touches it.

Frequently asked questions

What is agentic analytics?

Agentic analytics is analytics performed by software that can plan and carry out multi-step investigation rather than return a single query result. In its fuller form it also decides what to investigate without being prompted, and carries the finding to a decision with an owner. Most products currently on the market do the multi-step investigation but wait to be asked.

How is agentic analytics different from natural language querying?

Natural language querying translates a question into SQL and returns a result. Agentic analytics decomposes a question into sub-questions, follows evidence between them and revises its own path. The larger difference is initiation: a querying tool never starts, while a genuinely agentic system decides on its own that something is worth looking at.

Why do most agentic analytics products stop short of full autonomy?

Because the blocker is not model capability. To decide what to investigate unprompted, a system needs an explicit model of what matters to that company — thresholds, ownership and materiality — and most companies have never written those down. The missing input is organisational judgment, not intelligence.

Written by Meet Patel — founder of Company 8, building Dan (usedan.com). Dubai, UAE.

Read on themeetpatel.com