Nobody has this fully figured out yet. So we are working it out in public, together.
If you have ever waited a week for a number, and then waited again because the first one turned out to be wrong, you already know why we are here.
For years, getting an answer from your data meant joining a queue. You filed a ticket, you waited, and a number came back that you mostly trusted. The data team was not slow on purpose. They were buried. Everyone could see the system was broken, and nobody had a better one.
Analytics agents finally offer a better one. A person asks a question in plain words, and an agent does the work a careful analyst would do. It reads the question, writes the query, runs it, checks the result, and answers. When it works, the queue disappears and the data team gets to do real work again.
But an agent is only worth having if it is right, and being right is the hard part. It takes the business knowledge that lives in your data team's heads, the rules, the edge cases, the quiet history of what broke and when. This is a community for the people doing that work.
- 01
Reliability is the whole game.
An agent that is sometimes wrong, with no way to know when, is worse than no agent at all. People trust it, get burned once, and never come back. Trust is not a feature you bolt on at the end. It is the thing you are building.
- 02
Context beats model size.
Agents rarely fail because the model is weak. They fail because they do not know your business, that revenue is booked net of refunds, that the EU table is the one everyone actually uses, that last March had a data outage. The real work is teaching the agent what your team already knows.
- 03
The data team owns the context.
The people who know what the numbers mean should decide how the agent reasons about them. Not a vendor, not a sealed box you cannot inspect. When the data team owns the context, they own the quality, and that is exactly where it belongs.
- 04
We build in the open.
This is too important to work out behind closed doors. We share the methods, the benchmarks, and the things that did not work. A dead end mapped in public saves the next team from walking into it.
- 05
Self-serve only counts when it is correct.
Handing everyone an answer is easy. Handing everyone the right answer is the bar. A confident wrong number does more damage than a slow correct one, so we hold that line, even when it is the harder one to hold.
None of this is finished. We are all early here, figuring it out in public and changing our minds as we go. If any of this sounds like the work you care about, pull up a chair. You are already one of us.
Claire, Christophe, and the agentic analytics community
Started by the team at nao. Written to be argued with, so tell us where we got it wrong.