With groundbreaking tech comes FOMO. A vendor shows you a demo. A peer at another company mentions they're "exploring AI agents." Someone in a Slack community posts about asking their data warehouse questions in plain English and getting instant answers. The pressure to act builds fast. So your team or director, or CEO, starts asking: "Can we deploy agents?”
Well, not yet. Here are some reasons why.
- Your metrics aren't defined in one place. Ask three people what "active customer" means. You'll most likely get three answers: last 30 days, last 90 days, and made a purchase. The product team's definition doesn't match the one sales uses. This is manageable when humans are analyzing because they know how to ask. An agent doesn't know how or who to ask. It picks one definition and runs with it, usually without telling you which.
- Your business logic lives in people's heads. Some data team has rules that exist nowhere in the codebase. Always exclude test accounts. The first of the month is unreliable because of how billing posts. Refunds above a certain threshold get categorized differently. These are obvious to anyone who's worked the data long enough. They are completely invisible to an agent unless someone has written them down somewhere it can actually read.
- Your documentation is written for humans, not Agents. Open most dbt projects and look at the schema files. You'll find columns with no descriptions, or descriptions that restate the column name — customer_id: the ID of the customer. That was fine when analysts were the readers. Analysts can fill gaps with inference, ask questions, and cross-check. An AI reader treats your documentation as authoritative. Whatever's missing, it fills in silently with its best guess.
- Your data models aren't clean enough. Raw tables with cryptic column names. Joins that only one person fully understands. Five different views that each calculate revenue slightly differently. An agent pointed at this will produce queries. They will be wrong in ways that are hard to catch.
- You have no way to measure whether it's working. Most teams test the agent a few times, it looks okay, and they ship it. Then accuracy drifts once definitions change, source systems update, and users ask questions no one anticipated. If you're not running it against verified questions regularly and reviewing failures, you're not evaluating it. You're just hoping.
- Your leadership has bought into the vision, not the work. Saying "we need Agents" is easy. It costs nothing and signals the right things in the right rooms. What it rarely comes with is a restructured budget, updated team mandates, or any real accountability for making it happen.
How Far Back Are You?
The answer isn’t the same for everyone.
Some teams are close. The data foundation is solid, the models are clean, but the documentation is thin, and the metrics aren't formally defined anywhere. A few focused months of work, and an agent becomes viable.
Some teams are in the middle. A BI layer exists, dashboards are trusted, but nothing underneath is modeled consistently, and the business logic is still scattered. The gap is real but closable.
And some teams are further back than they realize. The warehouse exists, but it's mostly raw tables. Metrics differ by department. No one owns the definitions. An agent built on top of this wouldn't fail dramatically, but it would just quietly produce numbers that nobody can fully trust or explain.
Knowing which situation you're in is the first honest step.
The Honest Takeaway
The teams getting real results from analytics agents aren't using better models. They're teams that did the less glamorous work of first cleaning up data models, writing documentation that actually explains things, centralizing logic that was previously scattered, and defining metrics that the whole organization agrees on.
The question isn't whether you should build an analytics agent. Eventually, yes. The question is whether you've done the work that makes one actually useful.
For most teams, honestly, that work is still ahead of you. And it has a starting point. It just isn't an agent.
.png)
