AI agents won't replace your analysts — they'll kill the ad-hoc queue

Every data team I've worked with has the same silent tax: the ad-hoc queue. A Slack channel, a ticket board, a DM that starts with "quick question" and ends with a half-day rebuild of a number someone already had last quarter. It's the single biggest drain on analytical talent, and almost nobody measures it.

This is the queue AI agents are about to drain — and why the "agents will replace analysts" framing gets it exactly backwards.

The fear is aimed at the wrong job

An LLM with tools can take "what was net revenue retention for enterprise last quarter, by region?" and return a number with a chart. That looks like an analyst's job. It isn't. It's the lowest-leverage part of an analyst's job — the part that's repetitive, well-specified, and already answerable from existing models. Handing that to an agent doesn't remove the analyst; it removes the interruption.

What's left over is the work that actually moves a business: framing the question nobody thought to ask, deciding which metric is the honest one, pressure-testing a result before a VP acts on it. Agents are bad at all three, because all three require judgment about context the data doesn't contain.

What agents are genuinely good at

• Translating intent into queries against a well-defined model.
• Iterating fast — "now break that out by cohort, last 8 quarters" without a context switch for a human.
• Explaining themselves — surfacing the SQL and the definitions they used, so the answer is auditable.
• Being available at 11pm to the salesperson who'd otherwise file a ticket.

That's a real productivity unlock. It's also a narrow one.

What still needs a human

The moment a question is ambiguous, political, or novel, the agent's confidence becomes a liability. "Active user" has six definitions in most companies; an agent will happily pick one and never tell you it mattered. A human knows which fight they're walking into.

So the org chart doesn't shrink. It shifts: fewer hours spent pulling numbers, more spent on the modeling, definitions, and decisions that make the numbers trustworthy in the first place.

The bottleneck was never headcount

Here's the part teams miss when they rush to bolt an agent onto their warehouse: the constraint on self-serve analytics was never the interface. It was trust. People file ad-hoc requests because they don't believe the dashboard, or can't find it, or know the definitions drift. An agent pointed at a messy schema doesn't fix that — it industrializes the confusion, at the speed of autocomplete.

The teams that win with agentic analytics will be the ones who did the unglamorous work first: a clean model, governed definitions, a semantic layer the agent can actually reason over. That's the real project. The agent is just the new front door — and it's only as good as the house behind it.

That house is what the rest of this series is about.