There are three types of data questions and AI can really only handle the numbers.

• Tell me a number
• Diagnose a problem
• Tell me what to do

Tell me a number

If you can prompt a question, that question has already been answered (likely gathering dust in a dashboard somewhere). Self-service analytics got close but you still had to go get the number yourself. Now you can have AI just fetch all the numbers.

Diagnose a problem

This is what consumes 60-80% of a data & analytics team's time: what's wrong with my number?

Underneath the ambiguous question are four potential root causes. The first two mean the number is broken. The latter two mean nothing's wrong - it's just disappointing.

• Data - pipeline failure/delay; definitional drift
• Tech - feature release bug; missing event tracking
• Forecast - variance against Finance's plan
• Performance - variance against Operator's expectations

The modern data stack makes data remediation relatively swift and painless.

Tech issues often masquerade as data issues, making fixes a step slower. Persist long enough and tech issues become forecast/performance issues. Engineering teams with a culture of shipping data as product are better at mitigating tech-data issues.

Forecast issues tend to fly under the radar due to their black-boxy nature. Having deconstructed financial models with hardcoded data sources, endless tabs littered with undocumented formulas and color-coded cells that mean something, I can say that this step is ripe for AI disruption.

Performance issues are harder pills to swallow. Stakeholders consume validated findings and craft them into narratives voiced over Zoom or in closed rooms, gone like tears in rain.

Judgment leaves no artifacts.

Answering a diagnostics question is first a process of elimination, then an offering of assurances. It's rarely about delivering a number.

Tell me what to do

And then there are insights. Analysts are expected to deliver "proactive insights". But whether an insight lands has almost nothing to do with the insight. It comes down to timing, understanding and the operator's willingness to act.

I've never been fond of shipping insights. You clean the data, perform the analysis, create a deck with simple language and hope people understand, redo it a few times and pray. You feel like you've delivered value but you're not always sure an action was ever taken. AI takes the grind out of shipping insights. It can't make anyone act on them.

Answering data questions has never been about numbers. It's about instilling conviction in the people whose necks are on the line.

AI doesn't care about any of that. It leaves nothing but artifacts.