# Ask good questions

> How to phrase questions so the agent gives you a specific, usable answer.

The agent works best with clear, specific business questions. Ask what you'd ask a sharp data colleague.

## Be specific about the decision

| Instead of | Ask |
|---|---|
| Tell me everything about my customers | Which customer segments show rising cancellation intent this quarter? |
| What's the sentiment? | What are the top three reasons for one-star reviews in the last 90 days, by store? |
| Analyze the chatbot | Where does the bot hand off to a human most often, and what did customers ask right before? |
| What are people saying about billing? | How do billing complaints differ by plan tier and region? |

## Questions that work well

- *What are the top issues customers contacted us about last quarter?*
- *How often do conversations show churn risk, and what's driving it?*
- *Are competitor mentions increasing over time?*
- *Which product issues drive negative NPS among enterprise accounts?*
- *How does satisfaction differ between annual and monthly plans?*
- *Why is this segment behaving differently from that one?*

Questions about why, about change over time, and about the difference between two groups play to the agent's strengths: it joins what customers said with what they did and tests whether a difference is real. See [How we prove a driver](/causal).


## Give it context

Say who the answer is for and what they'll decide. *"This is for our head of support in North America, who's deciding where to add staff."* The agent shapes the analysis and the write-up to that.

If you give the same context every time, put it in the project's standing instructions instead. See [Projects and standing instructions](/agent/projects-and-instructions).

## Point it at the right data

You don't need table names, but you can steer: *"Use the predicted CSAT Dimension as the main measure."* *"Only use chat, not email."* *"Exclude anything before January."*

Give explicit dates for "this week" or "last month" when you ask late in the day. The agent's clock runs on UTC, which can already be tomorrow in the Americas.

## Start by exploring

If you're new to a dataset, ask what's there first:

- *What's in my workspace?*
- *Show me the tables and what each one holds.*
- *List my prompts.*
- *Which Dimensions have data, and for what dates?*
- *What metadata do I have?*
- *Do we capture competitor mentions anywhere?*

See [What the agent can see](/agent/what-it-can-see) for more.

## Ask, or plan?

For a quick answer, or to explore over a few turns, just ask. For a report, a brief or anything you'll repeat, ask the agent to write an analysis plan first. See [Quick answers and analysis plans](/agent/modes).

## Push back

It's a conversation, not a query. If a number looks off, ask how it was calculated. If a segment is missing, say so. Ask "why" as often as you need. You can refine, follow up, change direction or start over. When you've found something worth repeating, ask the agent to turn the chat into an [analysis plan](/agent/analysis-plans).

## When it says it can't answer

Take the refusal seriously. It names what's missing, such as an empty Dimension table, and offers a way forward. Reply *"use available data"* for an interim answer from the raw text, which will say it's a workaround, or fill the gap and ask again. See [Your first chat](/agent/first-session#4-when-the-agent-doesnt-have-enough-data).

## Name your files

If you've uploaded reference files (a plan, report execution guidelines, last month's report), name each one in your message and the order to read them. The agent plans before it looks around, so a file it isn't told about may not make it into the plan.

## What it won't answer

The agent reasons over your data, not your account. For login, billing, account settings or anything else, email support@dimensionlabs.io.