# Your first chat

> Open the agent, ask your first question, read the answer, and recognize when the agent needs more data.

This page walks through a first chat with the agent: what you see when you open it, what happens when you ask a question, how to read the answer, and what it looks like when the agent doesn't have enough data.

## Before you start

The agent works from the data in your workspace, so it needs something to read: one file of your own or a sample dataset. If you haven't loaded either, follow the [Quickstart](/get-started/quickstart).

There's nothing else to set up. You don't need Dimensions, plans or instructions for a first chat. You may soon want them, and the agent will tell you when they'd help.

If you came in through sign-up, your first chat already has the question you wrote as its first message, and the agent is at work on it. Every chat you start after that opens empty.

## 1. What you see when you open the agent

Four things are worth knowing before you ask anything.

| Part | What it does |
|---|---|
| **The chat input** | The *Message Agent...* field at the bottom. Type a question in plain English and press Enter. |
| **Chat history** | The left rail lists your past chats. In a new workspace it holds only your first chat. |
| **File Manager** | Opens from the message box. It shows your project's files, including finished reports in **Final Reports**, and has the button that publishes a file to **Artifacts**. |
| **Settings and voice** | The other icons under the input open the agent's settings and turn on voice input. **Prompt Designer** is one of the agent's settings. You can skip them in your first chat. |

The **Help** button in the app's left navigation opens these docs.

## 2. Ask a real business question

Type a question about your business and send it. For example:

> What is driving customer dissatisfaction?

The agent works through the request in four stages, and you can watch each one:

1. **It makes a plan.** Before answering, it lays out the steps it intends to take as a short task list: what it will inspect, what it will query, and what it will pull together at the end.
2. **It inspects your data.** It checks which tables, prompts and Dimensions are populated in your workspace, and what dates they cover. This is where it works out whether your question can be answered yet.
3. **It runs queries and pulls evidence.** It writes and runs the queries it needs: counts, filters, aggregations, transcript searches.
4. **It writes back with findings.** A headline answer, the method it used, evidence with numbers, quotes from the underlying conversations, and a recommended next step.

### Example response

Here is an excerpt of an answer to that question, from a sample workspace of customer support chats. The figures are the example's, not a benchmark (example from the Dimension Labs docs, May 2026).

```text
Plan
  ✓ Inspect available tables and Dimensions for sentiment
  ✓ Quantify how common dissatisfaction is and break it out by theme
  ✓ Pull representative quotes for each top theme
  ✓ Pull together root causes and recommendations
```

**Across 4,388 conversations (January 1 to February 1, 2026), the biggest drivers of dissatisfaction are billing, promo and credit problems, followed by account access friction.** Using a transcript-based proxy, 30.9% of conversations (1,356) show clear dissatisfaction language.

**Important caveat:** your usual sentiment Dimensions aren't populated, so this is based on patterns in the transcript text, not your configured Dimensions.

| Theme | Mentions | Dissatisfied | Dissatisfaction rate |
|---|---|---|---|
| Billing, promo and credits | 1,437 | 291 | 20.3% |
| Account access, login and PIN unlock | 1,605 | 289 | 18.0% |
| Setup and activation | 520 | 76 | 14.6% |

Representative quotes:

- *"BOGO promo isn't applying... being charged the full EIP..."*
- *"I just want to pay my bill, not troubleshoot an app."*
- *"I canceled... but I'm still being charged $55."*

A full response goes on to weekly trend charts, more themes and recommended actions.


## 3. Read the answer

Answers follow roughly the same shape. Once you've read one, the rest are quicker to scan.

| Part | What to look for |
|---|---|
| **Headline first** | One or two sentences with the main finding and the top-line numbers. |
| **Method disclosed** | If the agent used a proxy or worked around missing data, it says so, with wording like *"based on transcript patterns, not configured Dimensions."* |
| **Evidence with numbers** | Counts, percentages and time windows. Every claim comes with the number behind it. |
| **Quotes** | Direct quotes from the underlying conversations. They are the rawest signal, so read them. |
| **Recommended next step** | Usually either a deeper question to ask, or a setup step that would make the next answer more reliable. |

If something looks off, ask *"Show me the SQL you ran"* or *"How did you calculate that?"*. You can ask for the conversations behind any number.

## 4. When the agent doesn't have enough data

The agent would rather tell you it can't answer than guess. When it can't, it says exactly what's missing and what to do about it.

### Example refusal

> I can't yet answer "what's driving dissatisfaction" from your labeled Dimensions, because the tables that normally hold dissatisfaction and driver Dimensions are empty in your workspace.
>
> **What I checked**
>
> - `system_session_prompt` (expected sentiment, rating and reason Dimensions): 0 conversations
> - `support_reasons` (your custom prompt; expected issue category and predicted CSAT): 0 conversations
> - `transcripts` (raw transcript text): 4,388 conversations
>
> **What I can do next**
>
> Run a lighter theme extraction directly from the transcript text, which is less reliable but available now, or wait until a backfill has run the Dimensions on the full dataset.

Table names depend on your workspace: `support_reasons` here stands for a table named after one of your own prompts. See [What the agent can see](/agent/what-it-can-see).

From here you have two options:

1. **Tell the agent to use the available data.** It takes the interim path. That's how the example answer above was produced, and why it carries a caveat.
2. **Fill the gap, then ask again.** Run a [backfill](/dimensions/backfill) from the **Backfill** tab. An uploaded file and any new prompt need one, and the standard Dimensions (**System Prompts**) run every day only when Dimension Labs has turned on daily enrichment for your workspace. A backfill typically shows up within the hour, depending on how much data is being enriched. See [When enrichment runs](/dimensions/when-enrichment-runs).

:::tip[Trust the refusals]
Many analytics tools fail quietly and give you a number anyway. The agent fails out loud: it tells you what's missing, then offers a way forward.
:::

## 5. Optional: explore before you analyze

If you'd rather see what the agent can work with before asking an analytical question, start with discovery questions:

- *What's in my workspace?*
- *List my prompts.*
- *List my Dimensions.*
- *What metadata do I have?*
- *What Dimensions do I have for sentiment?*
- *Do we track competitor mentions anywhere?*

The agent lists what's available and tells you which kinds of questions it can answer well. More discovery questions are in [What the agent can see](/agent/what-it-can-see).

## It's a conversation

You can refine, follow up, change direction or start over. Treat the agent like a colleague: give it context, push back when an answer feels off, and ask why as often as you need to. See [Ask good questions](/agent/asking-good-questions).

## Next

- [Quick answers and analysis plans](/agent/modes): when to explore and when to build a repeatable deliverable.
- [Ask good questions](/agent/asking-good-questions)
- [Reports, decks and sharing](/outputs)