Quick answers and analysis plans
Two ways to work with the agent: ask for a quick answer, or ask for an analysis plan for structured, repeatable deliverables.
There are two ways to work with the agent: ask a question for a quick answer, or ask for an analysis plan when the work needs structure. There’s no mode to switch. You choose by what you type, and most work uses both, sometimes in the same chat.
Ask when you’re exploring. Plan when you’re producing.
Planning is always there
Section titled “Planning is always there”Even for a quick question, the agent plans. When you ask a real question, it lays out a short task list, inspects your data, runs queries and puts the result together. That happens whether you ask for it or not. See Your first chat.
An analysis plan goes a level deeper. Instead of letting the agent work out its approach in the moment, you ask it to write the approach down first, review it together, and save it for next time.
Plan an analysis of why customers cancel in their first 90 days.
- Inventory the tables and Dimensions
- Profile cancellations by tenure and channel
- Read the last conversation before each cancellation
- Rank the drivers by effect
- Draft the report and the deck
share of early cancellations, %
FindingBilling surprises drive the most early cancellations, ahead of onboarding gaps. The plan saves as a file you can rerun every month.
Compare the two
Section titled “Compare the two”| Ask | Plan | |
|---|---|---|
| Use it when | You have a direct question, or want to explore without committing to where it goes. | You have a complex question, or you’re producing a report, brief, deck or dashboard, especially a recurring one. |
| Path | Question, answer, follow up, change direction. | Make the plan, review it, save it, run it. |
| How to start | Type your question. | Type your question and ask the agent for an analysis plan. |
| Output | Answers in the chat, with charts and the rows behind them. A deliverable at the end if you want one. | A plan file, plus the deliverable it produces: a longer analysis in sections, or a set report structure. |
| Cadence | One chat. | One chat, or a repeatable report: weekly, monthly, quarterly. |
| Reuse | Turn the chat into a plan afterwards. | Built for reuse. Same plan, new data, comparable results over time. |
| Best for | Exploring, quick answers for stakeholders, building an analysis over several turns. | Deep dives that need structure at the start, and specific deliverables. |
For any recurring deliverable (weekly digests, monthly reports, quarterly reviews), use a plan. When the shape has to stay the same over time, that matters more than any single answer.
Ask in plain English. The agent turns your question into queries and joins your structured business data with your enriched conversation data. Ask when you need a quick answer, or want to build an answer over several turns.
Example questions:
- What are the top drivers of churn among enterprise accounts?
- How does satisfaction differ between annual and monthly plans?
- What’s the trend in billing complaints by region over the last six months?
- Which product issues drive negative NPS among enterprise accounts?
- How do billing complaints differ by plan tier and region?
Every answer comes from row-level data. Ask for the conversations behind any chart or number, or how a figure was calculated.
When you have what you need, ask the agent to compile the chat into a deliverable:
- a Markdown summary
- an HTML report
- a CSV of the underlying rows
- chart files you can paste into a deck
See Reports, decks and sharing for every output type and where they’re saved.
Give the agent a business objective. It builds an analysis plan, runs it on your current data, and produces the finished deliverable.
Start a plan
Section titled “Start a plan”Ask the agent to create a plan, name the deliverable, and say what kind of output you want. For example:
Using the available data, create an analysis plan for a monthly customer experience digest. Identify the top recurring issues, quantify how each has shifted over the past quarter, and surface representative quotes for each. Reference the relevant tables and columns. Output as a Markdown file I can save.
Asking for a file is what makes the plan reusable.
What comes back
Section titled “What comes back”An analysis plan has four parts:
| Part | What it holds |
|---|---|
| Approach | The research design: how the question will be tackled, including any statistical methods. |
| Data sources | The specific tables and columns the agent intends to query, named explicitly. |
| Analytical steps | The analyses in order. This is what the agent runs when you run the plan. |
| Output specification | The shape of the deliverable: sections, formatting, and how quotes and metrics are presented. |
Review it before you run it
Section titled “Review it before you run it”Push back on anything that’s off: the wrong table, a missing segment, the wrong way to measure something tricky. Reviewing the plan is the best way to see how the agent reasons, and it’s the cheapest place to catch mistakes.
Save it and run it again
Section titled “Save it and run it again”When you’re happy with the plan, keep the file in your project so you can run it again: same plan, new month, comparable output. Finished deliverables are saved in the project’s Final Reports folder. To share one, publish it to Artifacts from the File Manager, which gives you a link anyone can open. See Share it.
Analysis plans has the full step-by-step procedure: checking fields, tightening scope, asking for the “so what”, and running the plan in a fresh chat.
What a report includes
Section titled “What a report includes”A report built from a plan usually has:
- an executive summary with the key findings
- the method and the statistical analysis
- charts and segment breakdowns
- recommendations grounded in the evidence
Reports are repeatable: define the objective once and run it every month. You can download them as HTML, PDF, PowerPoint (PPTX) or Markdown, among other formats.
Bring supporting files
Section titled “Bring supporting files”Bring files when you have them. The agent uses what you give it and asks for the rest.
| File | What it gives the agent |
|---|---|
| A plan file | A description of the report you want |
| Report execution guidelines | How the output should look: charts, formatting, brand |
| A reference report | A previous report whose calculations and figures the new one should compare against |
| A template | A shape the output should follow |
Name each file in your message, in the order to read it. The agent plans before it looks around, so a file it isn’t told about may not make it into the plan. Context engineering describes every file type.
Use them together
Section titled “Use them together”A useful pattern: explore a question by asking, find something worth tracking, then ask the agent to turn what you did into a plan you can run on a schedule.
The reverse works too. Plans aren’t rigid. You can leave a plan part way through, follow a lead the data turns up, and come back to the plan when you’re ready. Treat plans as living documents, not fixed recipes.
Designing Dimensions with the agent
Section titled “Designing Dimensions with the agent”When a concept isn’t captured by your Dimensions yet, the agent tells you and can help you add it. For example:
Create a Dimension that captures billing-related frustration.
The agent samples your data and drafts a definition with you. Saving it takes Prompt Designer:
- Turn on Prompt Designer in the agent’s settings. With it on, the agent can design a prompt with you and save it. With it off, the agent can draft a definition in the chat but can’t save it.
- Review the draft and test it on a sample before you run it on everything.
- Save the prompt. It appears in the Prompts tab, and from then on it runs on new conversations from your live integrations every day.
- To cover conversations you already have, and anything you upload as a file, start a backfill from the Backfill tab. The agent can’t start one for you. A backfill typically shows up within the hour, depending on how much data is being enriched.
The result is a new structured field you can use in every later analysis, not a one-off answer. See When enrichment runs for timing.
Going further: context engineering
Section titled “Going further: context engineering”More advanced plans layer in extra context files. Each one is a guardrail or an instruction set that shapes how the agent works:
AGENTS.md: standing instructions read in every chat of a project- a skill: how to do a specific, repeated task
- report execution guidelines: how outputs should look
- a schema file: what your columns and Dimensions mean
- a failure-modes file: output habits you don’t want
Curating the right inputs so the agent produces the right outputs is called context engineering. Setup is covered in Projects and standing instructions.