---
title: "Best practices"
description: "The habits that make answers good once your data is in, and the features that help."
url: https://www.gatlas.ai/docs/best-practices
markdown_url: https://www.gatlas.ai/docs/best-practices.md
---

# Best practices

Once the data is in, these habits decide how good your agent's answers are.

## Where each fact goes

An agent knows only what you write down in Gatlas. Each thing you know about your business has one place it belongs, and the agent looks for it there.

| What you know | Where it goes | Where to find it |
| --- | --- | --- |
| Facts about a company: what it does, its fiscal year, its currency | The company's AI context | Settings, Company, or each company's page under Settings, Companies |
| What a dataset's numbers mean: what one row is, signs, what is left out | The dataset's AI context and column notes | Datasets, under Definitions in the sidebar |
| Why the numbers moved: board notes, memos, plans | A brief | Briefs, under Definitions in the sidebar |
| Habits that hold across the workspace, such as how you round figures | A note | Ask a Gatlas agent in Slack to remember it. Admins see its notes on the agent's Notes tab, on Gatlas agents, under Agents in the sidebar |
| The files themselves, such as PDFs and board packs | Files | Files, under Library in the sidebar, synced from a drive |

Start small. A line or two in each place is enough to begin; add to it where an agent gets something wrong.

Agents read a company's **AI context** when they start working with your data. Under Settings, **Company**, write it in **AI context**; with several companies, open each one on **Companies** and choose **Edit** on its **AI context** card. Write the rules and exceptions behind your numbers:

- What the company does and how it makes money.
- Its fiscal year, reporting currency and accounting system.
- Anything that trips people up: a second legal entity, a one-off write-down, a customer who pays annually in advance.

Admins and above can edit it. See the [onboarding guide](https://www.gatlas.ai/docs/onboarding.md) for where it sits.

## Dataset descriptions

On a dataset, **AI context** is the single field that changes answers most: every agent question about the dataset starts with that text. Write the rules and exceptions behind its numbers: what one row is, what each figure means, which currency it is in, and what the dataset leaves out, such as "Revenue excludes VAT and refunds; one row per company per month; amounts in the company's own currency". See [Datasets](https://www.gatlas.ai/docs/datasets.md).

## Dataset approval

Agents are offered approved datasets only. The switch is **Approved**, when you create a dataset and when you edit one. Approval decides what an agent works from. It can still read the synced tables of its key's companies through SQL, so the companies you grant a key set what it can read. A few rules for approving:

- Approve a dataset once you have checked one of its numbers against the books.
- Keep one dataset per question. Two datasets that both claim to be "revenue" give two answers.
- Name datasets plainly, with the company first: "Northwind Monthly P&L", "Northwind Invoices". Agents find datasets by name.
- Switch off approval on datasets you have replaced, or delete them.

## Questions to ask

Four habits make most answers better:

- **Name the company.** "Northwind", not "the retail one".
- **Name the period.** "the second quarter of last year" or "the last 12 months". Open-ended dates get vague answers.
- **Name the dataset when you know it.** "Use the Monthly P&L dataset" keeps the agent on your approved numbers.
- **Define your terms.** "Overdue", "active customer" and "margin" mean different things in different places. Say which you mean, or put the definition in the dataset's AI context once.

Prompts that work well to start from:

> Which datasets can you see, and what does each one answer?

> What was revenue by company for each of the last six months, and which company grew fastest?

> List Northwind's invoices more than 90 days past due, largest first, with the customer and the amount.

> Compare Northwind's gross margin this quarter with the same quarter last year, and explain the biggest change.

> Is our data current? Flag any company whose tools have not synced recently or whose books have gone quiet.

If an answer looks wrong, ask the agent which dataset and filters it used. That usually shows the problem in one step. How to connect each agent is in [Connect your agent](https://www.gatlas.ai/docs/connect-your-agent.md).

## Limits of the data

- **Synced books can still be behind.** A green sync means Gatlas reached the tool. For a company with a ledger dataset, if nothing has been posted in its accounting for more than 30 days, Gatlas warns the agent, and you should treat "this month" answers for that company with care.
- **Totals across companies count trades between them.** A trade between two of your companies appears in both, and Gatlas does not eliminate it for you.
- **Currencies need a rule.** Each company reports in its own currency unless a dataset converts it. Before comparing companies, check which rate the dataset uses.
- **Long lists come in pages.** An agent reads up to 500 rows of a dataset at a time and is told when there are more. For totals, ask it to calculate the total rather than add up a list.

## Recurring reports

If you ask for the same charts every month, ask your agent to build them once as an artifact: a dashboard, an A4 document or a deck that reads your datasets each time it opens. Share it with a link or print it. See [Artifacts](https://www.gatlas.ai/docs/artifacts.md).

## Briefs

Board notes, memos and KPI definitions hold the reasons behind the numbers. Sync a Google Drive folder per company and turn it into short briefs, so an agent can answer "why did margin fall?" with the reason as well as the figure. See [Files](https://www.gatlas.ai/docs/files.md) and [Briefs](https://www.gatlas.ai/docs/briefs.md).

## Alerts and summaries

On **Automations**, under Agents in the sidebar, choose **New automation**. An alert reaches Slack or email when a sync fails or a connection stops syncing. A summary sends the weekly digest to a Slack channel, and a scheduled post puts an artifact there on the days you choose. See [Automations](https://www.gatlas.ai/docs/automations.md).

## Access reviews

Access drifts as people join and leave. Once a quarter:

- Under Settings, **Members**: check each person's role and companies, and remove anyone who has left. See [Team and access](https://www.gatlas.ai/docs/team.md).
- On **Agent access**, under Agents in the sidebar: sort by **Last used**, and revoke keys nobody uses with **Revoke key**. Check that anyone who only asks questions has a **Read-only** key. See [Agent access](https://www.gatlas.ai/docs/agent-access.md).
- On **Setup links**, listed on the **Sources** stage page: revoke links that are still pending and no longer needed. See [Connecting tools](https://www.gatlas.ai/docs/sources.md).
- On each artifact you have shared: make sure the share link, and the link of each live edition under **Editions**, still belongs where it was sent. Withdraw an edition that does not.

**Audit log**, under **Settings**, shows who changed what, for the companies you can see.

## Feedback

Use **Feedback** in the app's sidebar, or write to [team@gatlas.ai](mailto:team@gatlas.ai); we usually reply within a business day. You can also ask your agent to send us feedback from inside a conversation, which reaches the same people.

> **Tip:** A missing tool, a number that does not match your books, or a question the datasets cannot answer yet are the most useful things to send.
