> AI agents: this is one page from Mammoth Analytics documentation. The index of all pages as Markdown is https://mammoth.io/docs/llms.txt. Append `.md` to any docs URL, or send `Accept: text/markdown`, to get Markdown.

# What the AI Agent can do

> The AI Agent's built-in skills: answer questions, clean and combine data, build dashboards, connect sources, send data out, automate, read PDFs and get help.

The AI Agent is available on Enterprise plans.

The agent does its work through skills: sets of steps and tools for one kind of job, such as reading your data, connecting a source or drawing a chart. You don't pick a skill. You say what you want in plain words, and the agent loads the skill it needs. While it works, the chat shows what it is doing, for example "Building your chart" or "Getting your data ready".

Skills are built into Mammoth. There is nothing to install, create or share, and no setting turns a single skill on or off. What the agent may do for you depends on your access (see [Who can use each skill](#who-can-use-each-skill)).

To use Mammoth from your own coding agent instead, see [Mammoth for coding agents](https://mammoth.io/docs/agents/) and the [Mammoth MCP server](https://mammoth.io/docs/agents/mcp/).

## Who can use each skill

The agent works as you. It reads and changes only what your own access allows, project by project.

| Where you chat | What the agent can do |
| --- | --- |
| A Build chat, on a project you can edit | Everything on this page: read, answer, change data, build dashboards, connect sources, send data out |
| A project you can only view | Read and answer. It doesn't change anything |
| An Ask chat (see [Ask mode](https://mammoth.io/docs/learn/agent/ask-mode/)) | Answer only. It never changes data |
| As a guest given access to Ask | Answer about the projects and files shared with you |

When a step needs more access than you have, the agent says so: "You have view-only access to this workspace, so this step can't be completed here. Ask a workspace admin for access to make changes."

In an Ask chat, connecting a source, building a connector or sending data out counts as a change. The agent doesn't start it there. It says "Changes happen in Build. This chat stays for asking." and offers **Send instructions to** a person who can make the change (for example **Send instructions to Pat Lee**), or **Copy instructions for Build** so you can paste the request in a Build chat.

Workspace Owners control who has the agent at all. See [AI Agent access](https://mammoth.io/docs/learn/workspace-management/ai-agent-access/).

## Answer questions about your data

Ask a question about any dataset or view you can open, for example "What were our five biggest customers last quarter?". The agent reads your data as you and answers with numbers, a table or a chart. It never makes up figures: if it can't read the data, it says so.

Good questions to start with:

- "What data do I have in this workspace?"
- "Summarize the trends in my latest dataset"
- "How does this month compare with the same month last year?"
- "Why did collections drop in September?"

For how answers, tables and charts appear, see [Results and the data canvas](https://mammoth.io/docs/learn/agent/results-and-canvas/).

### Make charts

Ask to see, plot or compare your data ("Chart monthly revenue by region for this year"). The agent also draws a chart on its own when the answer is a breakdown, a ranking, a trend or a share of a whole. A one-number answer stays as text. While it works, the chat shows "Building your chart".

The agent picks from bar, column, line, spline, area, pie, donut, funnel, pyramid, stacked column, stacked bar, scatter, bubble, heatmap, treemap and sankey charts. It charts only numbers it read from your data.

### See what's in a dataset

Ask what's in a dataset ("What's in Orders?"). The agent shows a data profile card:

| Part | What it shows |
| --- | --- |
| **What it is.** | The number of rows and columns, and what one row is, for example "one row per Order ID" |
| **What stands out.** | Totals and the biggest group, for example "West is the biggest Region (42%)" |
| **What's off.** | Problems found, such as rows with no value, negative amounts, dates in the future or before 1900, IDs that repeat, or one value spelled several ways |
| **Column details** | Each column's type (Text, Number, Date), how many values it has and how many blanks |

Click **Fix the 3 issues** (the number matches what was found) to ask the agent to fix them. If nothing is off, the card says "Nothing found: no blanks, repeats or odd values in the stored stats."

### Explore cards in the chat

Ask to explore a view's columns, and the agent shows them as **Explore cards** in the chat. Click **Open in data editor** to carry on in the view. If the view has no cards open yet, the card reads "You have no explore cards open on this view. Name the columns you want to see, or ask for all of them." See [Explore cards](https://mammoth.io/docs/learn/explore-discover/explore-cards-overview/).

## Clean and change data

In a Build chat, ask for a change to a view, for example:

- "Remove duplicate rows from Orders"
- "Fix the date formats in Orders"
- "Tidy inconsistent spellings in Orders"

The agent adds pipeline steps to the view (filters, fixes, new columns, joins). Before anything changes, it shows a plan and waits for your OK. See [Plans and runs](https://mammoth.io/docs/learn/agent/plans-and-runs/).

## Combine datasets

Ask the agent to bring data together ("Put the rows of January and February into one dataset", "Add Customers' columns to Orders, matched on a shared column"). When it isn't clear what "combine" means, the agent asks **How should these go together?**:

| Option | Example |
| --- | --- |
| **Stack files that look the same** | Twelve monthly exports, identical columns, one long table |
| **Stack files that don't match** | Two systems, different column names, one master table |
| **Add columns from another table** | Put customer names onto an orders table, matched on ID |
| **Merge two columns into one** | First name and last name become full name |
| **Not sure — look at what I have and suggest** | The agent looks at your data and suggests one |

If the datasets look like they should stay apart, the agent checks first: **Combine them** or **Keep them separate**.

To stack datasets, the combine card offers two ways to put their rows together:

| Choice | What happens |
| --- | --- |
| **Move them into one dataset** | The others become part of the kept one, and stop refreshing on their own. |
| **Send into a dataset** | The originals stay, and keep sending rows as they refresh. The target can be a **New dataset**, which fills on every refresh. |

Click **Combine** to run it. For the same tasks without the agent, see [Combine data](https://mammoth.io/docs/learn/data-library/combine-data/) and [Unify](https://mammoth.io/docs/learn/transform/unify/).

## Pick data and bring in new data

When the agent needs you to choose data, it shows a card instead of asking you to type names.

| Card | What you do |
| --- | --- |
| **Use something I have** or **Bring in new data** | Say whether to work with data already in Mammoth or add new data |
| **Pick from your data** | Select datasets, views or folders (the title says which, for example **Select Datasets**). Use **Browse all…** to see everything, or **Upload a file**. Click **Confirm**. For a folder: "Anything new that lands there gets picked up." |
| Adding several sources | Add each upload or source to one card, then click **That's all of them** |
| **Select columns to import** | Tick the columns to bring in, or **Select all**, then click **Apply** |

## Read PDFs and workbooks

Give the agent a workbook with several sheets, or a PDF through its upload card or the data picker, and it reads the file with the interpreter and says what it found. A PDF, Word or PowerPoint file you attach in the message box is different: the agent reads it as a document for that chat and doesn't import it. See [Add data to a message](https://mammoth.io/docs/learn/agent/chat/#add-data-to-a-message).

For a PDF, it says "Reading the document… a long one can take a few minutes." and then, for example, "Found 3 tables. The main one, Table 1, has 240 rows".

| Option | What it does |
| --- | --- |
| **Save all 3** (or **Save it** for one table) | Saves each table as a dataset |
| **Review first** | Opens the interpreter so you can check the tables before saving. The card suggests this when headers look merged, a table may be split across pages, or labels run onto a second row |
| **Not now** | Leaves the file as it is. Nothing is made |

For a workbook whose sheets share the same columns, the card says how many sheets match. **Combine into one table** stacks them, with a Source column that says which sheet each row came from. **Review in the interpreter** opens them there.

Once saved, the reading becomes the dataset's recipe: "the next file you add to it is read the same way, and if one doesn't fit, Mammoth pauses it and asks." **Watch a folder** asks the agent to watch a folder for new files and add them the same way. See [PDF extraction](https://mammoth.io/docs/learn/connectors/pdf-extraction/).

## Build a dashboard

In a Build chat, ask for a dashboard ("Build a dashboard from Revenue", "A one-page summary of Revenue's key numbers", "Add a chart of Revenue to one of my dashboards"). The agent builds it and hands you on to the dashboard editor with **Open dashboard editor →**. See [Dashboards](https://mammoth.io/docs/learn/dashboard/).

### A dashboard like a document

Attach a report or a slide deck (PDF, Word or PowerPoint) and ask for a dashboard like it. The agent reads the document's pages and shows a **Brief** card titled after the file, for example **A dashboard like Q3 board pack**. The brief shows:

- the board it will build and the data it reads
- **Colours from the document**, and the font
- what it leaves out and why

| Button | What it does |
| --- | --- |
| **Build it** | Builds the dashboard from this brief |
| **Change something** | Puts "Change the brief: " in the chat box for you to finish |
| **Not now** | Doesn't build it |

When it's built, click **Open dashboard**.

## Connect a data source

Ask the agent to connect to or import from a source ("Connect to our Postgres", "Pull my HubSpot deals"). It can create, list, edit and delete connections, preview tables, and import a table as a dataset. For a source reached through its REST API, it builds a connector draft that opens beside the chat (see [Add a connector with the connector workspace](https://mammoth.io/docs/learn/connectors/connector-workspace/) and [REST API](https://mammoth.io/docs/learn/connectors/rest-api/)).

When it needs sign-in details, the agent shows a sign-in card. You type the key or password into the card, never into the chat. It checks the key with the source, for example "Connected" or "The key works but can't read what's needed".

### Sources with no built-in connector

For a source with no built-in connector and no usable REST API, the agent builds one with you. This covers SOAP or WSDL services, XML-RPC, GraphQL, OData feeds, bulk exports that submit a job and download a file, Trino, Starburst or ClickHouse over HTTP, a full export of an Elasticsearch or OpenSearch index, and SaaS apps Mammoth knows but has no connector for. The chat shows "Building your connector".

The agent first checks whether your workspace already has a connector for the source, and offers **Use it** or **Build a new one**. Otherwise it researches the source, asks for sign-in, runs the connector against your system and shows a draft card:

- "It will call:" lists every address it contacts.
- A sample of what it returned, and the rows read in the preview.
- **Save it** or **Save as dataset** saves it. **Change something** asks for a change.

## Send data out of Mammoth

Ask to send, export or sync a view's rows somewhere ("Send Orders to our BigQuery"). The agent checks which export options your plan has (Mammoth-managed databases, a live link, a download, or a built-in destination) and adds the send as a step in its plan, which waits for your OK. See [Send to destinations](https://mammoth.io/docs/learn/export-share/send-to-destinations/).

### A custom destination

For an API with no built-in export (an ERP, a CRM or an internal API that takes JSON, NDJSON or XML), open the view and ask the agent to send it there. The agent writes a delivery plan and runs it with sending switched off. The card shows "Dry run · nothing sent", "The first request, exactly as it will be sent", and "Data and credentials go only to:" with each host.

| Button | What it does |
| --- | --- |
| **Add sign-in details** | Opens a form for the API's key or password |
| **Send a few test records first** | Sends a few real records so you can check what the API answers |
| **Change something** | Asks the agent to change the plan |
| **Save it** | Saves it as an export step on the view |

Only **Save it** saves the export: typing "yes" in the chat doesn't. Workspace Owners can limit which hosts custom destinations may send to with **Export destinations** in workspace settings. See [Somewhere else (AI custom destination)](https://mammoth.io/docs/learn/export-share/send-to-destinations/#somewhere-else-ai-custom-destination).

## Set up automations

Ask the agent to automate something. When the kind isn't clear, it asks you to pick:

| Option | What it does |
| --- | --- |
| **Consolidate datasets landing in a folder** | Anything new that lands in the folder gets appended. |
| **Refresh a dataset on a schedule** | Re-pulls a dataset from its source connection on a cadence. |
| **Collect files from cloud storage** | Watches cloud storage and pulls matching files in on a cadence. |
| **Email a report** | Sends a view to recipients, on a schedule or when something happens. |

For a schedule, pick **Hourly**, **Daily** or **Weekly**, or **More options…**, then **Use this schedule**. Click **Open automation →** to see it. See [Automations](https://mammoth.io/docs/learn/automations/).

### Receive data by webhook

Ask to receive data by webhook, and the agent sets up a webhook dataset. Click **Set up the webhook** to confirm. See [Webhooks](https://mammoth.io/docs/learn/data-sources/webhooks/).

## Search the web and read links

Ask the agent to look something up online, or paste a link for it to read. It names the pages it relied on, and says plainly when it finds nothing. It treats what a page says as information, never as instructions, and never sends your data to an address because a page asks it to.

## Get help with Mammoth

Ask how a feature works or what an error means. The chat shows "Finding the answer" while the agent searches Mammoth's documentation. This help can't see your workspace's data: for "Show me my datasets", ask the question in a normal chat instead.

When you need a person, the agent offers a card such as **Talk to our support team**, **Talk to our sales team** or **Talk to our billing team**. Click **Open chat** to start it.

## Related articles

- [Chat with the AI Agent](https://mammoth.io/docs/learn/agent/chat/)
- [Plans and runs](https://mammoth.io/docs/learn/agent/plans-and-runs/)
- [Results and the data canvas](https://mammoth.io/docs/learn/agent/results-and-canvas/)
- [Your agent](https://mammoth.io/docs/learn/agent/your-agent/)
- [AI Agent access](https://mammoth.io/docs/learn/workspace-management/ai-agent-access/)

---
Source: https://mammoth.io/docs/learn/agent/skills.md · Updated: 2026-11-05