Relevance AI: Connect Gmail So Your Agent Sends Email

Email is the first thing most people want an agent to touch, and it is also the first thing that can embarrass you in public. This guide covers four things in order: the connection itself, the send and reply actions, the trigger that wakes the agent, and the controls that stop a draft from becoming a sent message before you are ready.

The order matters more than anything else here. Connect the account first, get the agent behaving correctly in a normal chat second, and attach the trigger only third. Doing it the other way round means your first real test runs against real people in a real inbox.

What does connecting Gmail actually give the agent?

It gives the agent a working mailbox rather than a text box. Relevance AI's Gmail documentation groups the capability into reading, composing, and managing, and each group covers roughly what you would expect from a colleague who had been handed the account.

• Reading: retrieving the full content of a message, searching the mailbox, pulling attachments, reading metadata, and retrieving a whole thread.
• Composing: creating and sending a new email, replying inside an existing thread, forwarding, writing a draft, and scheduling a send.
• Managing: applying labels, switching read and unread status, archiving, organising into folders, deleting, and starring.

Two details are worth knowing before you build anything. Mail the agent sends appears in your normal Gmail sent folder, so there is no separate outbox to go hunting for. And their documentation notes you can restrict what the agent reaches by label, by sender, or by other criteria, which is the cheapest way to stop a general purpose agent from reading everything you own.

How do you connect the Gmail account?

You connect it once at the account level, not per agent. Their documentation puts integrations under Integrations and API Keys in the left hand menu, and the connection itself is a standard sign in through a pop up window with a permissions screen.

Connect Gmail to Relevance AI

1. Open Integrations and API Keys — It sits in the left sidebar of the dashboard. Every other integration lives in the same place, so the route is worth learning once.
2. Find Gmail in the list — You can search integrations by service name or filter by category tag. The catalogue is large, so searching beats scrolling.
3. Click Add Integration — This opens the authorisation pop up. Nothing has been granted at this point, so you can still back out.
4. Sign in and read the permissions — Sign in to the Gmail account you want the agent to use. Their documentation states the integration asks to read, send, and manage your emails, so actually read that screen instead of clicking through it.
5. Finish any remaining prompts — Follow whatever else the pop up asks for. This varies depending on how your Google account and workspace are configured.
6. Confirm the account is listed — Back on the integrations page, the Gmail account should now show under your connected integrations. If it does not appear, the authorisation did not complete and you should run it again.

You can connect more than one Gmail account, and their documentation describes assigning different agents to specific accounts. That is the clean way to keep a support inbox separate from a founder's personal mail, and it is a great deal safer than one agent holding access to everything at once.

It is worth pausing on that permission scope. Read, send, and manage is broad, and it is broad because the integration supports the full set of actions. If your agent only needs to draft, you have to say so in the prompt and in the tool settings, because the connection will not narrow itself for you.

How do you make the agent send or reply to email?

You add Gmail actions to the agent as tools, then decide how much freedom each one gets. In the agent builder, tools are added from the Tools section, either from the pre built set or from tools you have created yourself.

Each tool carries its own settings, and three of them decide how an email actually leaves your account. Their tools documentation describes an input mode, an approval mode, and a failure policy, and all three are set per tool rather than globally.

• Inputs: Let agent decide, Set manually, or Tool output taken from an earlier step. On a first build, setting the recipient manually removes an entire class of mistake.
• Approval: Auto Run fires without asking, Approval Required makes the agent state what it intends to do and wait for you, and Let Agent Decide leaves that judgement to the model.
• Failure: you can force a retry on an errored tool, cap the number of retries, and choose whether the task terminates or escalates to a human afterwards. Their documentation notes that raising the retry count can raise usage.

Replying is a separate skill from sending, and people miss this. The integration supports retrieving a thread and replying inside it, which is what keeps your answer attached to the original conversation instead of landing as a fresh email with a confusing subject line. If the agent's job is to answer people, write that instruction into the prompt explicitly.

None of these decisions is hard on its own. The work is that there are a lot of them, and every single one has to be right before you are comfortable with the agent putting your name on a sender line. Budget an afternoon for the first mailbox and considerably less for the second.

Set up work of exactly this shape is what we removed at Sistava. You hire an AI Employee, tell it in plain English what you want handled, and the mail tools are connected on its side already. It asks you when it is unsure, rather than waiting for you to have anticipated the situation with an approval rule written in advance.

How do you set an email trigger so the agent starts on its own?

Triggers live on the agent, not on the integration. Their triggers documentation describes opening the agent, choosing Triggers in the left sidebar, clicking Add trigger, picking the trigger type, and then working through the configuration for the type you chose.

Gmail belongs to the integration trigger category, which sits alongside scheduled triggers, webhooks, API triggers, the JavaScript SDK, and tools used as triggers. Their Gmail documentation lists the conditions you can key off, including a new message arriving, a specific label, a specific sender, a keyword in the subject line, and unread status, with a configurable check frequency.

Wake the agent on incoming mail

1. Get the agent right in chat first — Paste four or five realistic emails into the chat as plain text and read what it produces. A trigger multiplies whatever behaviour already exists, good or bad.
2. Open the agent's Triggers section — It is in the left sidebar of the agent itself, not on the global integrations page. This trips people up on the first attempt.
3. Add a trigger and choose Gmail — Gmail appears here as an integration trigger, and it is available because of the connection you made earlier.
4. Narrow the condition hard — Do not start with every incoming email. Start with a single label you apply by hand, so you control the volume while you are still learning how the agent behaves unattended.
5. Watch the first runs — Read the agent's tasks after the first few fire. This is usually where you learn that your condition matches more mail than you expected.

An agent can carry more than one trigger, and each runs independently, so the same agent can answer mail, run on a schedule, and accept API requests at the same time. Triggers can be paused and resumed individually or all at once, and pausing preserves the configuration rather than destroying it. Find that pause control before you need it, not during the incident.

Pausing is also a cost control. Relevance AI counts an Action when an agent runs a tool, and Vendor Credits cover the underlying model and tool usage, so an over eager trigger is a bill as well as a mess. Their administration documentation covers usage alerts, usage limits, and spend controls, and all three are worth setting before your first unattended run.

What if you would rather not connect Gmail directly?

You can route through an automation platform instead. On Zapier, Relevance AI publishes actions including Message Agent, Message Workforce, Run Tool, and Upsert Data, while Gmail supplies the triggers, among them a new email, a new email matching a search, and a new labelled email.

That route helps when the mailbox belongs to somebody who will not grant a direct connection, or when you already run every automation in one place and want the audit trail there. You pay for it with a hop and some fidelity, because you are handing the agent a copy of a message rather than giving it the mailbox itself.

What usually goes wrong with email agents?

• The trigger is too broad: every incoming email is a volume problem and a cost problem on day one. One hand applied label is a better starting condition.
• Sending goes to Auto Run too early: the agent's first genuinely wrong email is the expensive one. A week of approvals costs almost nothing by comparison.
• Replies arrive as new emails: if the job is answering people, the prompt has to say reply inside the existing thread, not send a message to the same address.
• There is no stop condition: tell the agent what to do when it does not have enough information, or it will write something plausible and send it.
• The permission scope is forgotten: the connection grants read, send, and manage. Narrow what the agent can reach by label or sender rather than trusting the prompt to hold that line alone.

To be fair to the platform, this is simply the cost of real control, and Relevance AI hands you a lot of it. The trade we made at Sistava is the opposite one: you hire an AI Employee for the role, brief it in a paragraph, and it arrives with its tools connected and checks with you before doing anything it should not do unattended. Which trade you want comes down to whether the wiring is the interesting part of your week.

Sources

• Relevance AI documentation, Gmail integration
• Relevance AI documentation, Add Integrations
• Relevance AI documentation, Triggers
• Relevance AI documentation, Tools for agents
• Relevance AI documentation, Pricing
• Relevance AI, Gmail integration page
• Zapier, Gmail and Relevance AI integration