# Agentic CRM: what AI agents should do in your CRM, and what they shouldn’t

Migaru, Oct 31, 2026 · Perspectives

> An agentic CRM lets AI agents do work, not just answer questions. Which jobs to hand them, which to keep, and the guardrails that make the difference.

An **agentic CRM** is a CRM where AI agents do work on the customer record rather than only answering questions about it. They read conversations and update records, research accounts, draft follow-ups and outreach, and carry out tasks of several steps, with people approving what matters.

The interesting question isn’t whether a CRM has agents. Most do, or soon will. It is which jobs they are given, how much they decide alone, and what they are working from.

## From chatbot to agent

AI has arrived in CRMs in roughly three forms:

- **Assistants** answer questions and write text when asked: summarise this account, draft an email to this person.
- **Automations with AI in them** run fixed steps on a trigger: when a call ends, update these fields; when a deal reaches a stage, write a note.
- **Agents** are given a goal and decide the steps: find companies like our best customers and draft a first email to each; prepare me for tomorrow’s meetings; tell me which deals slipped this quarter and why.

An agentic CRM is built for the third kind, so that agents can work safely on the same record people do. [CRM automation](https://www.migaru.ai/blog/crm-automation) covers when the second kind is the better choice.

## An agent is only as good as the record

An agent has no memory of the deal and no instinct that a field looks wrong. It works from what the CRM holds. If the last-contact date is from August and the summary was written four calls ago, it will prep a meeting from stale context and chase someone who replied yesterday.

That is why the agentic CRM and the AI-native CRM are one idea. Before agents can do useful work, the record has to keep itself current from the email, meetings and calls where the work actually happens. [What is an AI-native CRM?](https://www.migaru.ai/blog/what-is-an-ai-native-crm) covers that half.

## A ladder of autonomy

It helps to sort what an agent might do by how hard it is to undo:

- **Read and answer.** Search records, calls and threads, and answer with sources. Nothing changes, so nothing needs approving.
- **Write facts.** Who was on a call, when someone last replied, a summary of the thread. These are observations, cheap to correct, and safe to write automatically.
- **Suggest judgements.** A deal’s stage, value and close date, or a company’s industry. These are somebody’s call, so the agent proposes them with a reason and a person accepts or dismisses them.
- **Draft for a customer.** Follow-ups, sequence steps and anything else the customer will read. A draft costs nothing to throw away, so it waits for a person to edit and send.
- **Act outside the CRM.** Sending email, posting messages, changing another system. The hardest to undo, and the rung to give an agent last, with limits.

Most of the argument about agentic CRMs is about where to draw the line on this ladder. Some products let agents climb all of it. Our view is that the record should keep itself, judgements should be suggested, and a person should press send, except for campaigns a person has written, launched and put limits on.

## Guardrails to look for

Whatever line a CRM draws, these are what make agents safe to use on real customers:

- **Citations.** Every answer and every suggested change says which call, meeting or thread it came from, and a claim with no source is dropped.
- **One review queue.** Suggested changes, drafts and proposed records wait in one place, each with its reason, to approve or dismiss in batches.
- **Field-by-field control.** An admin decides, for each field, whether the agent writes it, suggests it or leaves it alone.
- **A history on every field.** What changed, when, and what changed it: a person, the agent, a sync, an import or the API.
- **Permissions per action.** Proposing changes, building lists, running code, searching the web and drafting email can each be switched on or off.
- **Hard stops on sending.** Sending stays inside set hours and daily limits, and a reply, bounce or unsubscribe ends a campaign for that person.
- **A place to try things.** A sandbox copy of the workspace, where workflows and agents run without reaching a customer.

In Migaru, the first five are on every plan, the stops hold wherever it sends, and [sandboxes](https://www.migaru.ai/help/try-changes-in-a-sandbox) come with Enterprise.

## Agents from outside the CRM

Not every agent that touches your CRM will live inside it. Assistants like Claude and ChatGPT can connect to a CRM over MCP, the Model Context Protocol, and read or change records from a conversation there. The CRM’s own agent can go the other way, signing in to the MCP servers of the other tools a team uses, such as Notion, Linear or Stripe, to answer with what they hold.

The same rules apply at that boundary. An admin should decide whether a connected assistant can only read, or create and change records too, and every change it makes should show up in the record’s history. In Migaru, an assistant connects by adding https://app.migaru.ai/api/mcp as a custom connector and signing in, an owner or admin chooses what it may do, and its changes are marked as made through the API. [Ask the agent](https://www.migaru.ai/help/ask-the-agent) walks through it, and [using your CRM from Claude or ChatGPT](https://www.migaru.ai/blog/mcp-crm) covers what to decide before you connect one.

## How agents are priced, and why it matters

Pricing shapes how much an agent gets used. CRMs with agents charge in four ways:

- **Per seat**, with the agent included: the bill grows with the team, not with how much the agent does.
- **Per workspace**, one price for the team: the same, with fewer decisions about who gets a seat.
- **Credits**: AI work draws on an allowance, and work beyond it is billed or stops. [Lightfield](https://www.migaru.ai/compare/migaru-vs-lightfield), [Attio](https://www.migaru.ai/compare/migaru-vs-attio) and [Clarify](https://www.migaru.ai/compare/migaru-vs-clarify) each use credits, in different forms.
- **Per agent**: you pay for each agent you deploy, as [Day.ai](https://www.migaru.ai/compare/migaru-vs-day-ai) does.

Credits and per-agent pricing put a meter on every question. That can be fair for heavy use, but it teaches a team to ration the agent, and makes next year’s bill hard to predict. Migaru charges a fixed price, per user on Starter and one price for the workspace on Pro and Growth, so asking the agent costs the same however often you do. [See pricing](https://www.migaru.ai/pricing).

## Where Migaru draws the lines

- **Written by itself:** new contacts and accounts from your conversations, summaries of accounts, contacts, deals, meetings and threads, last-contact dates, and prep before external meetings.
- **Opened by itself:** a deal, when a company is plainly buying, with its owner told why.
- **Suggested:** a deal’s stage, value and close date, and a company’s details, unless an admin changes that field by field.
- **Drafted:** follow-ups and sequence steps, into a review queue. An approved email goes to your own drafts for you to send.
- **Sent by itself:** only Outreach campaigns a person writes and launches, on Growth, from mailboxes connected for them and within their limits.
- **Never:** a message on LinkedIn, or a post in a Slack channel.

[The AI sales assistant](https://www.migaru.ai/product/ai-sales-assistant) shows the agent at work, and [how records keep themselves up to date](https://www.migaru.ai/help/how-records-keep-themselves-up-to-date) lists every field it writes or suggests.

## Frequently asked questions

### What is an agentic CRM?

A CRM where AI agents do work on the customer record, not only answer questions about it: reading conversations and updating records, researching accounts, drafting follow-ups and outreach, and carrying out tasks of several steps, with people approving what matters.

### What is the difference between an AI agent and a CRM automation?

An automation runs fixed steps on a trigger. An agent is given a goal and decides the steps, such as researching an account, choosing who to contact and drafting what to say. Because its path isn’t fixed in advance, an agent needs more guardrails: sources for what it claims, a review queue, and limits on what it can change or send.

### Should AI agents send emails for my sales team?

It depends on how much you trust the drafts and how costly a mistake would be. Many teams start with agents drafting into a review queue and a person sending, then let campaigns they have written and launched send within set hours and daily limits, stopping for anyone who replies, bounces or unsubscribes.

### Can Claude or ChatGPT work with my CRM?

Yes, if the CRM has an MCP server. In Migaru, add https://app.migaru.ai/api/mcp as a custom connector in Claude, ChatGPT or another assistant that takes MCP connectors, and sign in. An owner or admin decides whether it may create, change and archive records or only read them.

### What do AI agents in a CRM need to work well?

A record that keeps itself current from email, calendar and calls, a source for every claim, and clear limits: which fields they may write, which they may only suggest, and whether they may send anything a customer reads.

---

Canonical page: https://www.migaru.ai/blog/agentic-crm
Last updated: 2026-10-31