AI CRM: What an AI-Native CRM Does and How to Judge One

An AI CRM is a CRM in which a language model works with your customer data: it reads contacts, deals and conversations, answers questions about them, changes records and drafts the next message. Products differ less in whether they have AI than in what the AI may do. In some, it writes text in a side panel and a person still makes every change. In an AI-native CRM, the AI uses the same actions and permissions as a user, and it also works when nobody types a prompt: on a schedule, or when a record changes.

Customermates, an AI-native CRM, publishes this guide and appears in its own labelled section below. Facts about other vendors come from their own pages, checked on 8 October 2026.

What AI does in a CRM today

Most of what gets sold as AI in customer relationship management falls into seven jobs. The first five are language-model work and pay off in the first week, even with a few hundred records. The last two depend on data volume and on outside sources.

A scoring model learns from your won and lost deals. A consultancy that closes 30 deals a year gives it very little to learn from, so for small teams the first five jobs save the most time.

AI-native, add-on AI and smart CRM: what the labels mean

Vendors use these labels loosely. Here is what each one usually means, and the question that tells them apart.

A quick test for "AI-native": list what a person can do in the CRM, then check how much of it the AI can do through a tool. If the AI can only read, or only writes text that you paste somewhere, it is an add-on. The agentic CRM guide goes deeper into agent actions and the approvals around them.

Eight questions that sort AI CRMs quickly

Ask these in a demo or look them up in the documentation. The third column shows how Customermates answers, so you can see what a concrete answer looks like.

MCP, the Model Context Protocol, is an open-source standard for connecting AI applications such as Claude or ChatGPT to external systems. With an MCP server, the AI client you already pay for works in the CRM directly.

A ten-minute trial test

Run the same test in every trial, with your own data instead of demo records. The Customermates trial runs 7 days with Pro features, so the inbox is part of it.

1. Load a few real records

   Import 20 contacts and five open deals. In Customermates, import runs from an XLSX workbook, one list at a time.

2. Give one compound instruction

   Paste a real client email and write: "Find the deal for this client, move it to the next stage, add the three main points of this email to the deal's notes and save a reply as a draft. Do not send it."

3. Check what actually changed

   Open the deal and look at the stage, the notes, the activity history and the draft. Count what the AI did, not what it said it did.

4. Ask a question that needs arithmetic

   "Which open deals have the highest value and no message from us in 14 days?" A good answer covers every open deal, not the first page of a list.

5. Try a deletion

   Ask it to delete a test contact and see whether it stops for your approval.

How the large CRM vendors package AI

If your team already works in one of these CRMs, switching on its AI is usually cheaper than migrating. The table names each vendor's AI and how it is billed, as described on the vendor's own pages.

Microsoft says the Sales agent in Microsoft 365 Copilot will become the preferred way to use Copilot in Dynamics 365 Sales. If you plan to connect Claude or ChatGPT to an existing HubSpot account, the HubSpot MCP guide covers what that connection can read and write.

Customermates as an AI-native CRM

Customermates is an open-source CRM built in Germany for small client-facing teams: consultants, agencies, other service businesses and founders. Contacts, organizations, deals, services and tasks live in one workspace; on Pro and higher, the email, LinkedIn and WhatsApp conversations about them sit alongside. AI reaches that data in two ways, and self-hosting changes which of them you get.

Mate is the assistant inside Customermates Cloud. It reads your workspace Knowledge Base, where you keep company facts, processes and tone of voice, and cites the pages it used. For questions that need arithmetic across many records, such as the median deal value per lead source, it runs a read-only analysis over the matching records instead of reading one page of results.

Your own AI client connects to the MCP server: Claude and ChatGPT through a custom connector with OAuth sign-in, Claude Code, Codex, Cursor and Gemini CLI with an API key. The server asks clients to confirm before deleting or sending, but whether you see a confirmation depends on the client.

A worked setup for a five-person consultancy

• Deal stages: Lead (10%), Qualified (30%), Proposal sent (60%), Won (100%), Lost (0%). The percentages are stage probabilities and drive the weighted pipeline.
• Custom field on deals: Source, a single choice field with Referral, LinkedIn, Website and Event.
• Saved view: "My open proposals", filtered to the stage Proposal sent and to deals assigned to me, sorted by value. Saved views are personal, so each consultant saves their own.
• Inbox: each consultant connects 1 account on Pro, usually the work email or LinkedIn.

Each consultant then adds two routines. Both run in Customermates Cloud and use hosted AI credits.

The first, "Proposal follow-ups", runs On a schedule, Every day at 09:00:

Look at the open deals assigned to me in the stage "Proposal sent".
For each deal, read the latest conversation with its contacts.
If our message was the last one and is more than five days old,
save a short follow-up as a draft in that thread, in the language
of the conversation. Do not send anything.
End with a list of the drafts you saved.

The second, "Won handover", runs When something changes, on the event Record updated for the Deals list, with Changed fields set to the stage:

If this deal is now in the stage "Won", create a task
"Kick-off: [deal name]", link it to the deal and assign it to the
deal's assigned users. Then add a five-line summary of the latest
conversation to the deal's notes. If the stage is anything else,
do nothing.

A routine runs as its owner, with the owner's channels and permissions. An unattended run refuses everything that would need approval, and changes made by a routine never trigger another routine. Mate sends messages when asked to, which is why the first instruction says "Do not send anything".

What it costs: on Pro, this team pays €145 per month for the inbox, 1 connected account and up to 5 routines per person, with monthly hosted AI credits per active user for Mate and the routines (see pricing). Starter costs the same team €60 per month and includes Mate, MCP and 1 routine per person, but no inbox, so the follow-up routine would have no threads to read. Billing is monthly.

What Customermates does not do

• No lead scoring, no data enrichment and no revenue forecasting. Weighted pipeline values come from the stage probabilities you set.
• No email sequences or bulk sending. Routines have no recipient lists or sequence steps, and Mate sends individual messages.
• No calendar booking. Pro and higher read connected calendars without writing to them.
• No native mobile app.
• No inbox, Mate or routines when self-hosted. Self-hosting gives you the AGPL core with REST, webhooks and MCP.

If predictive scoring for a large sales team is why you want AI, Customermates is not the fit: look at Zoho CRM with Zia or at Salesforce, whose own pages describe predictive features.

Frequently asked questions

Vendor details checked on 8 October 2026 against Salesforce Agentforce pricing, HubSpot's AI CRM page, HubSpot Credits, Pipedrive AI, Zoho CRM Zia, Copilot in Dynamics 365 Sales and Attio pricing.