What is an AI native CRM?
An AI native CRM is designed with an AI as its main user. Its tools, limits and safeguards assume the caller is a model: writes come in batches, retries do not duplicate records, long lists come in pages, and text from outside is marked as data. An AI native CRM is built for that caller, not given an add-on.
By Alfonso Trocolí, founder · Updated
| Trait | What the AI gets | Where it lives |
|---|---|---|
| Tools designed for an AI caller | 140 tools, filtered by the member's role; a read-only link sees no write tools | crm_schema, then the tool list for the connection's role |
| Batch writes | Up to 100 rows per call, with a result for each row | create_leads, update_deals, create_tasks |
| Safe retries | The same idempotencyKey with the same arguments within 24 hours returns the first result | An optional idempotencyKey on every write tool |
| Cursor pagination | A nextCursor on long lists; the list is complete when it is null | list_leads, list_deals, list_tasks |
| Tool annotations | Each tool says whether it only reads, whether it is destructive and whether it reaches outside | readOnlyHint, destructiveHint and openWorldHint in the tool list |
| Outside text marked as data | A notice on each result that its text is data, not instructions | The first lines of every tool result |
| Audit log | Who called which tool, on which records, with what result | list_mcp_activity |
| Rate limit | 120 calls a minute per connection, 20 of them for tools that reach outside Sashimi | The connector's per-connection limit |
What makes a CRM AI native?
A CRM is AI native when its main user is an AI and it is designed around that. The tools, the limits and the safeguards all assume the caller is a model, not a person clicking through screens or a developer writing a script.
A model behaves differently from both. It works in batches, it retries when a call times out, it reads long lists in pieces, and it reads text that strangers wrote, such as emails and notes. A CRM built for an AI handles each of those on purpose.
Is an AI first CRM the same as an AI native CRM?
Not quite, though the terms are used loosely and often interchangeably. AI first usually describes an app that leads with AI features: an assistant in the sidebar, summaries, agents inside the product. AI native describes how the system is built: an AI is the main caller.
An AI first CRM can still be an app you click through. An AI native CRM may have no app you need at all, which is why the idea overlaps with a headless CRM.
Why is having an MCP server no longer enough?
Because almost everyone has one. By October 2026 most major CRMs have an MCP server or a Claude or ChatGPT connector (as of 2026-10-18). Working with Claude or ChatGPT no longer sets a CRM apart.
What still differs is the design behind the connector. An MCP server can be a thin wrapper over an existing API, or a set of tools designed around how a model works. You cannot tell them apart from a list of logos. You can tell them apart from the traits below.
For Claude in particular, the guide to the best CRM for Claude sets out what to check in a CRM's MCP.
What does a CRM built for AI do differently?
It gives the model tools that match how a model works. These are the traits to look for when you compare products, whatever the vendor:
- Tools designed for the caller. A schema tool first, so the model learns the fields and stages before it writes. The tool list shrinks to what the member's role allows.
- Batch writes with per-row results. One call creates or updates many records. Rows that fail come back with their reason, and the rest are saved.
- Idempotent retries. A write can carry a key. Repeating the same key and arguments returns the first result instead of writing twice.
- Cursor pagination. Long lists come in pages, with a cursor that says when the list is complete.
- Tool annotations. Each tool says whether it only reads, whether it is destructive and whether it reaches outside, so the client can ask before risky calls. They are hints: whether the client asks is up to the client.
- Outside text marked as data. Notes and emails come back labelled as data, not instructions, to resist prompt injection.
- An audit log. Every AI call is recorded with who, which tool, which records and the result.
- Rate limits per connection. One runaway loop cannot flood the workspace.
Who pays for the AI in an AI native CRM?
It depends on the vendor. Many AI CRMs meter their AI in credits. As of 2026-10-18, HubSpot's Breeze agents are paid with HubSpot Credits, and Attio's plans include seat and workspace credits (from their pricing pages).
If the CRM runs on the AI you already pay for, there are no AI credits to buy. The reasoning happens in your own Claude or ChatGPT plan, and the CRM only stores and checks the result.
How does Sashimi work as an AI native CRM?
Sashimi was designed around these traits: it is a sales CRM built as an MCP server, with no UI by default. The table above maps each trait to where it lives in Sashimi.
In practice:
- The connector has 140 tools. A connection only sees the tools its member's role allows, and a read-only personal link sees no write tools.
- Every business you run has its own workspace, and a connection reaches only one of them.
- Bulk create and update tools take up to 100 rows per call and report a result for each row.
- Every write tool accepts an optional
idempotencyKey. The same key with the same arguments within 24 hours returns the first result. - List tools return a
nextCursor, and the list is complete only when it is null. - Each connection can make 120 calls a minute, and at most 20 a minute to tools that reach outside Sashimi.
list_mcp_activityshows every AI call.
Sashimi runs on your own Claude or ChatGPT plan, so there are no AI credits to buy. The only credits are Apollo data credits, used for prospect import and enrichment on your own Apollo key, with an estimate before you spend. The Sashimi MCP server page lists the tool families, and the security page covers roles and the activity log.
You connect it from a Claude plan that allows custom connectors, or from ChatGPT with developer mode on the web; on Claude Team or Enterprise, an owner adds it. The steps are on how Sashimi connects.
Access is by invitation during the private beta, and it is free for now. The Solo plan, with 3 workspaces, is planned at $15 a month at launch. Nothing can be bought today.
Related: what a headless CRM is, and what an agentic CRM means when you bring your own agent. Setup guides are for Claude and ChatGPT. Comparing products? See the CRM alternatives, or go back to the home page.
Frequently asked questions
What is an AI native CRM?
It is a CRM designed with an AI as its main user. Its tools, limits and safeguards assume the caller is a model that sends batches, retries after timeouts and reads long lists.
Is an AI first CRM the same as an AI native CRM?
The terms are used loosely and often interchangeably. A useful distinction: AI first usually describes an app that leads with AI features, while AI native describes a system built for an AI to operate it.
How is a built-for-AI CRM different from a bolted-on MCP server?
A bolted-on MCP server wraps an existing app's API so an AI can reach it. A built-for-AI CRM designs the tools around the model: batch writes with per-row results, idempotent retries, cursor pagination, tool annotations, outside text marked as data and a log of every AI call.