AI CRM: what it is, what AI-native changes, how to choose

An AI CRM is a customer relationship manager where software agents, built on large language models, read what reaches your firm and write it into your client records. A classic CRM waits for someone to type: the call note, the new address, the date on a signed contract. An AI CRM lets an agent take that work: it reads an email thread or a scanned document, finds the household or company it concerns, updates the fields, and prepares the next task for a person to check. Salesforce defines it on its own page as a CRM "that integrates artificial intelligence (AI) to automate tasks" (Salesforce, read October 2026). That definition covers a wide range of AI-powered CRMs, from a chat box that summarizes a record to the agentic CRM, where agents run a whole intake. For a firm with real client files, the definition matters less than two questions: which changes to your data does the AI make without asking, and can you see afterwards who made each one?

An AI-native CRM gives its agents the same records, permissions and history as your staff; a CRM with AI bolted on adds an assistant beside a database designed for people to type into. Attio draws the same line in its July 2026 ranking, read in October 2026, contrasting AI "bolted onto a system designed for humans to type into" (Attio). When a client sends a new proof of address, a CRM with AI bolted on still needs someone to file it and update the record. In an AI-native CRM, an agent can read the document on the record and update its fields. The table compares the two designs on the points that decide whether you can hand that work over.

In Standards, agents read the PDFs and scans attached to a record, check a document and copy its information into the record's fields. A supporting document is stored in a field that carries the properties you declare, such as its type and expiry date. A scanned identity document or a signed contract goes through optical character recognition first, so the agent reads the text of the scan. The same agents can read, search, create and update records, documents and files, one at a time or in bulk. You can also ask an agent, in plain words, to find records or answer a question about a client file, and it searches the records you are allowed to see before it answers. Standards agents act with the permissions of the person who starts them: an agent started by an adviser sees what that adviser sees.

An agent starts when a record is created, updated or deleted, when a form is submitted, on a webhook, on a Slack message or on a schedule. An incoming email does not start an agent by itself in Standards: copying its attachments to a record is a tool the agent calls during its work, inside a conversation or a routine you set up.

On the email side, Standards connects a Gmail inbox and files each thread on the records of its participants. On the Standards cloud, as of October 2026, Gmail and Google Calendar are the email and calendar connectors available; Outlook runs on a dedicated instance or on your own infrastructure.

An AI CRM should ask before any action that removes data or leaves the firm: a deletion and an outbound send. Edits to your data sit in between, and you decide whether they wait for a go-ahead. In Standards, the two modes in which an agent acts on your data are labeled Manual and Auto in the app. Manual "asks for your approval before each change to your data". Auto, the default, "changes your data without asking". By default, in both modes, a deletion or a send stops and waits for a person, and the approval card states its impact, such as the number of records a deletion would remove. Switching to Auto does not lift that stop. In Standards, as of October 2026, a deleted record goes to the trash and can be restored for 30 days.

None of the three pages Google ranked first in the US for "AI CRM" in October 2026, the AI CRM pages of Salesforce and HubSpot and Attio's July 2026 ranking, describes a rule of this kind. Salesforce describes agents responding to "common customer questions without oversight" and a governance layer of "permissions, guardrails, and audit trails" (Salesforce), but no list of actions that always wait for a person. HubSpot writes that "every AI-driven update stays visible, editable, and under human control" without naming an approval step (HubSpot).

An AI CRM you can audit keeps, on each record, who changed what and when, and says whether the author was a colleague or an agent. Standards shows that history on every record, with the value before the change, so you can see that an agent replaced an address on Tuesday and what the address was before. That matters the day a client asks why their file says something they never told you, or the day a compliance review asks who updated a document date. HubSpot offers a centralized audit log that tracks, in its own words, "every action taken by humans, automations, integrations, and AI" (HubSpot, read October 2026). The question to put to any vendor is where that history lives: on the record your team opens every day, or in an admin console only an owner reads. A history kept on the record lets the adviser who works the file correct an agent's mistake in the same place, without asking an administrator to search a log.

Choose an AI CRM by what its agents may do on your client files, and by what they must ask you first. Ask each vendor to answer the five questions below in writing, with the setting names, before a trial.

1. Data model. Can you define your own objects, such as households, contracts or policies, and their relations, or must your firm fit a sales pipeline?

2. Documents. Do the agents read the PDFs and scans attached to a record, or only emails, calls and typed fields?

3. Approval. Which actions does the AI take without asking, which ones always wait, and what is the default?

4. History and permissions. Does every change name the person or agent behind it, and does an agent act with the permissions of the person who started it?

5. Data handling. Where are your data and the AI calls hosted, and does the vendor let model providers train on them?

Standards' own answers on model training, hosting and billing are in the questions at the end of this page.

Standards fits a firm of five to fifty people whose work runs on client files, with their meetings, emails and supporting documents. You define the objects those files live in, such as households, contracts or assets, and their fields. Developers reach the same data through an SDK, a CLI and an MCP server, described on the developer page, so Claude, Cursor or your own scripts read and write the records your team uses. If your client data still sits across a spreadsheet, a drive and several inboxes, start with our guide to centralizing client data. Standards has no bulk email, landing pages or SEO tools, so a marketing-led team will keep a separate tool for campaigns. Outlook needs a dedicated instance or your own infrastructure. As of October 2026, setup happens with Standards engineers, and self-serve signup is on a waitlist.