"Attio has introduced a new CRM architecture that lets AI agents work alongside revenue teams on the same customer data," @testingcatalog posted on October 15, listing an SQL engine that runs joins and calculations inside the CRM, live connectors for BI tools, and "agent scale" of up to 100 million records per object. Attio CEO Nicolas Sharp called it "a new shape of software" in the launch post: "the first hybrid architecture engineered for people and agents in equal measure."
Most owner-led service businesses will never need a hundred million records. But the reaction on X was not really about scale. It was about a shift in who uses a CRM. As @DReshtei put it: "the next saas user might never open your app... if your 'ai teammate' needs a stale csv to answer a question, you've built an expensive intern with amnesia." That line applies to a 12-person landscaping company on HubSpot as much as to an enterprise sales team.
TL;DR
- Attio's October launch added an SQL engine, live data connectors, agent-scale capacity and an activities beta, so agents can query CRM data directly under the same permissions as people.
- The lesson for small businesses is not "switch CRMs." It is that agents are only as good as the structure and freshness of your customer data.
- Before any agent reads your pipeline, fix the fields it will rely on: stage, source, owner, last contact, and one record per customer.
- Let agents read and draft first. Writes to the CRM, and anything that reaches a customer, get reviewed. AI drafts, your team approves.
What Attio actually shipped
Attio's launch post lists four capabilities: SQL Engine, Data Connectors, Agent Scale, and Activities in beta. The SQL engine lets agents and people ask questions that used to need an export, with examples such as "find which lead sources produce the highest win rates on deals over $50,000" and "identify open opportunities associated with churned customers." Attio says every query can be inspected, and that SQL also works inside its Workflows, so a result can trigger an action.
In an engineering post, CTO Alexander Christie describes the engine, code-named Centaur, as Postgres-compatible and transactionally consistent, so many agents can query at once without conflicts. Agents reach it through Ask Attio, MCP, the REST API and read-only connectors. Attio also stresses that agents work "with the same permissions and access controls that govern the rest of Attio."
Strip away the architecture and two design choices stand out. Agents read through a structured query layer instead of scraping screens. And they inherit permissions instead of getting a master key.
Do not migrate because of a launch
The most useful counterweight came from SaaStr's Jason Lemkin, who posted that when they asked "10K" what CRM to use, "it said to stick with Salesforce," with HubSpot and Attio as options only if starting over, and "most importantly, it said not to leave or migrate."
That is sound advice for a small business too. The CRM you already use almost certainly has an API, and most major ones now offer an official connector for ChatGPT or Claude. Migrating costs months of messy data and retraining. Cleaning the CRM you have costs a few afternoons and pays off whether an agent ever touches it or not.
What makes a CRM usable by an agent
An agent asked "which lead sources close best for us?" can only answer if the data supports the question. Here is what to check:
| Field or practice | Why an agent needs it | Common small-business gap |
|---|---|---|
| One record per customer | Otherwise totals double count | Duplicates from web forms and imports |
| Deal stage with defined meanings | Pipeline math depends on it | Stages used differently by each person |
| Lead source, filled every time | Answers "where do good jobs come from" | Blank on half the records |
| Owner on every open deal | Routing and follow-up drafts | Deals owned by people who left |
| Last contact date that updates itself | "Who has gone cold" questions | Updated only when someone remembers |
| Won or lost reason | Teaches the agent what wrong looks like | Rarely recorded |
| Notes in the record, not in texts | Context for drafted replies | Real history lives on phones |
None of this is new CRM advice. What is new is that the cost of skipping it went up. A person who knows the history can work around a blank field. An agent fills the gap with a guess.
A practical cleanup plan
Most service businesses can do this in two or three sessions:
- Merge duplicates. Start with contacts that share an email or phone number.
- Write down what each stage means. One sentence per stage, agreed by whoever moves deals. Retire stages nobody uses.
- Make lead source required on new records. Backfill the last 90 days by hand. Older history can stay blank.
- Reassign orphaned deals. Every open deal gets a current owner.
- Turn on automatic activity logging. Connect email and calendar so last contact updates itself. As one reply on X noted, getting people to keep a CRM updated was always its hidden cost.
- Add a lost reason field with five fixed options. Free text is fine as a second field, but the fixed list is what an agent can count.
- Test with three questions. Ask an assistant connected to your CRM questions you already know the answers to, such as last quarter's top lead source. If it is wrong, the data is the problem, not the AI.
Permissions are part of the data
Attio's point that agents inherit "the same permissions and access controls" as people is worth copying even on a basic CRM plan. When you connect an assistant, it acts as whichever user authorized it. If the owner connects it with an admin login, the agent can see and change everything the owner can, including commission notes, HR-adjacent records and every deal in the company.
A safer setup takes a few minutes:
- Create a dedicated user for the integration instead of connecting with the owner's account, if your plan allows it.
- Give that user read access to contacts, companies and deals, and nothing administrative.
- Keep private notes, pricing exceptions and employee records in fields or objects that user cannot read.
- Review the connector's granted scopes once a quarter, and remove any you are not using.
If your CRM plan does not support a separate user, connect with the account that has the narrowest access that still answers the questions you care about.
Read first, write later
Attio's examples are mostly questions: win rates, weighted pipeline, churn links. That is the right order for a small business too. Let an agent read and summarize before it writes.
- Read: "Which estimates over $5,000 have had no follow-up in 10 days?"
- Draft: "Write a follow-up for each, using the job notes."
- Write, with review: "Update the stage and log the follow-up once the estimator approves the message."
The first two are low risk and save real time. The third changes records other people rely on, so it waits for a person until the drafts have been right for a few weeks. That is the rule SpiAI builds every client workflow around: AI drafts, your team approves.
FAQ
What is a CRM for AI agents?
It is a CRM whose data agents can query and act on directly, through an API, MCP connector or query layer, under the same permissions as people. Attio's October 2026 launch is one example of a CRM rebuilt around that idea.
Do I need to switch to Attio to use AI agents?
No. Most major CRMs, including HubSpot and Salesforce, offer APIs and connectors that assistants can use. Cleaning and structuring the data you have usually matters more than switching.
What did Attio launch in October 2026?
An SQL engine for querying CRM data, live data connectors, agent-scale capacity of up to 100 million records per object, and an activities feature in beta.
What CRM data does an AI agent need most?
One record per customer, defined deal stages, lead source, an owner on every open deal, automatically updated last-contact dates, and a fixed list of won and lost reasons.
Should an AI agent update CRM records on its own?
Start with read-only questions and drafted follow-ups. Let it update records after a person approves, and only remove that review for low-risk fields once results have been reliable for several weeks.
How do I test whether my CRM is ready for an agent?
Ask a connected assistant three questions you already know the answers to, such as last quarter's top lead source. Wrong answers usually point to missing or inconsistent data.
Is it safe to connect ChatGPT or Claude to my CRM?
Use the CRM's official connector, grant the narrowest permissions that work, and keep customer-facing messages and record changes behind human approval until you trust the results.