Outbound agents arrive inside your CRM

The agent now lives where your records live

Next month, an outbound sales agent is due to ship inside a major CRM. It will sit beside the customer records your team opens every morning.

The vendor announced it on September 11, 2026. It's in pilot now, with general release set for November. It works from research to outreach, alongside your reps, over weeks and months. [1]

It's also the first agent on the vendor's new long-horizon runtime. In plain words, it can chase one goal for days or weeks, instead of finishing one task and stopping. [1]

Four days later, a data provider said the agent can reach its network of 240 million contacts. Customers get that even if they don't buy from the provider. [2]

So one agent can read your records, find new people and write to them, all from one screen. That's the appeal, and it's also the risk.

An agent built into the CRM is very easy to switch on, and much harder for your team to check.

The launch post says one pilot customer now gets 60% of its sales pipeline from the agent. That's the vendor's own claim. We found no method, time period or definition of pipeline behind it. [1]

Pipeline is also not revenue: a buyer can skip a booked meeting, and a held meeting can still go nowhere.

So judge the agent on your own records, with your own scorecard. Our pilot scorecard scores research, messages, replies and held meetings apart, against the full cost.

My view: pilot it if you want to. But pilot it on one segment, with every message reviewed and the stop rules written first.

Check the data it reads before it writes

Admins in a Reddit forum for this CRM are asking the right question. One says leadership wants a pilot on records full of duplicates, free text and half-filled fields. [3]

Another admin asks which part breaks first when messy form data feeds the agent. Is it the advice, the automatic actions or the summaries? [4]

We learned our own answer the hard way. On September 24, our CRM sync passed every pretend test, then broke in seven places on a real copy.

You don't need to clean the whole database first. Fix only the fields this one task touches. Those are company identity, contact owner, the stop list and what the agent may change.

The readiness post has the full sheet.

A huge contact pool makes this matter more, not less. New people pulled in from outside can copy records you already hold. They can also land on a buyer your team is already talking to.

Size doesn't fix fit, either. In one real lead search, our system judged 1,635 new public posts and kept only 70. Most of what any agent finds should be dropped.

Decide which actions need a person

The vendor says its agent follows guardrails for when it can act alone and when a seller must approve. It also says every agent works within the customer's business rules and permissions. [1]

Those rules don't write themselves. Someone on your team has to decide every single one, action by action, before the pilot reaches a single buyer.

Research can usually run freely. Sending a message, changing a customer record and making a commercial promise each need their own rule.

Some writes can't be taken back. Converting a lead in a CRM is one of them, so a named person should confirm it every time. Our approval map goes through each action.

In Mach, you pick how much of each playbook a person checks. The choices are every message, the first per company, until trusted, or none.

Until trusted means a set number of approvals without edits, from 1 to 20. Start any pilot at every message.

Write the stop rules on day one

A stop rule pauses the agent automatically, so nobody has to notice a problem first. Write yours before the first send, not after the first complaint.

You need four at the very least. Set a spend cap and a limit on research errors. Then make a reply stop that person's steps, and make an opt-out hold forever.

Then test each one on purpose, before a real buyer can set it off. Send a reply at an awkward moment, and check that the next step really stopped. The handoff post lists the cases worth testing.

In Mach, a real reply stops that person's steps and gives your rep a task to answer. Ask the agent inside your CRM to prove it does at least that much.

Don't let someone else own the rules

Here's the mistake to avoid. A big rollout can arrive with the vendor's own engineers doing the setup. That's fast, but you get rules your team didn't write and can't easily change.

An analyst firm warned about this on September 29, 2026. It predicts 70% of enterprises will abandon agentic AI built by a vendor's forward-deployed engineers by 2028. [5]

Forward-deployed engineers are the ones a vendor sends in to build the system for you.

So keep the rules where your team can read and edit them. Who the agent writes to, what it may send and when it stops should all sit with you.

If only the vendor can change them, the pilot is theirs, not yours.

Bring three numbers to the pilot call

Before the pilot call, export the one segment you plan to test. Count the duplicate companies, the contacts with no owner and the people on your stop list.

Bring those three numbers and your written stop rules to the call. How the vendor answers will tell you more than the demo does.