The Best Salespeople Are Rarely Systems People

Meta launched Muse this week. Connect your email, your calendar, a payment method, tell it what you want, and it goes.

That instinct isn’t unique to Meta. xAI just took Grok Bot from beta into enterprise availability, same pitch: assign it a role, connect your tools, expect results. OpenClaw and Hermes, the two open-source frameworks most of the current agent wave is being built on, compete with each other on how many channels they reach and how well they remember how you work, not on whether the system they’re plugged into is telling the truth. NVIDIA built NemoClaw to make OpenClaw safer for enterprise use, real progress, but it governs what an agent is allowed to do, not whether the data it’s acting on makes sense. Salesforce’s and HubSpot’s agents stay inside their own CRM entirely, which solves a different problem: theirs never even sees the parts of a business that live outside the CRM.

We’ve built agents on all four of the major frameworks ourselves, OpenClaw, Hermes, CrewAI, and Grok Bot, and the same thing has been true every time: none of them ask, on their own, whether the record underneath is coherent. That question doesn’t come from the framework. It has to come from whoever’s deploying it.

For a personal agent booking a dinner reservation, that gap doesn’t matter much. Worst case, you get a table you didn’t want. For a business connecting an agent to its own customer data, the calculation is different. The agent doesn’t know the CRM is a mess. It just executes against whatever’s there, faster and at a higher volume than any person would, and it never once raises a hand to say something looks off.

The mess isn’t the salesperson’s fault

Here’s the part that doesn’t get said out loud enough in real estate, or in sales generally: the people who are best at the actual job, reading a room, building trust, closing, are rarely the people who are good at, or interested in, the systems underneath it. That’s not a character flaw. It’s a specialization. Somebody who’s great with people at a kitchen table is not usually the same person who wants to sit down and design a CRM taxonomy. Most sales teams never had anyone whose job it was to do that second thing, so the tags accumulated however whoever was dealing with the problem that week decided to use them, and nobody ever came back to make it coherent.

That’s the actual argument for what a revenue agent should be. Not a faster way to answer emails. The systems person the salesperson never had time to become, so the salesperson can stay exactly where they’re valuable: in the room, with people.

The pattern, wherever you run it

Whichever framework it’s built on, a revenue agent worth trusting needs the same five things underneath it before it touches anything:

• A taxonomy. A written, shared definition of what your records actually mean, agreed on by the people who use them, not inherited from whoever was in a hurry the day a field got added.
• A company brain. A structured record of who you are, what’s true, how you talk, and what you won’t say, the thing an agent actually answers from instead of guessing. We call ours Groundwork.
• CRM access, not CRM control. The agent can see records and add to them. It doesn’t delete, and it doesn’t overwrite, without a person saying so.
• Inbox access, same limits. It reads, it drafts. It doesn’t send anything nobody’s seen.
• Real analytics. Visibility into where people actually go, not just what they typed into a form, so the agent knows what someone’s been looking at before anyone has to ask them.

We’ve run this exact pattern in two places now. Ben Olsen, at BrightWork Realty, named his agent Trevor. Before Trevor could qualify a single lead, the taxonomy had to be built from nothing, tags that had drifted for years into meaning five different things to five different people, and the company brain underneath it had to define everything from how Ben actually talks to where fair housing law draws a hard line on what any agent is allowed to say.

We run the same pattern inside our own company. Ziggy, our internal agent, works inside our CRM and reads Scott’s inbox, off the same company brain that generates our proposals and site content. When a prospect opens a live proposal instead of a PDF, Ziggy already knows which sections they lingered on and whether they came back for a second look, so that shows up before the follow-up call, not during it. Same five things. Different business, different day-to-day, same reason it works: knowledge first, integration second, autonomy last.

Check your own CRM before you connect anything to it

You don’t need to be deploying an agent to find out if you have this problem. Pull up your CRM and pick five tags at random. Ask two people on your team what each one means. If you get two different, equally confident answers for any of them, that’s not something you fix by turning an agent on and hoping it figures it out. That’s the actual first project, whether the thing you’re eventually connecting is an agent, a new hire, or anything else that has to trust what the system tells it.

We call the audit step two. What Trevor does with all of this now is its own story, and the numbers are here. This one’s about the pattern underneath, wherever you’re running it.