Your AI agent is a new hire who has no fear of merging records

Think about Claude connected to your HubSpot as a new RevOps hire.

A remarkable one. It works 24/7. It needs no onboarding. It never gets bored of a repetitive task. It can scan your entire pipeline before you've finished your coffee.

It also has some traits you would never tolerate in an actual employee:

• Zero fear of merging records.
• It will edit a property even if you didn't ask it to.
• It treats every field as the truth, including the one labelled OLD - do not use.
• It will hand you a beautiful report built entirely on bad data, and look proud doing it.

You would not give a person like that admin access on day one. So why do we keep doing it with agents?

"Headless" is usually theater

The pitch for fully autonomous, headless AI is seductive: the agent does all the work, the human is redundant, you wake up and it's done.

In production, on real CRM data, that's mostly theater. Not because the model is dumb – because the agent has no idea which of your conventions are load-bearing and which are scar tissue. Your team knows Lead_Score_NEW is the real one and Deal_Stage_old_2024 is dead. The agent sees two valid properties and picks one. Maybe the wrong one. Then it acts on that choice across 50,000 records.

But the opposite extreme is also theater. I've sat there clicking "approve" on changes I couldn't actually evaluate, pretending to be a checkpoint. That's not human-in-the-loop. That's a human being a rubber stamp.

Human-in-the-loop only works when the human can actually judge the loop. If you can't evaluate the decision, you're not a checkpoint. You're a delay.

The job isn't to keep a human in the loop for its own sake. It's to design a loop a human can meaningfully supervise.

How you onboard an agent

Same way you onboard a good hire who's fast but doesn't yet know your business.

Scope its access. A new hire doesn't get the keys to every object on day one. Neither should the agent. Give it the records and properties the task needs, nothing more. Blast radius matters more than convenience.

Give it the context your team carries in their heads. The tribal knowledge that made your portal usable for humans – which properties are real, what "Qualified" means here, which pipeline is the source of truth – is exactly what the agent is missing. Write it down. The agent can only respect rules it can read.

Make its work reviewable in batches. Don't approve line by line what you can't read. Have it run, then surface what it did, what it changed, and crucially what it refused to decide – so a human reviews judgment calls, not keystrokes.

Let it tell you when it's unsure. The best agent behavior isn't doing everything. It's flagging the records where it shouldn't. A hire who says "I wasn't sure, so I left these 200 for you" is more valuable than one who confidently guesses on all of them.

The trust is earned the same way

You don't hand a new hire the production database because their resume looked good. You give them a scoped task, you check the output, and the leash gets longer as the trust gets earned.

An agent is no different. It's the most capable, least careful colleague you'll ever manage. Treat its access, its context, and its review like you would any new hire with a lot of power and not much judgment yet – and it becomes the best one on the team.

Skip the onboarding, and you've just given admin rights to someone who has never seen your business and is afraid of nothing.