Two weeks ago I sat across from a man who built his company before CRMs existed. Calm voice, experienced hands, someone who did most things himself. In a few months he hands the company to his successor. He didn’t call me about a purchase. He called because he’d read somewhere that HubSpot is “no longer a CRM, but an agentic customer platform” - and it bothered him.
“Igor,” he said, tapping a finger on the sheet where he’d written the phrase down. “Explain this to me. Without your marketing babble. What is my successor actually supposed to do with it?”
I liked the question. Because it was honest. And because behind it wasn’t fear of technology, but something else: the discomfort of handing over something he didn’t fully understand himself. And yet this man understands it better than most thirty-year-olds who “grew up with AI”. He just doesn’t know it yet.
The notebook that started working
A classic CRM is a system of record. It logs who’s who, which deal sits in which stage, when you last called someone. Useful. But dead. The data lies there, waiting for someone to read it and decide. A human has to interpret. A human has to act.
HubSpot started turning that around back in 2024. Under the name Breeze, it built agents into the platform - agents that don’t just record, they work. Let me put it this way: a customer writes in at night with a question about their account. Until now, it waited for morning. Now an agent answers it at three in the morning, without anyone picking up a phone.
Here’s how I explained it to him. Your CRM used to be a notebook. Now it gets employees. Junior staff who don’t sleep, don’t take holidays and don’t ask for a bonus. The word “agentic” means just this: the software makes the first move instead of you.
Why he understands this better than his successor
He looked at me and said: “So I’m hiring people who aren’t people.” Exactly. And his face relaxed.
Because he’s done precisely this his whole life. He decided which task to trust to whom. Who gets a free hand and who stays on a short leash. Which decision he keeps and which he passes on. That’s not a technical question. That’s management judgement, and he has it down cold.
HubSpot even built this into the platform. You decide which tasks the agent handles alone and which stay with a human. Put in his terms: which five decisions the new trainee may make without you, and for which ones he has to knock on your door first. Every good manager draws that line in week one. Here it’s simply moved into the settings.
The real problem is elsewhere. Not that the successor won’t understand AI. But that he’ll trust it too quickly.
The danger isn’t AI. It’s trust.
This man trusts slowly. Before he gives anyone a free hand, he asks what happens if it goes wrong. That’s not suspicion. That’s thirty years of scars. His successor is different: fast, fluent with the tools, used to technology doing what it promises. And that’s exactly the trap.
An agent that never says “I don’t know” is a dangerous colleague. It delivers the right answer and the wrong one with the same confidence. Give it a free hand without oversight and you scale the mistake at the same speed as the success. His instinct to first ask “what if it screws up” isn’t a brake. It’s the most valuable thing he’s passing on. Worth more than all the passwords and access rights combined.
The agent inherits your mess, too
And now the elephant in the room that sales demos walk around. The successor doesn’t just inherit a tool. He inherits a database nobody ever cleaned.
I recently ran into this at a client. They run a well-known online store under their own brand. The company that owns and operates that brand sits in the CRM under a different name. The result: the sales deals are open under one record, the support tickets scattered across both. More than 200 tickets under one, 80 under the other, some attached to both at once. All for the same customer.
A human has no trouble with this. Everyone on the team knows the brand and the owner are the same customer, and makes the connection in their head. The agent doesn’t. It reads the structure literally.
Switch on the agent in support. A customer we’ve worked with for two years asks her next question and gets an answer from an agent that sees the eighty tickets under the brand - not the two hundred under the owner. It replies as if they’re meeting for the first time. Two years of shared history lie unread, filed under the wrong name. A human would recognise a returning customer. The agent won’t - until someone writes that connection into the data.
The numbers I believe, and the ones I don’t
On its product page, HubSpot claims its agents automatically resolve 70% of customer conversations and create four times more contacts. Nice number. I read it carefully, because it comes from the seller. And when I look at the same seller’s fine print, the 70% shrinks: in the announcement of its new pricing, HubSpot cites 65% of conversations resolved and 39% faster resolution across 8,000 customers. When a seller gives you two different numbers in two different places, you know the billboard is not your result.
So I’d rather tell you what we measured ourselves. Our client Leanpay cut the cost per resolved ticket by 98.6% with its assistant Lea, and brought response time down from six hours to under two minutes - numbers Leanpay published itself, on its own blog. Those numbers aren’t magic. They’re the result of clear boundaries and clean data.
One more detail won him over. In spring 2026, HubSpot moved these agents to outcome-based pricing: no licence up front, you pay when the agent completes the task - half a dollar for a resolved conversation, a dollar for a qualified contact. You pay the trainee only when he closes the case. For a man who watches every euro, that’s sound logic. And it forces you to measure whether the agent actually works or just looks good in a demo.
So what should the successor do
I came back to his first question. Three things, I said, and none of them technical.
First, clean and connect the database, because the agent will believe everything it finds in it. Then decide which tasks it may handle alone and when it must hand over to a human - and write it down, don’t guess. And third, measure the result, not the feeling: how many conversations it resolved, how many it botched. If you need someone who has already set this up in practice, I’m here. Write to me at igor.pauletic@frodx.com and we’ll look at a concrete case: where your “digital trainee” would do the most good - and where the damage.
The man folded the sheet and tucked it into his inside pocket. “I know how to do this,” he said. “I do it every day.”
Exactly. He’s not buying a CRM. He’s hiring. The successor will learn the software in a week. Knowing when to say “no, wait” isn’t in any settings menu.
