Paul Meinshausen asked Meta's Muse to check his notifications from marketing brands... the verdict was harsher than he expected.
In a conversation with CX Today, the Head of Agentic Decisioning at MoEngage and Co-Founder of Aampe, said:
"It gave a rundown on them that was sort of even more ruthless than I would have."
He had loaded up the personal AI agent while listening to a conference talk in New York, curious whether brands were responding to customers who now have software reading on their behalf.
For marketers, that small experiment points to a bigger problem. If a customer's agent is reviewing every message, relevance stops being a nice-to-have.
Everyone Is Getting an Assistant
Meinshausen believes the shift is bigger than a new channel. Executives, he noted, have long had human assistants who watch over their affairs "even more obsessively" than they do themselves. Agents are now making that available to everyone.
"Most people are not just a professional in their job, they're also a parent or a partner or family member and a friend," he said. They give each task a small amount of focus, whereas an agent can spend far longer on a single decision.
He believes that the result is a higher bar for brands:
"The expectation for detail, for relevance, and for personalization is just a lot higher."
"It's a good thing, but it just means that the rate of work has increased as well for everybody involved."
Activity Is Not Intelligence
If agents can choose the message, channel, and timing for each customer, what stops brands from simply sending more? Meinshausen's answer is that volume is the wrong goal.
Messaging across every channel is, in his view, "the previous wave of automation," where the aim was just to make machines do things.
"Activity should be intelligent activity," he said, which he defines as, "the ability to take an action, observe a result, and then update the action."
For marketers, that means giving agents a way to read the response, not only the authority to send.
A retailer wants a purchase, but Meinshausen pointed to a much wider set of signals before and after it. These include how customers read material, how they click around, and direct feedback such as whether something was relevant.
He compared it to everyday use of tools like ChatGPT, where people learn to prompt and then correct the system. Businesses, he said, need their agents to be correctable in the same way, "sometimes directly and explicitly and sometimes indirectly and kind of implicitly."
Why Learning Stops Over-Messaging
The same loop is what keeps an agent from becoming a nuisance. A negative reaction still counts as a response.
"One form of response can be a negative form," Meinshausen said. "It can be saying I don't like this. It can be unsubscribing. It can be not responding when you kind of were expected to respond. And those are really important to learn from."
In other words, the agent that over-messages is the one that ignores those signals.
CX Today has previously explored how suppression and next-best actions can reduce personalization fatigue, and the principle is similar.
No Time to Do Nothing
Looking ahead, Meinshausen sees "AI decisioning" arriving quickly in lifecycle and CRM marketing, and spreading into orchestration and content production.
There is, he admitted, "too much to do," so teams will need to prioritize and calibrate which adoption decisions they make first.
But he was clear on one point: "There's definitely no time to do nothing."
For marketers contemplating their own messaging strategies, consider how a customer's agent would grade your last week of messages and what it would say.