Program report · filed November 20, 2026
AI in SalesMeasured report
Meta Muse review shows customer agents grading brand messages
Paul Meinshausen’s Meta Muse test exposed a harsh standard for brand messages: customer agents could judge relevance and penalize over-messaging across every send.
By Rebecca Stone3 min read662 words
Program notes
- Paul Meinshausen asked Meta’s Muse to review marketing-brand notifications during a conference talk in New York.
- Meinshausen is Head of Agentic Decisioning at MoEngage and co-founder of Aampe.
- The account gives no sample size, test protocol or human-review comparison for the Muse experiment.
- Meinshausen defines intelligent activity as taking action, observing a result and updating the action.
- He expects AI decisioning to reach lifecycle and CRM marketing, then orchestration and content production.

By Rebecca Stone
When Paul Meinshausen loaded Meta’s Muse to review notifications from marketing brands while listening to a conference talk in New York, the agent’s verdict was harsher than he expected. “It gave a rundown on them that was sort of even more ruthless than I would have,” said Meinshausen, Head of Agentic Decisioning at MoEngage and co-founder of Aampe.
The test offers a concrete example, not a benchmark. The account provides no sample size, brand list, test protocol or comparison with human reviewers. It shows how one personal agent can read a stream of commercial messages and apply one customer’s standards. The episode therefore supplies a directional case, not proof of market-wide behavior.
What changes when customers bring their own agents?
For marketers, the mechanism is straightforward: when a customer’s agent reviews every message, relevance determines whether a brand earns attention. Meinshausen compares the shift with human assistants, which executives have long used to watch over their affairs, sometimes with more obsessive attention than they give themselves. AI agents make that form of delegated oversight available to everyone.
He also identifies a difference in attention. People divide focus among a job, children, a partner, family and friends, giving each task a small share. An agent can remain with one decision longer. In Meinshausen’s assessment, the shift is beneficial because it raises expectations for detail, relevance and personalization, but it also increases the workload for brands and the people they employ.
Customer-side agents make relevance a requirement, not a campaign preference.
Why send volume is not the goal?
If an agent can choose the message, channel and timing for each customer, the obvious temptation is to send more. Meinshausen rejects volume as the objective. He describes messaging across every channel as the previous wave of automation, whose purpose was to make machines perform tasks rather than improve decisions.
“Activity should be intelligent activity,” he said. He defines it as “the ability to take an action, observe a result, and then update the action.”
That definition changes the marketer’s job. Brands must give agents authority to send and a way to read the response. For a retailer, a purchase is only one signal. The wider evidence can include:
- How customers read the material
- How they click around
- Direct feedback about relevance
These signals can appear before or after a purchase and give the agent information for its next action.
The operating loop is act, observe and update, not maximize send volume.
How can an agent avoid becoming a nuisance?
People learn tools such as ChatGPT by prompting them and then correcting them. Meinshausen says businesses need their agents to be correctable in the same way, sometimes through direct, explicit feedback and sometimes through indirect, implicit behavior.
Negative reactions also count as responses, which keeps an agent from becoming a nuisance. A system should learn from:
- An explicit dislike
- An unsubscribe
- Nonresponse when a response was expected
An agent that keeps messaging despite those signals has not closed the learning loop. The principle matches the suppression and next-best-action approach CX Today has previously explored for reducing personalization fatigue.
The system must treat rejection and silence as data, not noise.
Where decisioning goes next
Meinshausen expects AI decisioning to arrive quickly in lifecycle and CRM marketing, then spread into orchestration and content production. He acknowledges that teams have too much to do and must prioritize and calibrate which adoption decisions come first, but he rejects waiting as an option.
He recommends a practical review: consider how a customer’s agent would grade the previous week’s messages and what it would say. Over the next phase, lifecycle and CRM teams will need to decide which messages, channels and timing decisions to delegate, while their agents learn from clicks, dislikes, unsubscribes and silence before changing the next send.
A useful agent improves the next send by learning from what happened after the last one.
via cxtoday.com (Source)
More from Rebecca Stone
Show full bio
Market editor covering people, launches and day-to-day practice in Sales technique and methodology, sales enablement at Quota Brief.
139 articles
Also rated
- AI Agents Are Now the Shoppers, and Brands Are Rewriting the Pitch
- US Brands Rewrite Sales Pitches as AI Agents Start Shopping
- Brands Now Sell to Two Audiences: Humans and AI Agents
- Dell Product Lead Maps Four Fixes Retailers Need Before AI Agents Shop
- Agentic AI Is Collapsing the Sales-Marketing Boundary