I recently turned off all my Muse notifications.

I’m sure I am one of many people who downloaded Meta’s Muse, tried it on a few use cases, and segmented themselves into an engagement degradation cohort on their way to churn. Part of my displeasure with Muse was its continuously unhelpful barrage of proactive notifications. Muse would read my emails and tell me someone had responded, 4 hours after it hit my inbox, as if I didn’t respond on my own in minutes. Muse would suggest changes to my overlapping schedule, as if I’m unable to manage my own calendar. Muse would provide me summaries of dinner reservation possibilities for an upcoming trip, as if I didn’t know exactly where I wanted to go. All of it was annoying and unhelpful. My annoyance made me distrust and dislike Muse more, and I ultimately stopped going to Muse as my daily driver personal AI agent and turned notifications off entirely.

I love using Town for work (we are big Townies at Gigi). One thing we admire about Town is the proactive, seemingly daily enrichment that Town performs behind the scenes on me, my team members, and our work together. It’s clear to me the more I use Town, the smarter and more personalized Town becomes. And while proactive background enrichment is where Town excels, Town, like Muse, fails at proactive actions. Recently Town began suggesting tasks and routines that I should create based on its perception of my upcoming work schedule. I’ve found zero of these suggestions helpful, and, similar to my experience with Muse, I would like to turn them off entirely. Worse, for an engaged, paying, and (clearly) advocating customer, they are degrading my experience and making me lose trust in the product.

Agent proactivity is hard.

The current trope is that most people are using AI agents as glorified search engines and are underutilizing AI’s vast capabilities. In legacy SaaS, you knew the features a product had because someone had to design and implement a prescriptive workflow to ensure adoption of a new feature. Agents work the opposite way. Their tools sit behind a simplified natural language interaction, so many of them stay invisible to early users of a product. A natural way to combat this is to proactively nudge users into realizing new capabilities, and through that nudge extract more value from the AI agent. But there’s clearly a slippery slope between nudging customers into value and annoying them to the point that trust is lost.

Underutilization of capabilities is a problem we see firsthand at Gigi. We recently visited a highly engaged and super happy customer who said, “It would be awesome if Gigi could bulk build campaigns with our exact naming convention and strategic structure from a template that we upload,” to which our team said, “Gigi can do this today! And has been able to do this for months.” Obviously our release note emails, Slack messages, and calls around new product releases weren’t doing enough. People need to see and feel the feature (like they did in legacy SaaS).

An interim solution that we’re implementing is conversation starters: contextually aware pills at the top of the chat experience that guide our customers to engage with use cases that we know are both high-value and underutilized. When discussing adding these conversation starters, a teammate opined that she “genuinely believed proactivity and timing could solve this in a more elegant manner.” In theory, she’s right. But I told her: “I think that you overestimate the proactivity capabilities that we can achieve in a delightful, non-invasive way.” While I believe that proactivity can provide a magical agentic experience, I’ve yet to see it implemented with success. Not to say that we can’t be one of the first to nail it, but if Meta and Town are struggling, then we shouldn’t assume we can concoct magical proactive agentic experiences for our customers.

I do, though, think it’s clear where proactivity has found its killer use case with agents: proactive context and memory enrichment.

Proactive about what it learns, reactive about what it does. Town became my daily driver because I felt it get smarter every day without once asking me to notice. Muse lost that spot the moment it started asking.