Albert is dead. Here's what he taught us.
TL;DR: Albert was a digital phone companion for seniors. Within five weeks, Lukas and I built him, reached out to nearly 1,000 potential partners, and had 13 people talk to him in 172 calls. Then we shut him down. In this post, we share what we learned.
Whether you’re a (future) founder, an investor, or a student, or you simply want to build your own Albert: our code, prompts, and system design are open source. You can find them on GitHub.
Motivation
Our society is aging, and aging brings serious challenges. One of the biggest is loneliness, and the cognitive decline that often comes with it. Chronic loneliness carries a mortality risk comparable to smoking up to 15 cigarettes a day [Holt-Lunstad et al. (2010); Murthy (2023)]. When I volunteered at a nursing home, I saw how many residents sat apathetically in their chairs, simply waiting for time to pass. Who would feel energized in such an environment?
Helping seniors has also been a big part of our own lives. Lukas spent weeks organizing his grandmother’s move from her home to a senior residence. I used to call my grandmother every week to keep her company, even while living on the other side of the planet.
At the same time, technology is making huge progress, yet very little of it is built for seniors. While large companies spend billions on AI agents for businesses and digital-native consumers, very few solutions are tailored to seniors. On top of that, our interviews showed that many seniors struggle with today’s smartphones: their eyesight is too weak to read the screen, their fingertips are too dry for the touchscreen to register a tap, and hidden menu options cause confusion.
We believe that the future interface for seniors is speech, not touchscreens, and that seniors should be able to benefit from recent advances in AI and interactive systems just as much as everyone else.
So why aren’t there more tech products for seniors? It’s hard to say for sure, but there are some clues. A quarter of US adults aged 65 and older never go online [Pew (2021)]. In more than 100 interviews, we also found that many of those who are online use their devices for only a few specific tasks, like playing games (solitaire, chess), making phone calls, and looking at pictures. That’s a paradox: the group that could benefit most from natural interaction with AI is underserved precisely because it is so hard to reach.
Our goal was to build an agent that seniors could reach through an interface they already use: the telephone. We believed that building on something familiar would remove friction and lead to quick adoption by a large number of people.
We were very wrong. More on that later.
The product
Albert was a digital companion for seniors who lived on the phone: no app, no internet, no setup, just a phone number. He was a friendly companion first, and the plan was for him to later become a helper for everyday errands.
User: “I listened to a new, um, Album. Today based on a documentary I watched and. The. Music was quite good, so I’m always out looking for. New. Music and. I do not shut down any music because I think all music is great.”
Albert: That’s a wonderful approach, [name redacted]. There’s so much to discover when you’re open to all kinds of music. Do you remember the name of the album or the documentary? What style of music was it?
User: Well, the name of the artist is Linda Perry, um. She. Was a lead singer for Four Nonblondes, so it was her documentary, and she’s just such a talented woman, and. Um, had. Such an interesting— And somewhat tortured life, but to see that the fact that she’s landed on both feet, um. Was. Really, really nice to see.
As a companion, Albert asked people about their day, how they were feeling, and what they were up to. He also talked with them about their past and came up with personalized stories, memory games, and riddles to keep them entertained and engaged. Research has shown that frequent conversations with other people reduce cognitive decline [Dodge et al. (2024)]. Whether the same holds for AI companions has not been scientifically proven, but we see promising signals in Asian markets like Naver launching a call service for seniors.
As an assistant, Albert could check the weather and suggest what to wear, search the internet, and look up local events to encourage people to go out and meet friends and family. In the future, we planned to give Albert more abilities and make him the universal interface between seniors and all kinds of digital and physical services.
We designed every process and every message around seniors’ needs: the speaking style, the choice of words, and a low-friction sign-up embedded in the conversation itself.
User: Also wir kaufen Brot meistens vom— Meistens vom Coop, äh, die haben dort. Eine Abteilung für Bäckerei. Und die haben gutes Brot.
Albert: Welches Brot habt ihr heute ausgewählt?
User: Ähm, es. War ein— also, wir hatten noch— Von. Gestern, äh. Einen Krustenkranz übrig.
We also let Albert handle recurring tasks on his own. For example, he onboarded new users end to end and offered to schedule the next call. For ourselves, we built a back office to review past and scheduled calls, check new sign-ups, and place calls with individual prompts.
Albert: Would you like to tell me the date and time of your next appointment? So I can remind you? Just let me know.
User: Ja. I have. I have an appointment on the 9th. Of November. For. An. Injection in my back. At 3:15.
Albert: One moment, I’ll put that in the calendar.
Albert: I’ve scheduled a reminder for your appointment on November 9 at 3:15 in the afternoon.
On the technical side, Albert was a voice agent built on vapi.ai and GPT-Live, with a custom backend, a database, and memory. To keep this post at a reasonable length, we won’t go into detail here. If you’re interested in the technical side, check out our GitHub. It contains everything you need to build your own version of Albert. If you get stuck, feel free to reach out.
In short, Albert was a digital companion that entertained people through the simplest interface there is, the phone, and that would eventually run errands for them too. Of course, our prototype had a few glitches (for some reason, Albert really liked calling people “Susie”), but overall, it worked well and did its job. So why did we decide to bury him? It turned out that a digital companion is more of a nice-to-have, and that the people it’s built for are extremely hard to reach.
Here’s what we learned after launching Albert.
Learnings
We reached out to nearly 1,000 institutions and people in our network and ended up onboarding 13 users, who made 172 calls over four weeks. Their responses and our conversations with them showed us where we had been wrong.
Product learnings
Interruptions and barge-ins: the key problem
This was our most important product learning. We first built Albert as a cascaded pipeline: a chain of models for voice activity detection (VAD), which decides when to speak; speech-to-text (STT); reasoning (LLM); and text-to-speech (TTS).
The core problem was interruption handling in the first step, VAD. Albert couldn’t tell a backchannel (“Yes, exactly”) from an attempt to take the floor (“Can you repeat that?”). At any sound, he simply stopped talking and listened, which made conversations feel extremely unnatural.
That’s why we quickly switched to a full-duplex model (GPT-Live). Full-duplex means the model can listen and speak at the same time. It decides for itself how long to wait during pauses and whether the user is trying to interrupt or just making a backchannel. The result is a much more natural conversation. This comes at the cost of losing control over the individual components, but that’s a topic for another post.
Latency: no worries
Interestingly, the problem everyone warned us about turned out to be a non-issue. Whenever people talk about the challenges of voice agents, one topic comes up again and again: latency. The common wisdom is that voice agents are too slow and need to respond faster. It’s true that nobody wants to wait three seconds for an answer, but today’s systems are already fast enough.
Localization: much more than language
We went as far as building a separate agent for each language. At first glance, this might seem unnecessary: today’s models are multilingual, so a single agent could easily switch between more than 100 languages. But that view is too narrow, because many other things change with the language: style and formality (speaking speed, directness, “du” vs. “Sie” in German vs. the universal “you” in English), units (Celsius and kilometers vs. Fahrenheit and miles), emergency numbers (112 vs. 911), and much more. A single agent gets you about 80% of the way. If you want the most natural interaction, you can’t ignore these details.
Communicating capabilities
Another problem was that people didn’t know what Albert could do. A phone companion has no screen to show options and commands, and you don’t want the agent to read out half a dozen suggestions before you learn that you can ask about the weather. Instead, we had Albert proactively suggest at most two activities, based on the context and past conversations (e.g., “You liked the towel riddle last time. Would you like another riddle, or would you rather just talk?”). This made conversations flow much better.
Go-to-market learnings
Go-to-market was extremely hard. Many seniors are interested in technology, but their trust is low, both because of bad experiences with scams and because of a general fear of AI.
Outreach channels
We emailed 888 institutions and made 59 cold calls. Only three of these led to a real conversation, and none of them closed.
We also spent about $160 on Facebook ads, which reached 5,000 people and led to 16 inbound calls. When we called them back, it turned out that they either didn’t remember calling or spoke only Spanish.
It’s a vicious cycle. The market is mostly offline, so email and online ads reach neither the gatekeepers nor the seniors themselves. You have to use the phone, but people hang up because they assume your call is spam.
Our main takeaway: this group is extremely hard to reach, and the only viable path besides word of mouth is to convince the gatekeepers: families, communities, doctors, insurers, and other institutions.
Why we decided to bury Albert
The tech is ready, but the market and the gatekeepers are not.
We still believe in the benefits of a companion like Albert, but today the need for one isn’t big enough, and fear is still a major blocker. Many gatekeepers are afraid of AI and would rather see seniors connect with other people. After contacting nearly 1,000 institutions and individuals, we didn’t see a viable path to our first 10,000 users.
So we stopped onboarding new users and personally talked to each of our active users before shutting Albert down. Since the experiment was short (only five weeks), no one had become deeply attached, and everyone took the news well.
For builders in the space
If you want to work on a similar project, please do. Loneliness and its health effects are important and growing problems, and readiness for helpers like Albert will likely increase over time.
Here are our recommendations for building in this space:
- Choose your words carefully. Avoid terms like “AI agent”. The term triggers fear and makes people stop listening. We said “digital companion” instead. Say “seniors,” not “the elderly.” Use empowering language that makes people feel valued: instead of “[the companion] will help you feel less lonely,” say “[the companion] wants to learn about the world and your life story.”
- Never assume all seniors are the same. They are a very diverse group, with different needs and attitudes toward technology. Narrow down your target group by digital affinity and by physical and cognitive abilities, and start there.
- Always let seniors decide. Family members and institutions are the best way in, but they should only make the introduction. Seniors need to discover the product for themselves: most won’t use something just because someone told them to.
Conclusion
We built Albert, a digital companion for seniors, and launched him in the US, Canada, and Switzerland. We found that such a companion is more of a nice-to-have, and we struggled to convince gatekeepers like institutions and family members to recommend him. So after five weeks, we shut the experiment down.
If you want to build your own version of Albert, go for it. All our code and documentation are open source on GitHub: https://github.com/mcbuehler/albert
Thank you for reading!
Acknowledgements
- Everyone who used and tested Albert.
- ETH Entrepreneur Club for their support.
- Anthropic and OpenAI for providing credits.
- Suno for generating Albert’s funeral song (used in the video).