Loneliness doesn't announce itself the way most problems do. There's no single trigger, no clean before-and-after. It creeps in through a retirement, a widowhood, a flight of stairs that's suddenly too much to manage, a phone call that never quite happens because everyone is busy. And it's not a fringe issue: millions of people in Germany report feeling lonely, with older adults among the groups hit hardest, and the effects reach far beyond mood. Chronic loneliness is linked to depression, weakened immunity, disrupted sleep, and elevated risk for heart attack, stroke, and dementia.
That's the backdrop a student team set against a much narrower, sharper question: could a product built on ChatGPT actually help? Not solve loneliness outright, nobody involved was under that illusion, but meaningfully soften it for older adults specifically. The resulting project, and the product it produced, AchtSAM, is a useful case study not just in what AI companionship might offer, but in where a team drew a hard ethical line partway through.
Starting With Empathy, Not a Feature List
The team followed the Design Thinking process end to end: Understand, Observe, Synthesize, Ideate, Prototype, Iterate. What's notable is how much of the early effort went into resisting the pull toward a features list before actually understanding the people involved.
The Understand phase started with a brainstorm surfacing real unknowns rather than assumptions: how replaceable is human contact by something like ChatGPT, actually? What do older adults consider a meaningful interaction? Why do some people seem more resilient to loneliness than others? From there, the team built an empathy map, a simple but effective tool organized around what a target user sees, hears, thinks and feels, says and does, and what they struggle with and hope for. For an older, isolated person, that map filled up with things like worry about family health, quiet hope for more contact with grandchildren, frustration at being overlooked, and a wish to still feel useful.
Recruiting actual lonely older adults for in-depth interviews turned out to be genuinely hard, both ethically and logistically, on a tight academic timeline. Rather than force it, the team pivoted to secondary research: expert interviews, published studies, podcasts, and journalism already covering the topic, organized around a few guiding questions. What causes loneliness in old age? What does an ideal social exchange actually look like? What already exists to help?
What the Research Actually Indicates
A few findings stood out clearly enough to shape the whole direction of the project. Loneliness in old age isn't simply about contact frequency. Someone can live with family and still feel isolated; what seems to matter more is finding peers, people at a similar life stage who can meet them as equals, not just company in general. Sudden illness was flagged as a particularly sharp trigger, since it restricts mobility for both the person affected and the friends who might otherwise visit them, and there often isn't an adequate substitute once that social contact drops away.
Life transitions carry real weight too: retirement removes structure and a built-in social network overnight, and widowhood often comes with a second, quieter loss, older adults reported being gradually excluded from social invitations once they were "the one without a partner." Financial strain matters, particularly in the context of old-age poverty, and there's a documented regional gap between eastern and western Germany. And shame plays a bigger role than you might expect: many older adults connect their own loneliness to a feeling that it's somehow their own fault, which makes reaching out for the very programs designed to help them a genuine act of courage rather than a simple logistical step.
What people actually wanted, when asked, was less abstract than "more connection." It was peer-level conversation, a sense of being needed rather than looked after, and the chance to pass on experience and values to someone younger. Purposeful activity mattered more than passive company. And several sources specifically noted that emotional closeness in a handful of relationships mattered more to older adults than a wide but shallow network.
Turning Research Into People: Three Personas
To keep the target group from staying an abstraction, the team built three personas, deliberately varied rather than a single stereotype. Margarete, 78, extroverted and warm, a former cook whose husband recently passed; her adult children introduce her to the product hoping she'll have someone to talk to during the hours nobody has time to call. Klaus, 81, the opposite temperament entirely: guarded, technically minded, a retired mechanical engineer who spent decades at Daimler and built almost no life outside work; a caregiver at his retirement home suggests the product specifically because he'd rather talk to something technical than to a person. And Gaby, 79, caring for a husband with dementia, worn down less by isolation itself than by the emotional toll of caregiving, looking for a low-effort way to switch off for a few minutes a day.
Three different reasons to want the same kind of product: company, a technically interesting distraction, and respite. That range turned out to matter later, because the same feature landed very differently depending on which persona (or real interviewee) was picturing it.
Sizing Up the Competition
Before settling on a direction, the team looked at what already existed. Dinsaw mini AI, a Thai product aimed at elderly users, offers facial recognition, video calls, and health reminders, but is language-locked to Thai and has no public pricing. Catalia Health's Mabu robot positions itself in therapeutic support and patient engagement, holding up to thirty-minute conversations while feeding data back to healthcare providers. ElliQ, probably the closest direct comparison, has built a genuinely loyal following, users report developing real attachment to it, and covers daily check-ins, mood tracking, and family connection, priced around $30 to $40 a month plus hardware, but only ships within the US. Replika and Character.AI sit further from the healthcare space entirely, positioned as general-purpose AI companionship and roleplay apps rather than tools built specifically around aging or isolation.
The gap the team identified: nothing served the German-speaking market directly, nothing combined a genuinely therapeutic posture with the flexibility to slot into whatever device an older user already owned, and nothing treated the family and caregiver network around the older adult as part of the product rather than an afterthought.
Building AchtSAM
The resulting concept, AchtSAM, whose name plays on the German word for mindfulness (Achtsamkeit) while the capitalized "SAM" doubles as the AI companion's own name, an acronym for sensitive, attentive, and empathetic, positions itself as a voice-first conversational companion built on ChatGPT, with speech recognition and text-to-speech handling the audio layer and a persistent, personalized memory of each user maintained in the backend. Rather than a single fixed device, it's designed to plug into voice assistants people already own, like Amazon Alexa, or run through a simple web chat interface for users with hearing difficulties or a preference for text.
A companion notification layer lets family members or caregivers opt into updates, via email, RSS, or webhook, about how the person they care about is doing, without requiring anyone to install a new app themselves. The explicit design principle throughout: AchtSAM doesn't try to replace a person's existing relationships. It's built to strengthen them, and to gently nudge someone toward new ones, a knitting circle, a volunteering opportunity, when the AI notices a gap.
The Prototype, and an Uncomfortable Fork in the Road
Building a believable prototype for a voice-only product turned out to be its own design problem: there's no screen to mock up. The team's solution was to generate sample conversations directly through ChatGPT, voiced using text-to-speech and, in early exploration, voice cloning technology, paired with AI-generated illustrations from Midjourney to help test participants picture the scenario before hearing it.
Buried in that early prototyping work is the most genuinely difficult decision in the whole project. One early idea, explored seriously enough to appear in test dialogues, was letting the AI take on the voice and persona of a user's deceased spouse, using their name and speech patterns to ease loneliness after loss. It's an idea with obvious emotional pull, and an equally obvious capacity for harm: an AI standing in for someone's dead husband or wife risks entrenching grief rather than helping process it, and raises real questions about consent, dependency, and what a healthy mourning process actually needs.
To the team's credit, this is exactly the fork where they stopped and turned back. The product's own conclusion states plainly that this kind of interaction could have harmful effects, and the final vision for AchtSAM explicitly excludes it, keeping the AI's role limited to supportive listening, encouragement toward real-world reconnection, and, in acute situations like anxiety or crisis, gently guiding someone toward professional help rather than trying to substitute for it. It's a good example of a team following an idea far enough to see its danger clearly, and having the discipline to cut it rather than ship it because it tested emotionally powerful.
What Real Users Actually Said
With the ethically safer version of the prototype in hand, the team ran seven solution interviews, playing back recorded sample dialogues and gauging reactions. The split in responses was sharp, and revealing.
Several participants, particularly those already comfortable with technology, responded well: a 72-year-old widower called the idea good and said he'd try it, reasoning that "current alternatives, like pets, can't actually talk back." A 72-year-old living alone said he could imagine talking to an AI and was open to the technology, though he flagged concerns about losing real human contact, something he chalked up partly to generational attitude. An 86-year-old widow said she could genuinely see herself using it if her circle of people to talk to ever shrank further, and specifically liked the idea of casual small talk, like a weather check-in.
Others pushed back hard. An 80-year-old found the sample dialogue too on-the-nose and said he probably wouldn't try it. An 83-year-old living alone said she found the whole idea of talking to a robot off-putting and simply couldn't relate to it. A 79-year-old flatly rejected the premise on principle, arguing that loneliness is a matter of personal responsibility and that the burden belongs with younger family, not technology. And one 78-year-old, married and living with her husband, put her finger on something the team hadn't fully anticipated: she said the dialogue simply didn't feel real to her, she couldn't get past knowing it was AI, even while agreeing the underlying idea, an always-available supplement to human caregivers, had genuine merit.
Across all seven interviews, a consistent thread held regardless of enthusiasm level: nobody thought the AI should or could replace human interaction. The most positive respondents saw it purely as a supplement, something available at 3am when no caregiver is on shift, not a substitute for the caregiver.
Turning Feedback Into Priorities
Synthesizing across the exploratory research and the solution interviews, three needs kept surfacing above everything else: social interaction (acknowledging up front that technology can't fully replace this), distraction and entertainment that shifts focus away from negative thoughts, and access to something resembling advice or support during difficult moments. The corresponding product hypothesis followed directly: an AI-driven product offering social contact and entertainment for older or isolated people could meaningfully improve wellbeing and quality of life, without ever claiming to fully replace human connection, particularly valuable specifically when human contact is limited or unavailable.
The clearest headwind, unsurprisingly, was trust and comprehension of the technology itself, especially among the least tech-familiar respondents, alongside the genuinely hard technical challenge of making an AI companion feel convincingly natural rather than obviously scripted.
Does the Business Case Hold Up?
On the numbers side, the team estimated roughly 221 person-days of development effort, landing an MVP cost around 82,000 euros once staffing rates and a contingency buffer were applied. Running costs per user came out modest, driven mainly by the ChatGPT API (a few cents to under half a euro a month depending on GPT-3 versus GPT-4 usage), text-to-speech, speech-to-text, and hosting, adding up to roughly 10 euros per active user per month. The proposed pricing was a three-month free trial followed by 19.99 euros monthly, deliberately structured as a flat subscription rather than usage-based billing, specifically so the price itself wouldn't discourage someone from reaching out to talk whenever they actually needed to.
Modeled out across five years with steadily growing user numbers, the projected break-even point lands at roughly three years and eight months, a plausible timeline for this kind of subscription SaaS product rather than an overnight hit, and honest about the significant losses expected in the first two years before the user base and recurring revenue catch up.
The Real Lesson Here
What makes this project worth paying attention to isn't really the specific product concept, voice-first AI companions for older adults are a crowded and rapidly evolving space, and AchtSAM is one modest student take on it. It's the process, and specifically the willingness to let genuine user feedback complicate a clean narrative. The team didn't get uniform enthusiasm, they got a real split between people who saw value in an always-available conversational partner and people who found the entire premise unconvincing or even a little sad. And when an idea (AI as a stand-in for someone's late spouse) turned out to be technically feasible and emotionally compelling in a demo, the team chose not to build it anyway, precisely because feasible and compelling aren't the same thing as responsible.
That combination, taking a genuinely hard social problem seriously, building something concrete enough to actually test, and being willing to walk back a feature once its risks became clear, is a better template for AI product design than most far larger, far better-funded companies manage to follow.