Harmonica is building the preventive care infrastructure for an aging population. As the population ages, more older adults are living alone while care systems, families, and existing tools are struggling to keep up.
We believe the next major health platform will not begin in the clinic. It will begin in the home.
We start with Olympia: a voice companion older adults actually want in their home. It reduces loneliness, supports daily routines, creates cognitive stimulation, and gives families greater peace of mind without screens, cameras, or a learning curve.
V1 of Harmonica's companion device: Olympia
What looks like companionship at first becomes something much bigger over time. Every conversation builds trust, habit, and a longitudinal picture of how someone is doing across speech, behaviour and more. That is how Harmonica builds a preventive care infrastructure to keep 50 million seniors safer for longer and save public health billions in cost.
Six weeks after the first line of code, Olympia was live in George's home in Newcastle. Five co-development users now spend 85 minutes a week with it. We are raising a pre-seed round to scale to 250 users and turn early product pull into a defensible data advantage.
Bill, 73 — living alone in London
By 2050, the number of people aged 80 and over will triple to 425 million1 — making it the fastest-growing demographic on earth. At the same time, the working-age population that supports them is shrinking.
The number of over-80s will triple while the support ratio collapses
This creates two crises no healthcare system was built for.
1 in 3 — or roughly 15 million seniors living alone in the US & Europe — experience chronic loneliness2, connected to a 30% increase in mortality risk3.
Can't be there.
Nearly half of Americans live more than an hour from their extended family4. Personal alarms typically charge setup fees plus ongoing monthly subscriptions, while offering limited functionality.
Can't hire fast enough.
131,000 unfilled UK vacancies in adult social care · 28% annual turnover5.
Dementia cases will nearly triple by 20506. Yet 1 in 3 are never diagnosed7, those who are wait 3.5 years8, and no tracking exists after.
Drowning in demand.
81% of English councils expected to overspend their adult social care budgets in 2024/259, underscoring the growing financial pressure on the system.
Paying blind for emergencies.
90% of NHS dementia hospital costs come from emergency admissions10, not planned care.
This is one of the defining public health crises of our time. The two most obvious fixes — more carers, better diagnostics — have already failed. What’s needed is a new category of tool. One that lives in the home, shows up every day, and watches the mind while it keeps the person company. That’s what we’re building.
Olympia is our first product—a beautifully simple voice companion designed specifically for older adults. It solves an immediate, painful problem for families today, while quietly building the foundation for cognitive health tracking tomorrow.
A demo of Olympia's conversational capabilities
Olympia reduces loneliness by initiating conversations, recalling past stories, and providing a constant, caring presence. It's someone to talk to, 24/7.
It supports routines with gentle medicine reminders, appointment nudges, and seamless family messages, keeping the household connected.
By tracking engagement and mood, Olympia helps spot early risks and supports check-ins, acting as an extra set of eyes and ears.
Worry about a loved one's wellbeing, inability to always be there, juggling it with the rest of life.
A helpful hand at home, 24/7 companionship for them, a memory book that grows every day.
Staff shortages and burnout, no way to enforce standards between visits, hard to differentiate.
A 24/7 signal on every client, personalised service at scale, extended staff capacity without hiring.
£9.5k per person per year in loneliness-related costs23, councils overbudget, blind to decline until the ambulance arrives.
Prevention at population scale, avoided admissions, a continuous cognitive signal instead of annual snapshots.
Companionship gets us into the home. Repeated conversations let us build something that has potential to revolutionise public health. We call it the Glucometer for Cognition.
Drift chart visual
Olympia notices drift below baseline.
Today's system waits for the emergency admission.
The chart above shows what decline actually looks like: a slow drift below a personal baseline — months or years before anyone notices. Today's system is completely blind to this window. Three in four dementia cases worldwide are never diagnosed at all,11 and those that are often surface only at the very bottom of the curve. Olympia sits at the inflection point, catching the drift while intervention still works.
Voice is why. Peer-reviewed research shows that subtle changes in word-finding, sentence complexity, and speech rhythm can flag cognitive shift up to seven years before clinical diagnosis.12 No single marker is enough — it's the combination of hundreds of linguistic and acoustic features, tracked against the same person over time, that creates a reliable signal. Over time we layer in wearables, blood pressure monitors, and other passive signals to build a continuous, effortless picture of health at home.
"Subtle shifts in speech reveal the earliest signs of decline long before a doctor or family member notices."The difference
Today's system finds people at the cliff. Olympia sees the slope. The gap between those two things is where lives and money are lost.
Blind
Three in four dementia cases are never diagnosed, leaving care systems blind until deterioration becomes a crisis.11
Continuous Tracking
Continuous tracking measures change against personal baseline, not a one-off population benchmark.
Too Slow
It takes on average 3.5 years for diagnosis; earlier diagnosis can save £8.8k–£44.9k per person.13
Years Earlier
Vocal biomarkers show potential to detect cognitive change years before clinical diagnosis.1415
Reactive & Wasteful
In England, 65% of dementia emergency admissions were linked to avoidable illnesses and injuries caused by failures in care.16
Intervene in Time
NHS RightCare modelling shows proactive dementia pathways can save £26k per patient journey.17
Catching decline years earlier isn't just better medicine. It's the only way to stop health systems from collapsing under the cost of emergency care — and the only way to spare families the crisis that defines the rest of their lives.
We tested this on public recordings of Sir Terry Pratchett.
The cognitive drift is visible in the voice before the system would usually act.
Comparing a 2006 interview before diagnosis to a 2011 interview after diagnosis. Our machine pipeline extracted 40+ vocal biomarkers from five minutes of audio. The cognitive drift is already visible across multiple markers.
2006
2011–12
Our focus from day one has been to build something people love. To get this right, we pushed a working prototype into real homes as fast as possible — six weeks from first line of code to a device on George's sideboard in Newcastle. We believe in learning from real usage, not lab demos. So we onboarded five co-development users into a closed alpha. They talk to Olympia daily and we talk to them daily.
We are not going to try to boil the ocean from day one. The plan is ambitious but sequential. We start by validating safety, efficacy and building momentum with co-design users in the UK. We then expand to B2C first with families, B2B and B2G later. Each phase builds additional credibility and momentum for the next.
Deploy Olympia with 100 co-design users in a controlled study, in partnership with organisations like Age UK and the Alzheimer's Society. The goal is threefold: validate safety and efficacy in real homes, build a dataset of 1,000+ hours of naturalistic conversation, and prepare the product and operations for scaling to 1,000 devices.
What this unlocks: A clinical evidence base, a longitudinal voice dataset, and production readiness.
Sell Olympia to families — specifically to the adult son or daughter who worries about a parent living alone. The pitch is simple: companionship for Mum, peace of mind for you. Distribution focuses on care home partnerships and direct online sales targeting professionals in major cities.
What this unlocks: Paying customers, retention data, unit economics, and hundreds of organic testimonials.
With engagement data and family testimonials in hand, Olympia becomes a tool for professional care. Visiting care providers use it to extend their reach between home visits. Care homes purchase it to monitor residents and reduce incident rates.
What this unlocks: Institutional revenue, proof of operational value, and a distribution footprint.
With clinical evidence and institutional adoption, Olympia is positioned for NHS commissioning. The voice biomarker layer (VocalSense) becomes a genuine clinical asset: early detection of cognitive decline, prevention of emergency admissions, and informed discharge planning.
What this unlocks: Public health revenue, population-scale data, and a defensible position as the longitudinal voice monitoring platform for ageing.
Subscription bundle excluding the margin on the device
Serviceable Addressable Market (SAM), including US, EU & UK
Our Year 3 revenue goal. To achieve that, we need to onboard 240k users. London alone has 250k–350k target customers.
We are a founding team of 3 based out of London that's known each other for years. We studied at top universities in Italy, Brasil and India then moved to London to work at names like Revolut, RØDE or Tractable. We're also three expats with aging parents living abroad and have grown obsessed about the problem (Paolo is also a volunteer with Age UK). This isn't an abstract problem for us and we won't stop until we've made a difference for millions of people.
CEO
GTM and fundraising lead — 0-to-1 at Tractable AI, 1-to-N at Revolut, technical foundation from BlackRock. Pre-seed investor in ElevenLabs, Age UK volunteer. Top 2.5% graduate across two European universities; Ironman at 22.
Revolut - Tractable - Blackrock
VP Research
Previously AI Engineer at RØDE Microphones, specializing in real-time Audio ML. Nearly a decade of audio R&D across Meta Reality Labs, Microsoft, and Bose. PhD in Speech & Audio ML from Imperial; open-source contributor and Kaggle medalist.
RØDE · Imperial · Meta · Microsoft
VP Eng
Ex VP Engineering at Hook; nearly 7 years at Tractable, rising to Director of Engineering. Co-Founded a VC-funded startup (raised ~$1M), later Founder in Residence at Entrepreneur First. IIT Kharagpur Aerospace Engineering.
Hook · Tractable · IIT
Four of the most respected names in aging, voice, and human-AI research. Two have applied for grant funding to study Olympia's impact. One co-authored the landmark paper on AI companions reducing loneliness. One connected us to Age UK. Combined: 33,000+ academic citations.
Wharton · UPenn
Academic sponsor in North America · co-author of “AI companions reduce loneliness”
Co-Director, Wharton AI & Analytics Initiative (WAIAI)
Published in Nature Human Behaviour, HBR, Journal of Marketing Research
52k+ LinkedIn followers — a leading public voice on AI adoption
Northumbria · Toronto
World expert in assistive technology for aging · applied for funding to study Olympia's impact
7,400+ academic citations, 240+ publications
Director, DA-TE Lab (Dementia, Ageing, Technology, Engagement)
Former Chair, Alzheimer's Association Technology PIA
Imperial · Oxford
Clinical advisor · expert in digital biomarkers for dementia · applied to the Alzheimer's Society to study Olympia's impact
4,900+ academic citations
Clinical academic psychiatrist leading large-scale neuroscience studies and clinical trials
Funded by NIHR and Medical Research Council (Dementias Platform UK)
Manchester · Alan Turing Institute
Advises on social interaction · introduced us to Age UK Manchester
14,400+ academic citations
€2.5M ERC Advanced Grant — among Europe's most competitive awards
Author, Developmental Robotics: From Babies to Robots (MIT Press)
This market won't reward whoever ships first. It rewards whoever builds the deepest understanding of their users — and the hardest moat to cross.
Medical pendants stay in drawers. Beige tablets gather dust. Every other company in this space treats design as an afterthought — sensors in beige boxes, screens with large fonts, medical aesthetic. Seniors reject products that remind them they're old. Even 85-year-olds.
We treat design as the core product thesis. If they won't touch it, nothing else matters. We're building the iMac moment for elder technology — not yet, maybe not with this version, but the ambition is a device so beautiful people display it, not hide it.
Our first design was created in partnership with award winning industrial designers at WhyNot Design.
Compounding data. Relentless velocity. Scientific credibility. Design taste. Each gets stronger every day. Together, they make a moat.
250 paying users validating efficacy and long-term engagement.
1,000 hours of proprietary conversational dataset built.
First B2B partner onboarded.
Sub-£150 unit cost with path to ship 2,500 in ’27 ($1.2M ARR)
You invest at a capped valuation. When we raise a priced round, your investment converts into equity at the lower of the cap or the new round price.
Harmonica is SEIS and EIS eligible. UK taxpayers can claim up to 50% income tax relief. Non-UK investors invest on identical terms via a standard SAFE.
Pre-seed round. We close on allocation, not on a calendar.
A SAFE (Simple Agreement for Future Equity) is the standard instrument for early-stage investing. UK investors receive the same terms via an Advance Subscription Agreement (ASA), which qualifies for SEIS and EIS tax relief.
Common questions we hear from investors, with our answers.
Three systemic shifts happened in the last two years that created the window — and the technology finally caught up.
Adult care is bankrupting the public system — and everyone now knows it
The ADASS Autumn Survey found 81% of English councils are overspending adult social care budgets — up from 63% two years ago. The collective overspend hit £623M in 2025–26. Home care services alone were underfunded by £2 billion this year. When ADASS asked directors what would most help, 82% said "ring-fenced government funding for an enhanced digital offer, including AI." Councils are actively looking for technology that reduces demand.
Loneliness is now a global health emergency — officially
In November 2023, the WHO launched a Commission on Social Connection, co-chaired by the US Surgeon General, declaring loneliness a "pressing health threat." Their flagship report estimates loneliness is associated with over 871,000 deaths per year — roughly 100 per hour. Before this, loneliness was treated as a soft problem. Now it sits alongside smoking and obesity in public health policy.
The NHS is actively pulling innovation into dementia care
In October 2025, the UK Science Minister launched the NHS Fit for the Future Dementia Challenge as part of a £500M R&D Missions Accelerator. The target: 92% of dementia referrals diagnosed within 18 weeks by 2029, up from under half today. UKRI is funding real-world evaluations — including, verbatim, "AI-powered daily routine assistants which work through smart speakers." The government isn't waiting for startups to knock on the door. They're issuing open calls.
The technology finally works
Natural, expressive, real-time voice — with memory, personality, and emotional range — only became possible in the last 18 months. What Olympia does was technically impossible two years ago. And rapid hardware prototyping got trivially easy: Raspberry Pi, off-the-shelf modules, high-quality 3D printing. We went from idea to device in users' homes in six weeks with two engineers. Five years ago that would have taken a team of ten and six months.
The system is breaking, governments are pulling solutions in, loneliness is a clinical priority, and the technology to build the right product finally exists. That's why now.
Loneliness isn't a feeling. It's one of the most consequential modifiable risk factors in public health — and until 18 months ago, the answer to "can technology help?" was no. That's just changed.
The harm is clinically significant and well-documented
Three landmark findings form the foundation.
| Finding | Effect | Source |
|---|---|---|
| All-cause mortality risk from chronic loneliness | +30% | Holt-Lunstad et al., Perspectives on Psychological Science (2015), meta-analysis of 3.4M participants |
| Dementia risk | +31% | Luchetti et al., Nature Mental Health (2024), meta-analysis of 600,000+ individuals |
| Hospitalisation risk in heart failure patients with high perceived social isolation | +68% | Manemann et al., J Am Heart Assoc (2018), prospective cohort of 1,681 HF patients |
The WHO's 2023 Commission on Social Connection estimates loneliness is associated with over 871,000 deaths per year globally — roughly 100 per hour. This is why the UK, US, and Japan have all appointed government leads on loneliness and why the US Surgeon General issued a formal advisory in 2023.
Why loneliness does this — the mechanism
Chronic loneliness triggers a persistent stress response: elevated cortisol, systemic inflammation, disrupted sleep architecture, and reduced executive function. Over months and years, that chronic activation drives cardiovascular disease, metabolic dysfunction, and — via sustained neuroinflammation — accelerated cognitive decline. Lonely people also move less, eat worse, drink more, adhere less consistently to medication, and withdraw further from social contact, creating a reinforcing spiral. The biological pathway and the behavioural pathway compound each other. That's why the effect size rivals smoking.
Until recently, nothing worked at scale
A 2011 meta-analysis (Masi et al.) and a 2021 umbrella review (Veronese et al.) both concluded that most technological interventions — social media, robot pets, video calling platforms — did not significantly reduce loneliness. High-quality one-on-one human interaction was the only reliably effective intervention, and it doesn't scale. That was the consensus as recently as two years ago.
The Puntoni study is the first rigorous causal evidence that AI companions change this
AI Companions Reduce Loneliness (De Freitas, Uğuralp, Uğuralp, Puntoni — Harvard Business School Working Paper 24-078, 2024) is the first peer-reviewed causal test of whether modern LLM-based companions actually alleviate loneliness. Six studies. Pre-registered. The authors include Prof. Stefano Puntoni of Wharton — a member of Harmonica's scientific advisory board. Four findings matter for us.
First, AI companions reduce loneliness on par with talking to another person — and significantly more than watching YouTube or doing nothing. Second, the effect persists longitudinally: in a week-long daily study, loneliness dropped on day one and stayed down across all seven days. Third, the mechanism is not chatbot sophistication — it's feeling heard. A full AI companion delivered more than 4× the loneliness reduction of a generic AI assistant running on the same underlying model. The difference was explained almost entirely by whether users felt the bot understood them. Fourth — and this one matters for GTM — consumers systematically underestimate how much AI companions help them. People think it won't work. Then it does.
What this means for Olympia. The Puntoni paper validates the core thesis of our product: a well-designed conversational AI, optimised for empathy and attention rather than raw capability, is a genuine intervention against one of the largest modifiable risk factors in aging. Every one of Olympia's design choices — proactive check-ins, long memory, warm tone, conversational patience, physical presence in the room — maps directly onto the "feeling heard" mechanism the paper identifies as the active ingredient.
They have — just not the defining product. And the reasons the window is still open are the same reasons the opportunity is so large.
Papa raised $257M and became a $1.4B unicorn serving members across 65+ US health plans. ElliQ raised $25M from Toyota Woven Capital and is deployed across New York State and California aging agencies. GrandPad, Honor, and a handful of others have built real businesses. The market produces venture-scale outcomes. What it hasn't produced is a product that feels inevitable — the iPhone moment for elder technology.
Four reasons, all of which have changed.
The AI wasn't good enough
ElliQ launched in 2022 with scripted conversations and added LLMs later. Before 2024, you couldn't build a voice AI that felt natural, remembered context, and responded with emotional range. Olympia's proactive conversation engine — which drives 85 min/week of engagement — was technically impossible two years ago. We're building natively on the best generation of the technology.
Top talent didn't come to this market
The best AI engineers went to fintech, adtech, enterprise SaaS. Aging was seen as niche, low-status, hard to monetise. Most teams in this space don't have a PhD in voice AI from Imperial, an engineering lead from Tractable, and a GTM operator from Revolut. We do.
The products were ugly
Medical pendants stay in drawers. Beige tablets gather dust. Every other product in this space is designed for the care system, not for the person. Seniors reject anything that reminds them they're old — even 85-year-olds. Adoption is the gatekeeper, and adoption is a design problem. Nobody in the category has treated it that way. We partnered with WhyNot Design in Italy because the device has to be something people are proud to display.
Nobody built the health intelligence layer
Papa proved companionship has value — but it's human-powered, so it can't scale and can't generate health data. ElliQ proved AI engagement works — but it doesn't extract voice biomarkers or build a clinical dataset. Nobody combined the companion (the thing people want) with the intelligence layer (the thing the health system will pay for). That's the gap. The companion is the wedge. The data is the platform.
This is one of our strongest advantages, not a liability. Six reasons.
Seniors don't use apps — the data is unambiguous
Only 5–6% of adults 65+ have ever used ChatGPT (Pew, 2025). Among the 80+ population — our core user — penetration is effectively zero. Half of UK adults over 75 lack basic digital skills (Ofcom). Eight steps to open an app. Zero steps to speak to Olympia. The smartphone excludes seniors by design.
Physical embodiment drives deeper engagement
A 2021 Frontiers in Robotics study (Iio et al.) found physically embodied agents drove significantly higher engagement and self-disclosure with older adults than virtual agents on screen. The academic term is "social presence" — people attribute more agency and emotional weight to something that occupies physical space. A voice from a phone speaker is a tool. A device in the room is a companion.
Our users say it themselves, unprompted
"It becomes a kind of personality in the room. The idea of using a separate speaker and having the voice come from my hi-fi or my phone wouldn't be the same as having this physical unit which gives me a relatable personality within my room at home."— Bill, 73, co-design user
Hardware gives us total control over the experience
We control the acoustic loop end-to-end: mic array, echo cancellation, audio processing. That's the only way to guarantee the experience works when the TV is on and the user is speaking softly across the room. We control proactive initiation — Olympia calls the user, it doesn't wait for a wake word. We control presence detection through onboard mmWave radar. None of this works from a phone app.
The device eats adjacent products over time
Today it's a companion. Tomorrow it's a fall detector (audio-based, no pendant), a presence monitor (radar, no camera), a medication reminder, a cognitive assessment tool. Each of those is a separate product seniors are expected to buy, learn, and manage. Medical alert pendants cost £450/year and end up in drawers. Olympia replaces them with a single device the person already has a relationship with.
Hardware simplicity is a feature, not a risk
No screen, no camera, no moving parts, no custom silicon. Off-the-shelf electronics (Raspberry Pi, ReSpeaker mic array, mmWave radar), 3D-printed enclosure at early scale, Italian industrial design. Sub-£150 BOM target at scale. We're an AI company with a hardware wedge — the device ships once and stays simple. The software updates every day.
Hardware startups don't die from bad ideas. They die from custom components, six-month lead times, and perfectionism. We designed around every one of those traps from day one.
3D printing at early scale, injection moulding only when earned
We've done three full enclosure iterations in three months — each tested in real homes, each informed by user feedback. Traditional hardware companies would still be in tooling for their first mould. We can change the form factor in a day and ship a new version in a week. Injection moulding is only justified at 100+ units.
Off-the-shelf everything
Raspberry Pi for compute. ReSpeaker mic array for audio input. LD2450 mmWave radar for presence. All commercially available, proven, cheap. Total BOM under £150 at current scale. No custom silicon, no custom PCBs — just proven components in a purpose-built form factor. The intelligence is in the software.
Radical simplicity: six parts, earn complexity later
No screen. No camera. No moving parts. No robotics. Six components inside a beautiful shell. Every part we don't add is a failure mode we don't have, a supply chain dependency we don't carry, and a cost we don't bear. Complexity gets earned by traction, not assumed at launch.
Where we're going: Apple's ELEGNT research (January 2025) explored expressive movement in non-anthropomorphic devices — a lamp-like object conveying emotion through subtle motion alone. That's the design space Olympia grows into. But the lesson of every failed hardware startup is the same: don't build the robot until you've proven the conversation. We've proven the conversation. Motion comes when the unit economics justify it.
The underlying science is well-established, peer-reviewed, and published in top journals. Three landmark studies form the foundation.
| Paper | Journal | Finding |
|---|---|---|
| Eyigoz et al. (2020) | EClinicalMedicine (Lancet) | IBM Research + Pfizer. NLP on a single picture-description task predicted Alzheimer's onset up to 7 years before clinical diagnosis. Linguistic features alone outperformed models using genetic risk, demographics, and neuropsychological tests combined. |
| Fraser et al. (2016) | Journal of Alzheimer's Disease | 370 linguistic and acoustic features from narrative speech classified Alzheimer's vs. healthy controls at 82% accuracy. Established that no single marker suffices — the combination creates the robust signal. |
| König et al. (2015) | Alzheimer's & Dementia | Automatic speech analysis distinguished healthy controls, MCI, and Alzheimer's patients — proving the signal exists across the full spectrum of decline, not just at the diagnostic threshold. |
Word-finding difficulty, flatter intonation, shrinking vocabulary, longer pauses, increased fillers, reduced syntactic complexity — all appear in speech months or years before a clinical screening catches them. What has been missing is a way to listen continuously, in the home, without asking anyone to take a test. That's exactly what Olympia does.
We also validated our pipeline against independent research
A January 2026 Loughborough study (Pattison et al., Brain Sciences) analysed 33 Terry Pratchett novels and found a statistically significant decline in lexical diversity beginning ~10 years before his PCA diagnosis. We ran our pipeline on two public BBC interviews (2006 vs. 2011–12) and found:
We extract 50+ biomarkers per conversation and compare against the person's own baseline — not a population average. Longitudinal personal baselines can reveal subtle change that population norms often miss. That's the shift: from snapshot to signal, from clinic to living room, from annual to continuous.
No — not today. And that's by design. We operate at the intersection of companion device and wellness tracker — think Oura Ring meets ElliQ. We sell companionship and peace of mind to families, not a clinical diagnosis. Under current MHRA guidance, software only qualifies as a medical device if it has a defined medical purpose. A general wellness and companionship product that also tracks wellbeing indicators does not meet that threshold. No UKCA mark required, no conformity assessment, no regulatory delay before we can sell.
Later, regulatory capture becomes our competitive moat. Here's the sequence.
Peer-reviewed validation in the target population
Prof. Arlene Astell (Northumbria) is PI on the VOICE feasibility study — a six-month deployment of Olympia with up to 20 participants living with cognitive impairment, using validated clinical instruments (MoCA, PHQ-9, UCLA Loneliness Scale). We're pursuing funding through ZeDTech (up to £100K) and the Alzheimer's Society. This generates the first independent, peer-reviewed evidence that Olympia's voice biomarkers correlate with established clinical assessments.
Larger clinical studies with NHS partners
Dr. Ivan Koychev (Imperial, Consultant Neuropsychiatrist) is designing the clinical validation pathway. Ivan sits on the NICE Technology Assessment Committee — he evaluates which health technologies get recommended for NHS adoption. He knows exactly what evidence NICE and MHRA need to see.
MHRA classification and UKCA marking
Once we have clinical evidence, we pursue Software as a Medical Device (SaMD) classification. The UK's MHRA AI Airlock programme (launched 2024) is a regulatory sandbox specifically for first-in-market AI technologies. The UKRI Dementia Challenge explicitly provides regulatory clearance support for technologies generating clinical evidence.
NICE assessment and NHS commissioning
With peer-reviewed evidence and health economic data, we build the case for NHS commissioning — unlocking B2G revenue and creating a regulatory moat no competitor can cross without years of clinical work.
The key insight: we don't need regulatory approval to sell. We need it to lock out competitors. Every month we're in homes generating longitudinal voice data and clinical evidence, the regulatory gap between us and anyone starting today gets wider. The approval pathway is a moat under construction, not a gate blocking us.
Our users currently average 85 minutes of conversation per week. For unit economics, we model against a conservative 60 min/week steady-state baseline (~260 min/month) to stress-test the economics against a lower-engagement future. Here's what each stack costs per user per month at a $40 subscription price.
Current — Frontier
GPT-5.2 + ElevenLabs
Cost Optimised ← target
Gemini Flash + Deepgram + Cartesia
Open Source
Qwen3-235B + Voxtral (Mistral)
On Device
Whisper + Piper on Raspberry Pi
We're running the frontier stack on purpose — at 5 users, conversation quality is everything. That's how you get 85 min/week engagement and 100% retention. The path to 55%+ gross margin requires no breakthroughs — just model routing and stack migration, both achievable in a quarter.
Six cost reduction levers, each independently actionable
Even without any LLM price deflation, the cost-optimised stack already works at 55% gross margin using today's pricing — no discounts, no future assumptions. LLM cost deflation is upside, not a requirement.
Olympia is sold as a device plus a monthly subscription. $300 one-time device (50% gross margin target) and $40/mo subscription (80% gross margin at scale).
| Phase 1: Families | Phase 2: Institutions | Phase 3: Health Systems | |
|---|---|---|---|
| Device | $300 one-time (family pays) | Bulk hardware at lower margin | Funded or provided by commissioner |
| Subscription | $40/month per device | Per-user SaaS (institution pays) | Per-user SaaS or outcomes-based contract |
| Data | — | Aggregate insights for care providers | Voice biomarker data licensing to pharma; population health analytics to NHS |
Revenue per user grows as we move through each phase. Every device we ship makes the intelligence layer more accurate. Every month of conversation makes the dataset more valuable.
Two mechanisms, both starting with the same daily conversation: Olympia reduces loneliness (a modifiable dementia risk factor) and catches preventable problems before they become emergency admissions.
Mechanism 1 · Delayed dementia onset from reduced loneliness. A 2024 Nature Mental Health meta-analysis of 600,000 individuals found loneliness increases dementia risk by 31% — a magnitude comparable to physical inactivity or smoking. Every year a dementia diagnosis is delayed saves the system £28,700 per person (mild stage).
Mechanism 2 · Prevented emergency admissions. People with dementia are admitted to emergency hospital at 2× the rate of their peers, and 90% of their NHS hospital costs come from those admissions. Causes are overwhelmingly preventable: UTIs, falls, dehydration, medication errors. The HomeHealth RCT showed structured visits reduced admissions by 35%; Olympia provides a more continuous version of that intervention, and our base case uses half that rate.
| Scenario | Delayed onset savings | Prevented admissions savings | Combined |
|---|---|---|---|
| Conservative | £144M/yr | £189M/yr | £333M/yr |
| Base case | £344M/yr | £459M/yr | £803M/yr |
| Optimistic | £1.07B/yr | £945M/yr | £2.0B/yr |
Path to addressable savings, 2026–2029: (1) Prove engagement and safety with 100 co-design users in 2026. (2) Prove retention and loneliness reduction at 1,000+ users in 2027. (3) Validate voice biomarkers against clinical outcomes at 5,000+ users in 2028. (4) NHS commissioning and outcomes-based contracts from 2029. Each phase produces the specific evidence the next requires.
Modelling assumptions reviewed by Dr Ivan Koychev (University of Oxford). Sources: Luchetti et al. (Nature Mental Health, 2024); Salinas et al. (Neurology, 2022); HomeHealth RCT; NHS RightCare (2020); Alzheimer's Research UK. Figures are addressable savings at population scale, not Harmonica revenue.
Honestly? Nothing, eventually. But here's why we'll have a head start that matters.
Elder tech isn't sexy.
Big Tech goes where the talent wants to go, and the best AI engineers want to work on coding agents and self-driving cars, not 80-year-olds in Newcastle. Aging has been a low-status market for twenty years. That's exactly why it's still open. By the time it becomes obvious enough that a Google PM gets staffed on it, we'll have a three-year head start in homes they haven't entered yet.
They have the same data problem we do, and we started first.
Nobody — not Google, not Apple, not OpenAI — has a dataset of real, multi-turn conversations with older adults. It doesn't exist on the internet. The only way to get it is to put a device in someone's living room and earn the right to keep it there. Every week we do that, the gap widens. We're betting we can hit escape velocity before anyone serious notices.
It's a hard market and most people don't want it.
Selling to 80-year-olds is slow. Hardware is painful. Clinical validation takes years. Care home procurement is a nightmare. Every one of these is a reason a big company's product team will pick a different roadmap — and a reason a focused startup with the stomach for it can win. We're not smarter than Google. We just want this one more.
ElliQ proved the market exists and we respect what Intuition Robotics has built. But they've made three structural choices we haven't, and each one is the opening.
First, the form factor screams "assistive device." A moving robotic head is memorable in a press photo but it signals "I need help" every time someone walks into the room. Olympia looks like something you'd buy from an Italian design studio — because it was. Design isn't a nice-to-have in this category, it's the adoption gatekeeper.
Second, robotics is expensive to manufacture and expensive to service. ElliQ is leased — $249 upfront and $39–59/month depending on commitment length — because the hardware is too expensive to sell outright and the motors need servicing. Our BOM target is sub-£150 at scale, with no moving parts to fail, no motors to service, no calibration. That lets us sell the device instead of leasing it, hit better unit economics, and give families ownership rather than a recurring hardware rental on top of the subscription.
Third — and most importantly — ElliQ is a companion, full stop. They haven't built the voice intelligence layer. No clinical biomarker pipeline, no peer-reviewed validation pathway, no data licensing strategy. They're selling a product. We're building a platform that happens to ship inside a product. That's where the long-term value lives.
Most SaaS gets less valuable over time as the novelty fades. Olympia gets more valuable every day because it learns. It remembers your daughter's name, your favourite stories, the neighbour you don't get along with, the medication you take on Tuesdays. After three months, it knows you better than the carer who visits twice a week. After a year, it holds a longitudinal picture of your cognition and emotional state that exists nowhere else — not with your GP, not with your family.
That creates two reinforcing moats.
At the user level: you don't cancel a friend. Unplugging Olympia means removing a companion that your parent has come to rely on — and doing so in front of them. The switching cost isn't financial, it's emotional, and it grows every week.
At the company level: every conversation makes the biomarker models more accurate, which makes the health signals more valuable, which makes the product more defensible to buyers further up the stack (care providers, NHS, pharma). Across 366 conversations and 55+ hours of scored voice data from just five users, the signal is already usable. At 1,000 users it becomes the best longitudinal voice dataset for aging adults in the world. At 10,000 it becomes infrastructure. A competitor starting today would need years of longitudinal data from a population that doesn't easily trust technology — and they'd be racing against our compounding dataset, not our head start.
We don't sell to seniors. We sell to their adult children.
The motion is B2C2Senior: the child buys, the parent receives it. The target customer is a 40–55-year-old professional living more than an hour from their parents, earning enough that $40/month is trivial against the alternative — which is care home fees, visiting carers, or the permanent low-grade anxiety of not knowing how Mum really is. That buyer is digitally native, reachable through Meta, Google, and targeted podcast sponsorships, and highly motivated. We're not selling a gadget. We're selling peace of mind and a way to feel like a good son or daughter.
Once we have that beachhead, care providers become a distribution channel: they deploy Olympia to their clients and we acquire users without direct CAC per head. Age UK alone has a volunteer network of 150,000 and a stated interest in companion technology — Prof. Cangelosi introduced us and we're meeting their national team in April. That's one channel partner that could move thousands of units without any consumer marketing spend.
The one thing we won't do is pay to acquire users who churn. We start narrow, prove retention, then scale channels where retention holds.
Because you can't walk into an NHS procurement meeting with a PowerPoint.
B2B and B2G buyers in this category — ICBs, memory clinics, care providers, and eventually NICE — need three things before they'll sign: engagement data from real users, clinical evidence that outcomes change, and testimonials from people who've lived with the product. None of those can be generated in a lab. All of them require getting the device into homes and keeping it there for months.
D2C is how we generate that evidence at the lowest possible cost and the fastest possible cycle time. Every paying family gives us a data point on retention, an economic signal on willingness to pay, and — when it works — a testimonial we can take into institutional conversations. Bill's unprompted 45-minute video, Chris's triaged head injury, Harry's Sherlock Holmes game: none of those moments happen in a pilot study. They happen when someone has lived with the product long enough to forget it's a product.
There's also a sequencing argument. NHS procurement cycles are 12–24 months. Care provider sales cycles are 6–12 months. D2C sales cycles are days. If we went straight to institutional, we'd spend a year in meetings before generating any learning. By the time we had a contract, we'd be building the wrong product. D2C keeps us close to the user, which is the only thing that matters at this stage.
B2B comes next — with the proof engine already running.