Most people never ask.
Fifty-five million American adults had a mental illness last year. Twenty-six million of them received no treatment, and four out of five of those did not believe they needed any. Maison is an AI that runs on your own machine and remembers your life permanently. It is already there on the day nobody thinks to ask for help.
Fifty-five million American adults had a mental illness last year. Half received nothing at all.
The curve is not still climbing. Most of these measures peaked around 2021 and have partly retreated since, from a baseline far worse than 2013. Two in five high school students still report persistent sadness or hopelessness, against three in ten in 2013.
The gap is the part that has not moved. Prevalence goes up and down. The share of people with a diagnosable illness who receive nothing has been roughly stable for a decade, through a pandemic, a telehealth boom, and about fifteen billion dollars of venture funding.
Everything built for this problem was built for the people who go looking. Half of them do not.
Of the adults who went untreated, 78.6% did not believe they needed treatment.
Each dot is one of the 26.9 million adults with a mental illness who received no treatment. The 25.5 million above is the narrower count who received no support service either. Among untreated adults with serious mental illness, the share who saw no need is still three in five.
Among those who did perceive a need, why they still did not go
Self-reliance beats cost. Cost beats access. The binding constraint at population scale is recognition, and you cannot solve recognition with more supply.
Americans stopped spending time with each other, and the clearest evidence for it is not a survey.
American Time Use Survey, minutes per day, 2003 to 2020
Behavioral, nationally representative, and independent of whether anyone calls themselves lonely.
- 49% of Americans report three or fewer close friends, against 27% in 1990.
- Lacking social connection raises the risk of premature death by about as much as smoking up to fifteen cigarettes a day.
- Fewer than one in five people who often feel isolated recognize it as a major problem.
- Whether the phone caused the decline is genuinely unsettled. That the hours moved is not, and they came out of rooms with other people.
Whatever else is true about the last fifteen years, people stopped being in rooms together. Anything built here has to put them back in the room rather than stand in for it.
The modal experience of American mental health care is a prescription written by someone who is not a specialist.
Who is prescribing
About 75% of antidepressant fills are written in primary care
Psychiatrists and psychiatric nurse practitioners write roughly a quarter. Every other mental health professional, under one percent.
What psychiatry does now
78.4% of psychiatrist visits contain no psychotherapy at all
Against 55.6% in 1996. Between 2010 and 2016, 53% of psychiatrists stopped providing psychotherapy entirely.
Whether anyone is reachable
65% of non-metropolitan counties have no psychiatrist
The threshold for being called a shortage area at all is one psychiatrist per thirty thousand people.
What a directory is worth
3.1% of calls to an insurer's own listed psychiatrists produced an appointment
All 192 in one New York City directory were called. Half never responded.
And whoever you do reach is working from what you can remember about the last eight weeks.
The median mental health app keeps three of every hundred people for thirty days.
Each square is one of a hundred people who installed. Measured across apps with more than ten thousand installs, so the successful end of the market rather than the long tail. Trials in this category lose about 26% of participants. The market loses 96.7%. The difference is not the software. Trials pay people, enroll them in person, and remind them.
What that does to every published effect size
App against a waitlist, as published
g ≈ 0.45–0.56Against a comparator that actually does something
g ≈ 0.22After correcting for publication bias
g ≈ 0.18After correcting for who is still there on day thirty
≈ 0Improving an effect size from 0.28 to 0.40 is worth almost nothing. Improving thirty-day retention from 3.3% to 20% is worth a factor of six.
So the thing to build is not a better intervention. It is something a person is still using in a year. That is a memory problem, and it is the one thing this category has never had.
A persistent partner that runs on your hardware, remembers everything, and does real work.
Not a chat window. An operating system for how you run your life and your business.
- Its own memory, its own computer, its own hands.
- Reads and writes files, sends email, runs code, ships software.
- Wakes on a schedule to check what you asked about last week.
- Desktop, web, phone and terminal. Same memory at every door.
It runs on your hardware. For anyone whose information is genuinely sensitive, that is not a feature. It is the precondition.
Memory
Every fact is its own file
Typed, cross-linked, indexed, loaded by relevance rather than by scrolling back through a transcript. Episodic, semantic, procedural.
Execution
A real computer, not a description of one
Shell, filesystem, network. It writes the script, runs it, reads the error, and fixes it.
Orchestration
It delegates to itself
Sixteen sub-agents at once. One finds, another verifies, a third tries to prove the first two wrong. You get the conclusion, not the transcript.
Every number above is a continuity problem wearing a different hat.
Four in five never ask
It is already there
Maison is in your day for work, logistics, and the things you are trying to get done. It does not require anyone to decide they are unwell in order to be present on the day they are.
Eight weeks of recollection
It was there for all of it
Not a summary of you held on somebody's server. The actual record of your weeks, on your machine, in your hands, for as long as you want it.
3.3% at thirty days
It does not need to be opened
Retention in this category means remembering to open an app about your feelings. Maison is where the work already happens, so being there is not a thing anyone has to sustain.
The most sensitive data there is
It never leaves the house
Nothing has to be uploaded for any of this to work. There is no central store of anyone's interior life, because there is no central store.
Everything built for this problem so far asked a person to come to it. Maison is already in the room.
Engagement platforms measure time captured. We measure whether you needed us less and saw people more.
The last fifteen years
- Optimize for time on platform.
- Success is that you came back today.
- A good day is a longer session.
- Revenue rises when a person stays.
The objective function here, inverted
- Optimize for your life outside it.
- Success is that the work got done, the appointment got booked, and you were with someone.
- A good day is a shorter session.
- Revenue is a flat subscription that does not move when you do.
The objective function is the mechanism, not the medium. Products built to maximize time on platform succeeded at taking time, because that is what they were for. The charter forbids engagement objectives outright, and that covenant is enforceable by a body that is not the board.
Same technology, opposite objective. That inversion is what makes a system that remembers you for years worth building rather than worth worrying about.
It is not a mental health product bolted on. It is the thing you already use, on the day it matters.
For the person
- It is in your day for work and logistics and the things you are trying to finish. Nobody has to reclassify themselves to use it.
- It still has March when you say something in September, so it knows the difference between a hard week and a pattern.
- It does the errand the hard week made impossible. The appointment, the form, the email you have not been able to write.
- It can say that a professional might help, and then find one, check the network, and book it.
For the people around them
- One machine, one household, and everyone in it gets their own mind. A parent's Maison and a teenager's Maison cannot read each other.
- Schools and employers get an organizational memory while every person keeps their own. The institution never inherits the individual.
- The isolation is architectural rather than a permission setting, which is the only version of it a family will believe.
- And it is built to send people toward each other. The measure it holds is time with other people, not time with it.
The most common reason people give for not getting help is that they should be able to handle it on their own. This is built for exactly that person, and it never asks them to say otherwise.
Your clinician is working from eight weeks of your recollection. We can hand them eight weeks of record.
- At your direction, Maison prepares what you would otherwise have tried to remember. Sleep, routine, what changed and when, and what you actually did after the last visit.
- Medication management means the logistics of a regimen. Refills, timing, the pharmacy call, and what you noticed and when. It is never a dose recommendation and never a prescription.
- The clinician stays the clinician. Maison writes nothing into a chart and directs no clinical action.
The line, written as engineering
- No scored clinical thresholds anywhere in the product. No instrument scores, no severity labels, no risk stratification.
- No diagnosis, screening, monitoring, or treatment claim, in the interface or in the marketing.
- No emotion analysis presented as clinical inference, which Illinois law reaches directly and Nevada reaches even as a representation.
- Crisis escalation is deterministic and out of band. It does not depend on a model behaving correctly on the night it matters.
A wellness product that makes a clinical relationship better. That distinction is the product specification, and it is enforced in the build rather than reviewed by a lawyer before launch.
Psychiatry has the worst odds in drug development, and a large part of the reason is measurement.
The problem the industry has
- Psychiatry has the lowest probability of success in both Phase I and Phase II of the twenty-one largest therapeutic areas.
- Placebo response in antidepressant trials runs 31 to 40% against a drug response near 50, and has been flat since 1991.
- Much of that noise is the instrument. Asked to recall the last stretch of their illness, depressed inpatients rated 7 of 13 symptoms more severe than they had rated them in the moment.
- And after approval, more than 94% of adverse drug reactions are never reported at all.
What a consented Maison cohort is
- Dense, longitudinal, ordinary-day data, kept for reasons that have nothing to do with any study. Data gathered because someone enrolled is shaped by the enrolling. Data that was already there is not.
- The cohort does not have to be recruited and does not decay, because the reason it exists is a product people want for their work. That is the 3.3% figure read as an asset.
- The path is named rather than hypothetical: FDA runs a Patient-Focused Drug Development guidance series for patient-reported endpoints, and a Real-World Evidence program directed by the 21st Century Cures Act.
A sponsor can buy a question. A sponsor can never buy influence over what Maison says to a person. That firewall belongs in the charter beside the three covenants, because it is the objection this business has to answer before anyone asks it.
Three layers. Only one of them ever leaves, only if you say so, and only for a question you agreed to.
One · Your data
Never leaves
On your machine, yours, deletable. Not conditional on anything. HIPAA does not cover a product you own and run, which is exactly why this is architecture rather than a promise.
Two · The lesson layer
Methods, never content
What one Maison works out about how to help, every Maison can use, abstracted before it travels. No person in it, so no consent question. Your product improves on days you never open it.
Three · The evidence layer
Consented, per study
You join a specific question, not a general condition, and you can leave. Deleting removes you from the store and from the denominator. A tally already counted was never you.
What governs this is the FTC Act, not HIPAA. BetterHelp paid $7.8 million in 2023 for handing health data to advertisers; Cerebral paid $7 million in 2024. Three covenants close off both business models: no data sales, no engagement objectives, full export and deletion on demand.
And a position on inferred data, which nobody in this category has yet had to take. The things the system concluded that you never said are yours. Visible to you, and deletable by you.
You are never the product. When your data is worth something to someone else, you decide, and you are the one who is paid. Every incumbent has to ask forgiveness, because they already took it centrally. We are the only ones who have to ask.
Three channels, one architecture, and a price that is designed to fall.
Direct
The person who owns the machine
Consumer subscription is where this category dies on acquisition cost, and we are not pretending otherwise. It is not what the business is built on. It is what the mission requires, and the other two channels pay for it.
Organizations
Schools, employers, and families
The one channel with demonstrated durability here. Talkspace survived by abandoning direct-to-consumer; Lyra is an employer business. It is also the larger product, because organizational memory is worth more than personal memory.
Clinical partners
Practices and networks
Where the adjunct sits beside a clinician. The longest sales cycle of the three, and the one that produces the evidence the other two are judged by.
The top two tiers pay for the descent that makes the bottom one permanent. That is written into the charter as a public benefit rather than offered as a plan.
The company is bound to a purpose beyond its own profit, and to a body with standing to enforce it.
The charter names three public benefits
- Access. The greatest capability we can build, available to people generally, and more of it for every dollar our users pay.
- Distribution. The capability goes to the people who use it, not to any provider, including what the systems learn about them.
- Capability, not capture. Systems that increase judgment rather than engagement.
- The tax treatment is unchanged. A Delaware PBC is a C corporation.
- Three covenants close off business models we could run. They are also the commitments customers are buying.
Governance
An independent Stewardship Trust
A charter binds the company. It does not bind a future board under pressure. So the charter is paired with a trust holding a class of stock that carries no economics and one power: consent. Twelve trustees, none with a dollar of interest in the company they guard.
Three vetoes, and nothing else. No amendment to the purpose or the covenants, no change of control, no exit from public benefit status. It cannot touch the chief executive, the budget, the roadmap or the product.
Nobody should have to ask first.
Twenty-six million people did not get help last year, and most of them never went looking. Not because care was unavailable, but because nothing was there on the ordinary day, before it was a crisis, that already knew them.
It is already there
No one has to ask
Present on the day it matters because it was present on all the others.
It remembers
Years, not eight weeks
The difference between a hard week and a pattern.
It is yours
On your machine
The most private thing about you never leaves the house.
It sends you out
Toward people
Measured on your life, not on your session length.
A compass, not a cage. Built so that the capability ends up in ordinary hands, and bound by a charter that makes that obligation enforceable by someone other than us.
Confidential, and furnished for the recipient's own evaluation. This is not an offer to sell or a solicitation of an offer to buy any security; any offering will be made only to accredited investors under Rule 506(b) and only by means of the Private Placement Memorandum and definitive transaction documents, which supersede this page in their entirety. Figures are drawn from public sources believed reliable and carry the year and denominator of the study they come from. Forward-looking statements reflect assumptions that may not prove correct.