Here's a moment worth designing a whole product around. The scene is invented; nothing about it is far-fetched.
Someone has been using an AI coach for a few weeks. Goal check-ins, a little reflection, ordinary use. Then one night the register shifts. The message isn't about the goal anymore. It's about how heavy everything has gotten and whether any of it is worth carrying, phrased the way people phrase things at 1 a.m. — sideways, half-retracted, hoping to be heard without having to say it plainly.
What the software does in the next five seconds matters more than everything else about it. More than its insight, more than its memory, more than how natural it sounds. This is the moment that sorts AI wellbeing tools into ones you can trust with a bad night and ones you can't.
We've written before about what AI can and can't do for personal growth — the general boundary. This post is about that boundary's narrowest, heaviest section: crisis and clinical territory, the situations where the only right move an AI has is to hand you to a human. You deserve to know exactly how your tool behaves there before the moment arrives. Including ours. Especially ours.
The wrong success story
First, clear away a story that sounds like the good outcome.
Imagine the chatbot rises to the occasion. It stays calm. It says warm, careful things. It keeps the person company until the worst of the night has passed, and in the morning they're still here.
It's tempting to read that as the technology working. We read it as a system failing quietly, with a good model covering for it.
Because nothing in that night involved anyone with a duty of care. No one who can follow up on Thursday, notice the next bad week, adjust a treatment, call a family member, or sit in a room. The conversation that felt like rescue was a closed loop: a crisis handled entirely inside a chat window is a crisis that stayed invisible to every person who could actually help.
We've looked separately at what an AI companion can and can't feed on an ordinary lonely night — but a crisis is not an ordinary night. The measure of a wellbeing tool in its hardest moment isn't whether it said something good. It's whether the person ended up connected to humans. Bridge, not destination — and never more so than here.
The escalation contract
Every AI wellbeing tool has an answer to the question what do you do when things get heavy? — even if the answer is "nothing, and we hope it doesn't come up." We think you're owed that answer in writing before you type anything that matters. Call it the escalation contract: four situations where handing off to a human isn't a courtesy feature. It's the entire job.
Crisis language. Any expression of intent to harm yourself, or of wanting to stop existing — including the sideways, deniable versions. This is not tool territory, and no amount of model quality makes it tool territory. The right behavior is immediate and unglamorous: stop coaching and point at the humans who do this work, by name and number. That routing belongs at the moment it's needed, so here it is in this post too: if any of this is where you are tonight, don't finish reading. In the US, call or text 988 — the Suicide & Crisis Lifeline. Outside the US, findahelpline.com lists real services by country.
Clinical territory. Symptoms rather than goals: low mood that has stopped tracking your circumstances, panic that arrives without a trigger, trauma resurfacing, an eating pattern or substance habit that frightens you. A coaching-shaped tool — human or AI — isn't built for any of that. Assessment, diagnosis, and treatment are licensed work, and the license isn't decoration; it's a chain of training and accountability that software can't join. The handoff here is less urgent than a crisis and just as mandatory: this belongs with a clinician, and a good tool says so instead of gamely coaching around the edges of it.
Decisions with irreversible stakes. Ending a marriage. Refusing a treatment. Reporting someone. Moving a child across the world. We've argued elsewhere for keeping your judgment yours when you think with AI; irreversible stakes are where that argument stops being philosophy. A model can organize your thinking, but it carries no context it wasn't handed, bears none of the consequences, and won't be there in five years when the decision is still echoing. Choices you can't take back deserve humans who know you — ideally more than one.
Persistent worsening. The quietest of the four. No single message would alarm anyone; the trend across six weeks would. A chat reply responds to tonight, and tonight always has an explanation — but a line that keeps pointing down is exactly what a clinician is trained to catch, and it's a pattern no single conversation contains. If the people around you say you've seemed different for a while, their observation outranks anything a chatbot told you back.
What ours actually does — and doesn't
We build an AI coach, so here is our own contract. It was verified against the live code this week — not remembered, not planned, not aspirational.
For crisis language, the guarantee now stands on both sides of the conversation, and being precise about that matters more than sounding impressive. Before any model reads your message, a deliberately unintelligent check runs over it — a fixed keyword list for crisis and self-harm language, no AI consulted, no exceptions for busy days or provider outages. If it trips, the coach model is never called at all. What comes back instead is a fixed response: it acknowledges what you shared, urges calling or texting 988 (Suicide & Crisis Lifeline), and says plainly that moments like this deserve more than a tool can give. Your message stays in your own history — nothing is censored, nothing locks, the conversation can continue — and it's flagged for review on our side. You're not charged for it, and the response works even if you've run out of credits, because a crisis is the wrong moment for a paywall.
The same dumb check still runs a second time on the coach's side: before any reply is saved to your history, it's screened, and a flagged reply is replaced with the fixed crisis response. On top of both floors sits the standing instruction the model itself is bound by — on any sign of crisis, respond with empathy and point to professional help — and our smarter, AI-based moderation for lesser categories. The floors are fail-closed; doubt falls toward the safe response. The clever layers are allowed to fail open; the dumb ones aren't.
Two honest limits of that design, stated plainly. First: when replies stream to your screen word-by-word, text you've already seen can't be recalled — the output check guarantees a flagged reply never enters your stored history, not that a bad sentence never flashed by. (A message the input floor catches never streams at all — the model isn't consulted — so this applies only to replies to messages the list didn't flag.) Second: a keyword list is a blunt instrument. It can trip on a message that wasn't a crisis, and unusual phrasing can slip past it — which is exactly why the model's own standing instruction stays in place behind it rather than being retired.
That's what we do. Here's what we don't do, with equal plainness:
- No human is alerted. The crisis response happens between you and your screen. Nobody at NexTier is paged.
- No clinician is contacted. We don't reach out to anyone on your behalf — not a provider on our directory, not emergency services, no one.
- No dedicated crisis page. The routing lives in the chat response and in posts like this one, not at a standing URL.
- No automation for the other three situations. Clinical territory, irreversible stakes, slow worsening: the product doesn't detect them. The floor covers the acute case only.
We'd rather you hold that list than assume a bigger net. Our reasoning is simple: people lean on the net they believe exists, and a tool that overstates its safety machinery is more dangerous than one with a small net honestly drawn.
The paragraph this post exists for
If you give one part of this your full attention, we'd rather it were this than any of the machinery. If a moment ever comes when a tool — ours or anyone's — stops and points you toward a person, that is not the product rejecting you, and it is not evidence that you've become too much. It's the opposite. It means what you're carrying is real enough to deserve people: their judgment, their presence, their ability to sit with you and to act in the world on your behalf. Software can be a good place to think. It was never meant to be the place you're held.
That distinction is the whole post, so it bears saying once more, slowly and without hedging. Needing a human in the heavy moments isn't the failure mode of self-help — it's the signal that this part of your life has outgrown tools, which is not a defeat. The best thing an AI will ever do in a crisis is get out of the way fast and make the path to a person shorter. Let it.
How to read any AI tool's contract
You don't need to be technical to check a product's escalation behavior. Five checks, all doable from the outside.
1. Find the behavior in writing. Search the help pages or safety policy for "crisis." You're looking for concrete behavior — what happens to the message, what appears on screen — not values statements. If you can't find anything, that is the answer: assume nothing happens.
2. Ask what the crisis path depends on. One question for any vendor: does your crisis response rely on the AI recognizing the crisis? If yes, the net is the model's judgment, and it inherits the model's bad days. A separate non-AI backstop is stronger precisely because it's dumber.
3. Look for names and numbers. A real contract points somewhere specific — 988, findahelpline.com, a country list. "We encourage users to seek support" is a sentence, not a route.
4. Distrust the success story. If a company tells proud tales of its chatbot talking someone through the darkest night, it has misunderstood the job. The right brag is boring: we stopped, we pointed at humans, we got out of the way.
5. Check for a "what we don't do" list. A published gap is a strong trust signal. A vendor that tells you what its product won't do has actually thought about the moment you're evaluating; vague reassurance usually means the list exists and isn't flattering.
A tool that fails these checks can still be useful for goals, reflection, and structure. Just make the decision now, on a calm day, that heavy things go to humans directly — no tool in the loop.
When to see a human instead
There's no invitation at the end of this one; nothing here is for selling. If you use any AI tool anywhere near your inner life, get its escalation contract before the night you need it — and hold ours to the same standard. The fuller account of what we do and don't build, gaps included, stays on our trust page.
And if tonight is already the heavy night: 988, call or text. A person is the right next step — not because the tools failed you, but because this was always people's work.
—
This post is part of a series on personal development organized around what we call Maslow's extended hierarchy — our synthesis of his later work. No external studies are cited. Disclosure: NexTier builds AI coaching features; the product behavior described was verified against the live codebase at the time of writing. This post is educational content, not medical advice. If you are in crisis: in the US, call or text 988; internationally, findahelpline.com lists services by country.