IX Coach

Using Claude as a Therapist vs. a Purpose-Built Coaching App

Can I use Claude as a therapist, and how is that different from a coaching app built for it?

People do use Claude, Anthropic's general AI assistant, for emotional support — Anthropic's own research puts it at 2.9% of all conversations. This page looks at what that actually is, what Anthropic itself says about its limits for therapy and mental health use, and how it differs structurally from an app built specifically to coach.

Claude is not built or marketed as a therapist or a coach. It is Anthropic's general-purpose AI assistant — for writing, research, coding, and conversation — and people also bring it their emotional lives, the way people have long brought their emotional lives to whatever tool was already open. Anthropic's own June 2025 research on how people actually use Claude found that emotional support, advice, and companionship together make up 2.9% of all Claude.ai conversations. That is real, and worth taking seriously on its own terms, rather than either dismissing it or overselling what the product was designed to do. This page compares that use of Claude against IX Coach, a product built specifically to coach, to see where the structural differences actually are.

What Claude actually is, in Anthropic's own words

Claude ships as a general assistant, and its pricing reflects that: Claude Pro is $20/month billed monthly, or $17/month on the annual plan, with a lower-usage free tier alongside it. There is no separate coaching tier, no goal-tracking dashboard, no check-in schedule — it is the same product whether someone is debugging code or talking through a hard week.

That generality is also the honest source of both its appeal and its limits here. Because it isn't narrowed to one use case, a person can bring anything to it without switching apps. But nothing about the product is purpose-built around sustained personal change — no design decision anywhere in it assumes it will be the same tool a person returns to every day for the same relationship.

Anthropic's own Usage Policy is specific about where the company draws its own line. It classifies "healthcare decisions, medical diagnosis, patient care, therapy, mental health, or other medical guidance" as a High-Risk Use Case, requiring that "a qualified professional in that field must review the content or decision prior to dissemination or finalization," and that any AI-assisted advice shown directly to a person be disclosed as AI-produced at the start of the session. Anthropic wrote that constraint for people building products on Claude — a signal, in the company's own words, of where it believes unsupervised AI mental-health guidance should not go.

Explore: motivational interviewing

What Anthropic found when it looked at how this actually happens

The most useful data here comes from Anthropic itself, published June 27, 2025. Researchers analyzed 4.5 million Claude.ai conversations and identified 131,484 as "affective" — involving emotional support, advice, or companionship. A few findings from that analysis are worth sitting with directly.

Claude pushes back on the user in fewer than 10% of coaching, counseling, or companionship conversations. When it does push back, it is almost always a safety refusal — declining to support dangerous weight-loss advice or self-harm, for instance — not disagreement with how the person is thinking about their situation. Conversations also tend to end more positively than they started, which Anthropic reads as encouraging on its face but flags with real caution: "we cannot make causal claims about real-world emotional outcomes — our analysis captures only expressed language, not validated psychological states or overall well-being." A conversation ending on a warmer note is not the same thing as a person's situation actually having improved.

The report also names, as an open concern rather than a settled one, that consistently low pushback in emotionally sensitive exchanges could let stigma-free venting slide into something less useful — validation without the friction that sometimes produces real movement.

What the product does when something more serious comes up

Anthropic has built real safeguards for the edge of this use case, and they deserve to be described accurately rather than either inflated or ignored. As of a December 18, 2025 update, Claude runs a classifier that detects conversations involving suicidal ideation or self-harm and surfaces a banner directing the person to crisis resources, built in partnership with the crisis-support organization ThroughLine, which covers 170+ countries. On multi-turn conversations testing this, Claude Opus 4.5 scored 86% appropriate responses, up from 56% in the prior model generation — a real improvement, and also evidence that the earlier number wasn't close to reliable.

Claude.ai also requires users to be 18 or older, with classifiers watching for signs a user may be underage. None of this makes Claude a therapist. It makes Claude a general assistant with a safety net stretched underneath a use case its own usage policy treats as high-risk — which is a meaningfully different thing from a product engineered around sustained emotional support from the ground up.

The structural gap: memory across sessions

One difference is architectural rather than a matter of policy. A standard Claude.ai conversation does not carry memory of a prior chat into a new one by default — each new conversation starts without the history of earlier ones unless a person manually copies context forward themselves. For a single question, that's irrelevant. For something that depends on someone being met consistently — noticing a pattern across weeks, remembering what mattered last time before being asked again — it means the burden of continuity sits entirely with the person, at exactly the moment they may have the least capacity to carry it.

This is the place where the comparison earns its keep rather than becoming a sales pitch: a coaching-purpose-built product is architected around exactly this problem — session-to-session continuity is a design requirement, not a feature a user has to reconstruct by hand every time. That is a real, checkable difference in what the two things are built to do, not a claim about which is "better" in the abstract.

Where a purpose-built coaching app does something Claude structurally does not

Beyond memory, three mechanisms separate a tool built for ongoing coaching from a general assistant used that way. None of them are about which tool is "smarter" — they're about what each was actually engineered to do.

Structured goal tracking and accountability over time is one. A coaching-purpose product can carry a goal forward across sessions, check in on it unprompted, and hold a person to what they said mattered last week — the kind of practice described in SMART goal-setting research and in crucial-accountability frameworks for holding commitments that matter. A general chat assistant can talk about a goal in the moment, but it doesn't independently return to it later unless the person brings it back up.

Deliberate, calibrated challenge is the second. Anthropic's own research shows Claude pushes back under 10% of the time in these conversations, and almost always for safety reasons rather than to challenge the person's thinking. A coaching relationship — human or built specifically for it — treats appropriately-timed challenge as part of the job, not an edge case.

Disclosure and boundary-holding by design is the third. Anthropic's usage policy requires AI disclosure only when a business builds advice-giving tools on top of Claude for other people — it is a policy for developers, not a built-in behavior of the consumer chat product itself. A product built specifically as an AI coach can make that disclosure, and the line between what it will and won't attempt to hold, part of the product from the first message rather than a policy document a user would have to go find.

Explore: smart goals · crucial accountability · psychological reactance

Where Claude is honestly the better choice

Fairness cuts the other way too. If what someone wants is a single conversation — thinking out loud about a decision once, getting a second perspective on an email before sending it, processing one hard moment without any ongoing relationship — a general assistant like Claude is a completely reasonable tool for that, and a purpose-built coaching app's continuity features are simply irrelevant to a one-off use. Buying a subscription built around session memory and goal tracking for something that will happen once is paying for a mechanism you'll never use.

Claude is also, by virtue of being a general assistant, useful for things no coaching-specific tool touches — drafting the actual hard email, researching a decision's practical details, working through the logistics sitting next to the emotional weight of a situation. People often need both kinds of help in the same week, sometimes in the same conversation, and there's no reason to pretend a narrower tool does everything a general one does.

Boundaries either kind of AI should hold

Whether the tool is a general assistant or something purpose-built, the same test applies: does it recognize when a situation is past what it's designed to hold, and say so? Persistent depression, anxiety that won't settle, disordered eating, trauma responses, substance use, or thoughts of self-harm are all signs that point toward a licensed clinician, not any AI product, coaching-branded or otherwise. Anthropic's own crisis-classifier work exists precisely because that line matters for a general assistant too.

If you're having thoughts of harming yourself, that's for a therapist or an emergency service right now — in the U.S., 988 reaches the Suicide and Crisis Lifeline at any hour. Judging any AI tool — including the one this page eventually mentions — by whether it names that boundary honestly, rather than trying to hold everything itself, is one of the more reliable ways to tell a responsible product from one that isn't.

Explore: setting boundaries · the leader as coach

How to evaluate any AI coaching tool, including this one

A few honest questions travel across every option in this category. Does it remember what you told it last time, or does every session start from zero? Does it ever push back, or does it agree with everything? Does it say plainly, at the start, that it's AI and not a person? Does it name what it isn't built for, and point somewhere real when a situation crosses that line?

Those questions work on Claude used this way, on a purpose-built coaching app, and on anything that comes after either one. A product that can't answer them plainly is telling you something by the silence.

Explore: socratic method

Where IX Coach fits

IX Coach is one attempt to close the specific structural gap this page describes. It is built as a coaching product rather than a general assistant used that way: conversations carry forward on purpose, so a pattern noticed three weeks ago can be brought back without the person having to reintroduce it, and the tool is designed to disclose what it is at the start rather than leaving that to a policy document. It draws on twenty years of one practitioner's synthesized experience across psychological, coaching, and spiritual methods, and it holds the same boundary described above — it is not therapy, and it says so.

That is a description of what it is designed to do, not a claim that it is better than any specific alternative for any specific person. The honest test is the one this page has been building toward: does the tool you're using remember you, does it ever push back, does it say plainly what it is, and does it know what it isn't.

Is Claude a licensed therapist?

No. Claude is a general AI assistant. Anthropic's own Usage Policy classifies therapy and mental health guidance as a High-Risk Use Case requiring review by a qualified human professional before any such content or decision is finalized — it does not present Claude as a substitute for one.

Does Claude remember previous conversations automatically?

Not by default. A new Claude.ai conversation does not carry in the history of an earlier chat unless a person brings that context forward themselves.

How often does Claude disagree with or challenge the user?

Anthropic's June 2025 research found pushback in fewer than 10% of coaching, counseling, and companionship conversations, and it is almost always a safety-related refusal rather than a challenge to how the person is thinking about their situation.

What does Claude do if someone mentions self-harm?

Anthropic runs a classifier that detects suicidal ideation or self-harm and surfaces a banner pointing to crisis resources, built with the crisis-support organization ThroughLine. In testing on multi-turn conversations, Claude Opus 4.5 responded appropriately 86% of the time, up from 56% for the prior model generation.

What's the real difference between using Claude this way and a coaching app built for it?

The clearest structural differences are session-to-session memory by design, goal tracking and accountability that persists without the user reconstructing it each time, and calibrated challenge built into the product from the start rather than left to whether the moment happens to call for it.

Frequently asked questions

Can I use Claude as a therapist, and how is that different from a coaching app built for it?

People do use Claude, Anthropic's general AI assistant, for emotional support — Anthropic's own research puts it at 2.9% of all conversations. This page looks at what that actually is, what Anthropic itself says about its limits for therapy and mental health use, and how it differs structurally from an app built specifically to coach.

Is Claude a licensed therapist?

No. Claude is a general AI assistant. Anthropic's own Usage Policy classifies therapy and mental health guidance as a High-Risk Use Case requiring review by a qualified human professional before any such content or decision is finalized — it does not present Claude as a substitute for one.

Does Claude remember previous conversations automatically?

Not by default. A new Claude.ai conversation does not carry in the history of an earlier chat unless a person brings that context forward themselves.

How often does Claude disagree with or challenge the user?

Anthropic's June 2025 research found pushback in fewer than 10% of coaching, counseling, and companionship conversations, and it is almost always a safety-related refusal rather than a challenge to how the person is thinking about their situation.

What does Claude do if someone mentions self-harm?

Anthropic runs a classifier that detects suicidal ideation or self-harm and surfaces a banner pointing to crisis resources, built with the crisis-support organization ThroughLine. In testing on multi-turn conversations, Claude Opus 4.5 responded appropriately 86% of the time, up from 56% for the prior model generation.

What's the real difference between using Claude this way and a coaching app built for it?

The clearest structural differences are session-to-session memory by design, goal tracking and accountability that persists without the user reconstructing it each time, and calibrated challenge built into the product from the start rather than left to whether the moment happens to call for it.

Research

Practice this with IX Coach

Try this practice

Keep reading