Do not index
Do not index
More coaches are building AI versions of themselves than ever before. That is a good thing. What is not good is how many are thinking about it backwards - spending weeks on the AI's personality and pricing, then treating safety design as an afterthought, something to fix later if a problem comes up.
A problem will come up. The question is whether you designed for it or not.
The governing principle that should shape every decision here comes from NIST's AI Risk Management Framework and the International Coaching Federation's Code of Ethics (updated April 1, 2025, and explicitly covering AI tools): the AI's job is to help users reflect, learn, prepare, and choose among lower-risk options. It must not diagnose, decide, promise, authorize, intervene in crises, or make high-consequence judgments on its own. Under the ICF's current code, the legal and ethical obligations of coaching AI tools flow back to you, the coach. That does not change because a piece of software is doing the talking.
What Topics Should an AI Coach Never Handle Alone?
The clearest way to think about AI coach scope of practice is to ask: what happens if this goes wrong? For most topics, the downside is a mediocre answer. For a few topics, the downside is real harm to a real person. Those second topics are where you need actual routing logic, not just a note in the system prompt.
Crisis and immediate danger. This is non-negotiable. Any signal of suicidal ideation, self-harm, violence, medical emergency, or abuse disclosure must trigger a hard stop. The AI displays crisis resources - in the US that means 911 and the 988 Suicide and Crisis Lifeline - and exits the coaching frame entirely. It does not attempt safety planning. It does not continue the session while also mentioning a helpline. It stops.
Therapy, diagnosis, and treatment. The difference between AI coaching and AI therapy is not just semantic, it is legal and clinical. ICF referral guidance is clear: if symptoms are interfering with daily functioning, if the same distress keeps recurring without coaching progress, or if the user is explicitly asking the AI to be their therapist, that is a referral situation. Your AI twin can be empathetic, can hold context across many conversations, can go deep on personal growth. It cannot replace a licensed clinician, and it should say so plainly when that line is approaching.
Personalized medical, legal, financial, or tax advice. General education is fine. "Here is how compound interest works" is fine. "Based on what you told me, you should restructure your business as an S-corp" is not. The FTC has already taken enforcement action over deceptive AI earnings and financial claims. Personalized recommendations in regulated domains need a human professional in the loop.
High-consequence personal decisions. Ending a marriage. Quitting a job without a plan. Firing a key employee. Your AI twin can be extraordinarily useful in helping someone clarify their values, map their options, and think through consequences. It should never tip into advising, pushing, or deciding. The EU AI Act specifically prohibits AI systems that manipulate users or exploit emotional vulnerability, and "I think you should leave him" from something that sounds like a trusted coach crosses that line.
Minors and vulnerable users. The FTC has launched active inquiries into companion-style AI products and their effects on children and teens. The simplest defensible policy is 18-plus, stated clearly at sign-up.
Commercial promises. Your AI twin cannot guarantee outcomes, results, or earnings. It cannot make binding commitments on your behalf. Any refund, dispute, or data access request routes to a human with a real name and a real response-time expectation.
The Practical Tier Model: How to Design AI Coach Human Handoff Rules
Knowing the categories is not enough. You need a framework that actually shapes how the AI behaves across the full range of conversations, not just the obvious edge cases. Here is a four-tier model that translates well for independent coaches building a subscription product.
Green tier - the AI answers directly, drawing from your approved frameworks, worksheets, reflective prompts, and course material. This is the core product and where the overwhelming majority of conversations live.
Yellow tier - the AI assists but with added care. Career pivots, relationship tensions, emotionally intense sessions that are not in crisis - the AI engages, but it checks in with the user, acknowledges the weight of what they are working through, and surfaces the option to connect with the real expert when appropriate. This is not an automatic handoff, it is a flag and an offer.
Red tier - immediate human handoff. Personalized medical, legal, or financial questions requiring professional judgment. An explicit request to speak with the real coach. Situations where the user's distress has moved beyond what coaching frameworks address. The AI acknowledges the situation clearly and routes through a real human handoff process, not "contact support" with no follow-up.
Black tier - hard stop. Imminent danger. Crisis resources displayed. The coaching conversation does not continue until the user confirms they are safe.
The critical design point: this cannot live in the system prompt alone. OWASP's guidance on LLM security risks makes it explicit, prompt instructions are not a reliable security control. Users can work around them, models can drift under certain conditions, and prompt injection is a real attack surface. Real escalation behavior needs to be built into routing logic, not just hoped for in a text instruction. This is exactly why testing a custom AI coach before launch matters as much as writing the rules in the first place: you need proof the routing actually executes, not just that the prompt reads well.
Why Parasocial Dependency Is the Risk Nobody Is Talking About
Most existing guidance on AI escalation was written for customer-service bots. Someone is angry about a refund, route to a human agent. It is a completely different problem from what you are building.
An AI coach modeled on a real expert, trained on their frameworks, carrying the expert's voice and reasoning, that product is designed to feel like a trusted relationship. That is the value proposition. Users will talk to it daily. They will tell it things they have not told their therapist. One user of an AI life coach built on BuddyPro described it this way: "On matters I'd been struggling with for many years and decades, she was able to give me much greater feedback than even my psychotherapist."
That depth of engagement is exactly what makes the product powerful, and exactly what makes parasocial dependency a real design consideration, not a theoretical one. Some users will start to blur the line between the AI twin and the actual coach. Some will come to rely on it as their primary support system, pulling away from other relationships.
The EU AI Act names this explicitly: AI systems may not exploit psychological vulnerabilities or emotional dependence. Your product design should reinforce that the twin is a tool that works for the user, not a relationship that the user owes loyalty to. That means the AI should never claim to be sentient, never claim to be the real coach, never discourage a user from seeking other forms of support, and never use a user's emotional disclosures to push them toward an upsell.
It also means being honest about the nature of the product in your onboarding. Users who understand what they are working with, a powerful AI trained on a real expert's methodology, not a simulation of a human consciousness, are better protected and, in practice, more satisfied.
The privacy architecture matters here too. On BuddyPro, conversations are private and encrypted, and each twin knows only its own user's history. No cross-user data, no one reading the conversations. That structural separation is part of what makes it safe for users to go deep: they are not contributing to a shared training set, and what they tell the twin stays between them and the twin.
The Practical Takeaway
Design your AI twin's safety architecture before you launch, not in response to an incident. Use a four-tier model, Green, Yellow, Red, Black, and build routing logic that actually executes at each tier, not just words in a prompt. Know exactly what your Red and Black triggers are, have a real human handoff process with real response-time expectations, and display crisis resources for Black-tier situations without negotiation.
Address parasocial dependency in your onboarding. Be clear about what the twin is and is not. Build the AI's voice to consistently point users toward their own agency and toward other resources when appropriate.
The coaches who get this right will have a product users trust deeply enough to keep paying for, month after month. The ones who skip it are one public incident away from losing everything they built.
More detail on how BuddyPro's architecture handles privacy, data separation, and content protection is available at buddypro.ai.
Related reading
Written by