On this page
Two Different Tools, Not One Replacing the Other
The framing that gets benefits teams into trouble is “AI vs. human therapy,” as though a workforce has to pick one. In practice the two are built for different moments and different levels of need. A human therapist is a licensed clinician who can diagnose a condition, adapt treatment session to session, hold legal and ethical accountability for the care they provide, and intervene directly in a crisis. An AI chatbot is software: available at 2 a.m., free of scheduling friction, and useful for structured, low-acuity support — but it cannot diagnose, it cannot exercise clinical judgment the way a trained person can, and, as the sections below lay out, it carries safety limits that a human clinician does not.
None of that makes AI support worthless as a benefit. It makes it a different category of tool, and the question for HR isn’t whether AI is “as good as” a therapist — it’s where AI support is appropriate on its own, and where it needs to function as a bridge to a human rather than a substitute for one.
What the Independent Evidence Shows
Start with what the marketing claims. A widely cited npj Digital Medicine study (2019) reviewed the top-ranked mental-health apps and found that 64% claimed clinical effectiveness, but only 3.4% cited any supporting research — and much of that thin slice of research involved people who had helped build the app in the first place. That gap between marketing claims and independent evidence is the starting point for evaluating any vendor pitch, not a footnote to it.
Once you look past the marketing to the independent research, the picture is more measured than “does it work” suggests. An independent review of 31 randomized trials (2025) found that chatbots produce modest but real effects on depression and anxiety symptoms. The detail that matters for anyone evaluating a specific product: scripted chatbots outperformed generative, LLM-based ones in that review, and the effectiveness of the generative tools — the category most current vendor products fall into — was called “inconclusive.” A more narrowly scoped JMIR meta-analysis of 14 generative-AI-chatbot trials (2025) reinforces that same pattern: it found only a small overall effect, the anxiety-specific effect was not statistically significant, and none of the 14 trials were rated low risk of bias.
Put together, the independent evidence supports a narrow, specific claim — scripted, rules-based chatbots have a modest, real effect on mild-to-moderate symptoms — and does not yet support the broader claim most vendors imply, that a general-purpose, LLM-based conversational AI is a clinically validated substitute for therapy. If you want the fuller breakdown of study design, effect sizes, and where the evidence is strongest and weakest, see what the research on AI therapy effectiveness shows.
The Conflict-of-Interest Problem
There is a structural reason to read every one of those effectiveness numbers with some caution: essentially every published study of a named AI mental-health chatbot has at least one author who is employed by, founded, or advises the vendor whose product is being studied. That is not a claim that any individual study is fabricated — it is a description of the research landscape as it currently exists. A study designed, run, and published by people with a financial or professional stake in a positive result is a weaker piece of evidence than an arms-length trial run by researchers with no connection to the product, even when the methodology looks sound on paper.
For a benefits team, the practical implication is simple: when a vendor hands you a “clinically validated” study, ask who ran it and who funded it before you treat the headline number as independent proof. The npj Digital Medicine finding above — that most of the thin slice of cited research involved people who helped build the app — is the same pattern showing up at the industry level, not an isolated case.
The FDA and Safety Line
Regulatory status is where the comparison to a licensed human therapist becomes unambiguous. Per an American Psychological Association health advisory (November 2025), no AI chatbot currently has FDA approval to diagnose or treat any mental-health condition, and the advisory describes these tools’ ability to safely guide someone in crisis as “limited and unpredictable.” A licensed human therapist, by contrast, operates under a professional license, a duty of care, and established crisis protocols — accountability structures that a piece of software simply does not carry, regardless of how it is marketed.
That limitation is not theoretical. In 2023, the National Eating Disorders Association replaced its human helpline with a chatbot; per NPR’s reporting (2023), the bot gave harmful dieting advice to people who had reached out for eating-disorder support, and NEDA pulled it offline within days. That case is the clearest available example of what “limited and unpredictable” crisis handling looks like in practice, and it is exactly the kind of failure a benefits program needs to design around, not discover after the fact. We cover the specific mechanics of how a vendor should handle crisis escalation in a dedicated guide — it is one of the first things to check before signing a contract.
When to Route an Employee to a Human
Given the evidence above, the safest operating rule for a benefits program is that AI support is appropriate as a self-guided, low-acuity layer — and a human referral is the default the moment any of the following is true:
- Any indication of crisis or self-harm risk. Per the APA advisory above, chatbot crisis handling is not reliable enough to be the last line of support. A qualified vendor routes to a human crisis line or emergency service immediately, with no ambiguity in the product flow.
- Symptoms that look moderate to severe, not mild. The strongest independent evidence (the 2025 review of 31 RCTs) supports chatbot effectiveness for mild-to-moderate symptoms; it does not establish that generative chatbots are adequate for more serious presentations.
- A diagnosis is needed. No AI chatbot has FDA approval to diagnose anything. If an employee needs a formal diagnosis — for treatment planning, disability accommodation, or medication — that requires a licensed clinician, full stop.
- The employee asks for or clearly wants a human. Comfort with a self-guided tool is not universal, and a benefits program that only offers an AI option has no fallback for the employees it does not work for.
- No improvement after a reasonable trial period. If symptoms are not improving with self-guided use, that is itself a signal to escalate rather than a reason to keep waiting on the tool to work.
Building those triggers into the vendor contract and the product’s actual escalation flow — not just into an internal HR policy document — is the difference between a safety net and a liability. Our AI mental-health vendor RFP checklist has the specific questions to put to any vendor before you sign, including how they document and audit crisis-escalation performance.
Bottom Line
The independent evidence supports AI chatbots as a real, if modest, tool for mild-to-moderate depression and anxiety symptoms — with the caveat that scripted chatbots have outperformed generative ones in the strongest available review, and that most published effectiveness research carries a vendor conflict of interest. None of that evidence supports treating an AI chatbot as a substitute for a licensed human therapist: no chatbot has FDA approval to diagnose or treat a mental-health condition, and the APA’s November 2025 advisory describes chatbot crisis handling as limited and unpredictable — a description the 2023 NEDA incident illustrates directly. The responsible design for a benefits program is not “AI or human” but AI as a self-guided front line, with clear, contractually specified triggers for routing an employee to a human the moment risk, severity, or a diagnosis need appears.
This page is general information for benefits decision-makers, not legal, clinical, or financial advice — confirm the specifics of any vendor you evaluate with your own counsel and clinical advisors before you buy.
Frequently Asked Questions
Are AI chatbots as effective as human therapists?
The independent evidence does not support that. A 2025 review of 31 randomized trials found only modest effects for chatbots, a separate meta-analysis found the anxiety-specific effect was not statistically significant, and no AI chatbot has FDA approval to diagnose or treat any condition.
When should an employee be routed to a human rather than an AI tool?
Whenever there is any sign of crisis or self-harm risk, symptoms that look moderate to severe, a need for a formal diagnosis, or when the employee simply wants a person. AI support is best kept to a self-guided, low-acuity layer.
Why should vendor studies of AI chatbots be treated cautiously?
Because essentially every published study of a named AI mental-health chatbot has at least one author employed by, founding, or advising the vendor — a conflict-of-interest pattern that the strongest independent reviews specifically avoid.
Can an AI chatbot safely handle someone in crisis?
The American Psychological Association’s November 2025 advisory says these tools’ ability to safely guide someone in crisis is “limited and unpredictable.” In 2023, an eating-disorder helpline chatbot gave harmful advice and was pulled offline within days.
Related Reading
- AI Mental Health Tools for Employee Benefits: An HR Buyer’s Guide — the fuller vendor-evaluation guide this page drills into
- Is AI Therapy Effective? What the Research Shows — the fuller clinical-evidence picture behind the comparison on this page
- AI Mental Health Crisis Escalation for Employers — how a vendor should handle risk and crisis situations, and what to check before you sign
- AI Mental Health Vendor RFP Checklist — the specific questions to put to any vendor, including safety and escalation performance