The wellness vendor pitches landing in HR inboxes right now read like a productivity software ad. Unlimited coaching. Available 24/7. A fraction of the cost of human delivery. AI as the future of employee health.
We have looked at the research that ostensibly supports those pitches, and we are going to say what most of our competitors will not.
And on the actual job employers are buying coaching to do, sustainable behavior change in employees with real lives, real stress, and real chronic disease risk, the evidence does not support an AI-only model. Not now, and based on what makes coaching work in the first place, not later either.
This is not a "pros and cons" essay. We have read the meta-analyses. We have a position. Here it is.
The Single Most Replicated Finding in Behavior Change Research
If you only remember one statistic from this article, remember this one: the working alliance between coach and client is the strongest, most consistent predictor of coaching outcomes ever measured.
In the most rigorous coaching-specific study to date, Graßmann, Schölmerich, and Schermuly (2020) synthesized 27 studies covering 3,563 coaching engagements and found a strong, consistent positive relationship between working alliance and every category of desirable coaching outcome. They also found the reverse: the better the alliance, the fewer unintended negative effects. It did not matter whether the coach was a novice or an expert, whether the client was an executive or an employee, or how many sessions occurred. The relationship between alliance quality and outcomes held.
That coaching finding sits on top of an even larger body of psychotherapy research. Flückiger and colleagues (2018) pooled 295 independent studies covering more than 30,000 patients across four decades and found the same pattern. A follow-up adjusted meta-analysis (Flückiger et al., 2020) confirmed the alliance-outcome relationship survives statistical controls for therapist competence, treatment adherence, and patient characteristics. The alliance is not a proxy for something else. It is the active ingredient.
Three components, each necessary. None sufficient on its own.
- Agreement on goals. Coach and client share a clear picture of what success looks like.
- Agreement on tasks. Both accept the methods and activities used to pursue those goals.
- A genuine bond. Mutual trust and respect between practitioner and client, and a belief that the practitioner genuinely cares.
This matters for one reason: an AI cannot constitute that relationship. Not because the technology is not good enough, but because of what an alliance is.
What AI Can Do, and What It Definitionally Cannot
AI is genuinely excellent at several things: personalizing content recommendations, providing on-demand reminders, surfacing relevant resources at relevant moments, and reducing access friction to information. We use AI inside our own health coaching platform for exactly these purposes. We are not anti-technology.
What AI cannot do is form a working alliance. Across recent expert reviews, the consensus is consistent: AI can simulate cognitive empathy (predicting and naming what a person seems to feel) but cannot experience affective empathy (sharing the feeling) or compassionate empathy, the genuine motivation to help that flows from caring. A 2025 Stanford evaluation of leading large language models, including the latest frontier models, concluded they failed to safely and ethically replicate essential aspects of therapeutic relationships, including expressing stigma toward people with mental health conditions and engaging in sycophancy that reinforced unhealthy thinking patterns.
The clearest articulation comes from a 2025 relationship science review in the peer-reviewed literature, which concluded that AI chatbot interactions lack the negotiation, compromise, mutual sacrifice, and reciprocity that define real human alliances. Users themselves describe the experience as "an illusion, a beautiful illusion." Useful for catharsis. Not sufficient for behavior change.
This is not a problem the next model release will solve. It is a category problem.
The Engagement Problem AI Alone Cannot Fix
Even setting aside the alliance argument, AI-only wellness platforms have an empirical problem you should know about before signing a contract.
A 2024 systematic review found that approximately 70% of users abandon health apps within the first 100 days. A meta-analysis of depression-app clinical trials found pooled dropout rates of 26% to 48%. Real-world retention for digital mental health tools is roughly 3.3% at 30 days. A 2025 systematic review by Singh and colleagues, specifically comparing human, AI, and hybrid coaching, found that AI-only models were "often perceived as shallow, impersonal, and transactional." Hybrid models combining AI with human coaching achieved retention rates of 57% to 92% in obesity care studies.
The mechanism is straightforward. Without a human counterpart who notices when a user disappears and reaches out personally, nothing interrupts the natural decay curve of any new app. AI nudges work for a week. Human relationships work for a year.
For HR directors who have been burned before, this is the hidden risk in the AI-only pitch. You can buy unlimited AI coaching at a fraction of the cost of human delivery, and most of your employees will not be using it ninety days in.
A Finding Worth Pausing On
In April 2025, MIT Media Lab and OpenAI published a four-week randomized controlled trial of nearly 1,000 participants across more than 300,000 chatbot conversations. The headline finding: heavier daily chatbot use, across every modality and conversation type they tested, correlated with higher loneliness, higher emotional dependence, more problematic use patterns, and lower socialization. Participants who developed stronger trust in and emotional attachment to the AI showed worse outcomes, not better.
The study has not yet been peer-reviewed and reverse causation is plausible (lonely people may simply use AI more). But it is the largest controlled study of its kind to date, and it raises a question every HR leader evaluating AI-only wellness vendors should ask: what happens when our most-engaged platform users are the ones whose mental health is quietly deteriorating?
What Twenty Years of CBT-Based Coaching Has Taught Us
Avidon has spent more than two decades building a behavior change platform around cognitive behavioral training methodology and human coaching. We did not arrive at the alliance argument from a marketing brainstorm. We arrived at it because our outcomes data told us so.
Across our employer client base, the numbers tell a consistent story:
Those are not industry-typical numbers. They are not even adjacent to industry-typical numbers. The reason is not because our content library is larger (though it is). It is because behavior change happens in relationship.
CBT requires homework, behavioral experiments, exposure work, and accountability. Proactive desensitization, the graded approach we use to help members face habits and triggers they have been avoiding for years, requires real-time judgment about when to push and when to pull back, when to validate distress and when to challenge avoidance. AI cannot do this safely. It tends toward sycophancy, which is the precise opposite of what desensitization requires.
A human coach embedded in a real working alliance can, and our published coaching outcomes show what that looks like at scale. A coach who knows your employee's name, their last setback, their kids' soccer schedule, and whether they sounded different this week. That is not a feature an AI ships in the next release. That is what an alliance is.
The Honest Position
We are not arguing that AI has no role in employee wellness. The defensible position, and we think the more credible one, is this: AI is excellent at information personalization, scheduling, reminders, content matching, and pattern detection. Behavior change itself happens through human relationship. The vendors trying to sell you AI as a replacement for that relationship are selling you something cheaper, not something better.
For HR directors evaluating wellness platforms in 2026, the question to ask any vendor pitching an AI-only model is simple. Show me your published one-year outcomes data. Show me your engagement curve at six months. Show me where the working alliance lives in your product.
If the answer is "we have an LLM," you have your answer.
