542 Dead Mental Health Startups: What the Failure Data Says About AI That Replaces Therapists

In 2021, Pear Therapeutics was worth $1.6 billion. It had FDA clearance for prescription digital therapeutics, real clinical trials, and a marquee leadership team. Two years later it filed for bankruptcy, and its assets sold at auction for $6 million. The regulator had said yes. Nobody agreed to pay.

Pear is one row in an unusual dataset published this year: a catalogue of 542 digital mental health companies that shut down, went bankrupt, pivoted, or were absorbed between 2000 and 2026. The author, a founder building in the space himself, coded each company on who paid, what the product claimed to do, whether clinical evidence existed, and how it ended. The result is the closest thing this field has to actuarial tables.

One finding matters more than the rest for practicing clinicians, because it settles a question therapists are asked constantly: what happens to products built to replace the therapist?

They die. At roughly twice the rate of everything else.

The replacement finding

The study coded each product into a role. "Clinical replacement" means treating a condition without a human clinician in the loop: autonomous chatbots, prescription software, passive monitoring framed as treatment. "Care delivery" means a human clinician treats through the product. "Augmentation" means the product strengthens work a human clinician is already doing.

The mortality gap is stark. Products that replaced the therapist died 53 percent of the time. Products with a human clinician in the loop died 27 percent of the time.

The obvious objection is that replacement products were sold to consumers, and consumer products die more often for reasons that have nothing to do with clinical architecture. The author checked. Inside the institutional-payer column, where an employer, insurer, or clinic pays the bill, replacement products still died at 50 percent while care delivery died at 9 percent. Same payer, more than five times the mortality. The failure belongs to the architecture, not the go-to-market.

The causes tell the same story. Among replacement companies, the leading recorded cause of death was failure to find product-market fit, followed by the regulator. The market either could not be convinced the product worked, or the product was prohibited from claiming it did.

Woebot closed the argument

If one case study captures the pattern, it is Woebot. A pioneering, rule-based CBT chatbot with roughly 1.5 million lifetime users and $124 million raised, it retired its app in June 2025. Its founder told STAT that the cost of meeting FDA marketing authorization requirements, combined with the absence of any regulatory pathway for large language model treatment, made the model unsustainable. Free general-purpose chatbots were doing a version of the same thing for nothing, and no payer had a reason to fund the clinical version.

That is the replacement trap in full: regulators will not authorize autonomous treatment, payers will not fund what is not authorized, and free tools undercut whatever remains. Funding does not escape it. The study found companies that raised over $100 million still died at 25 percent, and the replacement cohort collapsed regardless of what it raised.

We wrote about the clinical evidence behind this pattern in what 2026 research says about AI mental health chatbots. The graveyard data adds the business half of the argument: even where autonomous tools show symptom effects in trials, nobody durable pays for them.

What survives

The surviving architectures share one property: a human clinician stays in the loop, and the technology makes that clinician's work stronger.

Care delivery platforms, where licensed clinicians treat through the software, died least. Augmentation products, which support work between sessions under clinician oversight, sat well below replacement too. And the study's single strongest variable was not the product at all but the payer: products paid for by institutions died at 21 percent against 53 percent when an individual consumer paid.

Published clinical evidence also mattered, cutting mortality by ten points, and it turned out to be the asset that outlived companies. SilverCloud and MindBeacon were both acquired largely for their evidence bases.

None of this says technology has no place in mental health care. It says the durable place is underneath the clinician, not instead of the clinician.

What this means when you evaluate tools

Therapists now field products claiming to handle between-session support, documentation, and client engagement. The graveyard suggests the questions worth asking before you build workflow around any of them:

• Who actually pays for this product, and why? Consumer-subscription tools face 53 percent mortality. If the vendor's revenue depends on individual app subscriptions, plan for its disappearance.
• Does the product claim to treat autonomously, or to support treatment you deliver? The first category is the one regulators keep closing and payers keep declining.
• Is there a human accountability chain? Products that route risk to a clinician survive commercially for the same reason they perform better clinically: someone answerable is in the loop.
• Has the vendor published evidence? Not testimonials, published data. It predicts survival and it protects continuity even in the bad scenario.

This is the architecture we chose for Citt.ai: between-session support, screening, and documentation that run under a licensed therapist's supervision, with the clinician holding the treatment relationship. Our trust page describes the boundaries in plain terms, because the data says those boundaries are not a limitation of the model. They are why the model survives.