Woebot (a governed app wind-down)

Woebot was a smartphone chatbot app from Woebot Health that delivered self-help cognitive behavioral therapy (CBT). Its replies followed conversation paths its authors wrote in advance, so it could not invent a new clinical claim. It stored each user's conversations as transcripts, sensitive records that had to be dealt with when it shut down.

How it worked

Woebot Health, formerly Woebot Labs, of San Francisco, was founded by Alison Darcy. The app ran from 2017 to 2025. Woebot Health offered the app directly to the public and through employers and health systems.

The conversation felt natural, but each reply followed a decision path written in advance. Roughly 1.5 million people used Woebot over its lifetime. That is a cumulative figure reported in press coverage, not an audited count of active users at one time.

What the evidence showed

The foundational study in its peer-reviewed record was a 2017 randomized controlled trial in the journal JMIR Mental Health. In such a trial, people are assigned by chance to the app or to a comparison. Its authors were Fitzpatrick, Darcy, and Vierhile, Woebot's own people. It enrolled 70 people aged 18 to 28 for two weeks. The comparison group received an information-only self-help ebook.

The Woebot group's depression symptoms fell significantly on the PHQ-9 questionnaire. The authors called it a moderate difference between the groups, an effect size of 0.44. They concluded that a chatbot can be a feasible and engaging way to deliver CBT.

That is an efficacy signal, an early sign the app might help, not regulatory validation. The trial was small and short. It was unblinded, so participants knew which group they were in. Woebot's wider published evidence includes further studies not gathered here. What the sources lack is an independent, arm's-length evaluation of the consumer app.

The prescription version, WB001

WB001 was a separate, investigational version, available only on prescription. It was an eight-week smartphone treatment for postpartum depression, to be used under a clinician's supervision. It combined CBT with elements of interpersonal psychotherapy.

In May 2021 it received a Food and Drug Administration (FDA) Breakthrough Device Designation. That status speeds up review. It is not marketing authorization, the FDA's permission to sell a device. WB001 entered a pivotal trial as medical-device software, with the first patient enrolled in January 2023. A pivotal trial is the main study meant to support an FDA decision. This one ran at several sites and was double-blind, so neither patients nor researchers knew who got which version. It stayed investigational and never received marketing authorization. The consumer app that shut down is a different product.

How it was funded

In July 2021 Woebot Health closed a $90 million funding round, called a Series B, co-led by JAZZ Venture Partners and Temasek. That brought its total funding to $114 million, per the company's press release. A secondary industry summary calls the round a Series C and puts the total at about $107.5 million. The difference is noted here, not resolved.

The funding is context for the founder's later argument. She said the cost of authorization, not a lack of investment, ended the product.

How it was wound down

In April 2025 Woebot Health announced it would shut the app down. It did so on a published schedule, set out in its own FAQ. The app was retired on June 30, 2025. Users could request a transcript of their conversations until July 15, 2025. The FAQ said all account data would be anonymized as of July 31, 2025.

Anonymizing removed personally identifying information. The service was not quietly abandoned, and identifiable data was not kept indefinitely. After the retirement, no new accounts could be created, and earlier accounts could no longer be opened.

What each side says

The founder and chief executive, Alison Darcy, attributed the shutdown largely to the cost of meeting the FDA's requirements for marketing authorization. She also pointed to a mismatch between a fast-moving AI field and regulatory frameworks that lacked a clear pathway. She framed the exit as economic and regulatory, not a clinical failure. That is her own on-record account, not an independently audited finding.

Commentators read the exit as a signal about how AI in mental health is regulated. A clinically studied, transparently governed tool that did not generate text was shut down. Meanwhile, unregulated general-purpose chatbots drew far larger audiences without clinical validation, because they are not marketed as health tools.

What to watch

The danger here is not a runaway mistake by the chatbot. It is what a running service does with the sensitive records it built up, once the business reason to keep going runs out. It is also whether an early efficacy signal gets quietly read as validation on the way out.

Watch the transcript store, and watch what the exit claims about the evidence. This network follows the pattern the case file describes. It does not reconstruct the app, and the people who used it are not in it.

What the available tools can and cannot address

A failure pathway is a link between two parts of the network where a mistake made by one part can be passed on to the other. Here a mistake is, for example, a flaw in a pre-written reply. Closing a pathway means mistakes stop passing along it. The work along it goes on.

This case has a budget of 10 units. Each tool costs the same at every target level.

Explore (No Targets) sets no targets. Service Targets Only asks for two things. The network must be self-correcting, meaning its mistakes are corrected rather than building on each other. The chatbot must also be helping the work. Before any tool is used, the chatbot is helping. But the network is at a tipping point, on the edge between correcting its mistakes and letting them build up.

Two tools meet Service Targets Only on their own, for 2 units each. Escalate checks raises checking when monitoring flags trouble. Mark AI-written records labels what the chatbot wrote in the transcripts. Store less data does it at its stronger setting, for 5 units.

Lingering effects is a Lab setting in which mistakes and reliance on the system stay after their cause is gone. The Lab starts with it on. While it is on, Upgrade model also does it at its stronger setting, for 5 units. With lingering effects off, Store less data does it at its standard setting, for 3 units, and so does Vet connections, for 3. Upgrade model then does not. More than 150 different sets of tools within the budget meet this level.

Under Service and Safety Targets, the targets can be met within the budget. That level asks you to close every failure pathway, among other targets, and it runs with lingering effects off. Three are open before any tool is used. They are the pathways named Users' answers shape the script, Conversations saved per user, and Staff manage and anonymize transcripts.

The cheapest way costs 7 units. Store less data closes the two transcript pathways. Mark AI-written records closes the pathway named Users' answers shape the script, and Vet connections can close it instead. This level also asks the staff to keep up with their work. The tools that close those three pathways leave the staff behind their work. Adding Escalate checks or Review the riskiest first keeps them up.

That level can be met with 14 different sets of tools, and every one includes Store less data. Review on schedule is in none of them.

Under All Governance Targets, the targets are not fully addressable with the available tools. That level keeps lingering effects on. There, closing all three pathways takes Store less data at its stronger setting and Mark AI-written records, for 7 units. That leaves the staff falling behind. Adding Escalate checks or Review the riskiest first keeps them up. But the benefit the chatbot delivers then falls below what this level asks. Every combination within the budget was checked, and none meets every target.

This case does not offer Understand the system, the tool that lets you aim other tools. Without it, Store less data and Vet connections work at reduced strength while lingering effects are on.

More is not better here. Every tool at its strongest setting costs 34 units, more than three times the budget. It closes every failure pathway, but the chatbot then no longer helps the work, so it meets no level's targets.

No tool offered here adds the Arm's-length evaluation, the independent check of the app's evidence.