The rise of AI chatbots

While ideally every patient with mental health needs would have access to a high-quality mental health provider, we know that this is not the case—and likely won’t be in the future. Almost one-third (23%) of the U.S. population has a mental illness, but only half receive treatment. Similarly, 17% of the population has a substance use disorder but only a quarter receive treatment.

These persistent shortages of qualified clinical staff have pushed some organizations to consider alternative support tools less reliant on a one-on-one connection. In recent years, several companies have released AI-powered online chatbots capable of messaging back-and-forth with patients about their conditions and challenges. While most would claim these tools aren’t perfect, some health system leaders are convinced that an imperfect tool is better than nothing at all.

Even before the current wave of GenAI chatbots captured the public’s attention, AI chatbots were already attracting substantial interest from patients. A 2021 survey found that 22% of U.S. adults had already used a mental health chatbot, and 47% expressed interest. We’ve heard that some patients sometimes even prefer chatbots over other forms of human support and many form a close bond with them (many patients will, for instance, have a strong conviction about the gender of their chatbot).

These chatbots offer a variety of capabilities. In 2021, the U.K.’s National Health Service began using Wysa, an app-based chatbot capable of e-triage, waitlist management, and guided CBT exercises. The app has already achieved substantial scale: the company claims to have facilitated 500M conversations with 5M users across 95 countries, including 2M CBT sessions facilitated by AI. More recently, chatbots focused on mental health like Youper and Earkick have rolled out with GenAI features that take advantage of the latest LLM advances to generate more fluid and varied conversations.

While AI chatbots are a promising new direction for the industry, there are also potential pitfalls. These include:

• Some experts worry that misleading marketing claims lead patients to overestimate therapeutic benefits and underestimate technological limitations.

• Mental health professionals worry that chatbots won’t refer patients to more appropriate forms of human care in crisis situations: a recent study of 25 AI chatbots (all built on top of OpenAI’s ChatGPT model) found that many offerings are unable to quickly refer patients to human intervention or suicide hotlines.

• There is some fear that patients might try mental health apps, find them lacking, and then not pursue further treatment.

• Finally, many mental health apps have poor data privacy practices and share sensitive health information with advertisers (many of these companies are not subject to HIPAA).

A case study with Virtua Health

To learn more about how health systems might consider deploying these tools, we interviewed Virtua Health’s senior vice president and chief digital officer, Dr. Tarun Kapoor, to discuss the system’s ongoing pilot with an AI chatbot app called Woebot.

As part of that partnership, Virtua’s primary care physicians are recommending the app to adult patients with mild-to-moderate signs of depression or anxiety. Doctors ‘digiscribe’ the suggestion, ordering it in Epic, which generates a free activation code for patients. After patients log on, Woebot uses rules-based AI to guide patients through conversational exercises using concepts from cognitive behavioral therapy (CBT), interpersonal psychotherapy (IPT), and dialectical behavior therapy (DBT). The company also offers a technology platform that integrates with the EMR and helps clinicians track usage.

Here are four things we learned from our interview:

1. Virtua chose Woebot Health over other competitors because of their willingness to work on an intensive pilot.

Virtua is Woebot Health’s first public health system partner and was appealing as they expressed a willingness early on to work closely with Virtua to design and implement their pilot.

Virtua was also impressed by the extensive research that Woebot Health has produced (18 studies with more than 2,000 participants, according to Woebot Health’s Chief Commercial Officer Brad Gescheider) and more generally their company culture of “meeting patients where they are.” This is even evidenced by its name. Woebot’s somewhat glum name is based on market research suggesting that patients find mental health tools with cheerful names off-putting.

2. Woebot is popular among patients, but Virtua’s clinicians were initially hesitant to recommend it.

Among the 277 patients initially referred for Woebot in Virtua’s pilot, 109 patients signed up for the service (an impressive 39% activation rate). Since beginning the pilot earlier this year, patients have completed 862 sessions with Woebot (average session length of 6.8 minutes) and 94% of users have given the app a positive rating. Kapoor also noted that the most popular window for Woebot sessions was between 2am and 5am—outside normal working hours for human therapists when people often need the help the most. Seventy-seven percent of all engagement happened outside of office hours.

However, Virtua’s primary care physicians were initially reluctant to recommend the chatbot because of questions about adequate follow-up and liability. In particular, physicians were worried about an item in the chatbot’s patient questionnaire that asks users if they are experiencing thoughts of self-harm. Ultimately, system leaders worked to convince doctors that the chatbot was an appropriate addition to the system’s broader ecosystem of mental health support.

3. Virtua is hoping to negotiate FFS codes with payers to get paid for Woebot’s support.

While Virtua is finalizing data analysis now on the impact of Woebot on patients’ mental health, the initial data looks very promising. The primary outcome that Virtua is measuring in the pilot is improvements in PHQ scores (as these put a numerical rating to users’ self-assessed depression and anxiety symptoms).

Virtua hopes to convince payers to cover Woebot with a new dedicated FFS code by demonstrating that the tool helps users reduce their mental health symptoms before potentially developing a more acute diagnosis that could require more intensive interventions. According to Kapoor, at least one insurer has said informally that they would be willing to offer such an arrangement if the pilot can demonstrate this impact, but this arrangement has not yet been negotiated into a contract.

4. Woebot Health’s traditional approach to AI is more predictable than newer GenAI products.

Woebot Health’s chatbot was built using older approaches to AI that uses predetermined rules and patterns to deliver pre-written responses. As part of their market research, Virtua considered several alternatives including some chatbots that use GenAI features. Ultimately, they concluded that LLM chatbots aren’t ready for primetime patient-facing work because of their tendency to “hallucinate” and produce inaccurate or unpredictable responses.

Woebot Health leaders told us that they’re currently studying how GenAI can be applied to mental health in a clinical setting with an IRB-approved study. Ultimately, they’re seeking “data to show that the risk benefit ratio is appropriate for patient and health outcomes.”

So What for Health Systems?

Given the worsening state of Americans’ mental health and severe access shortages, behavioral health continues to be a major sticking point for most health system strategists. Considering the full market landscape we’ve mapped out in Part I and Part II above, here are some considerations for health system strategists trying to decide how their organization fits into the bigger picture on virtual behavioral health:

1. With behavioral specialists in short supply, primary care doctors are picking up much of the slack.

As we look at the landscape of strategies health systems are adopting to scale access to behavioral healthcare, we note that most are integrating it into primary care. This is not new: primary care physicians handled nearly 40% of all visits for depression or anxiety pre-Covid, and prescribe 50%+ of anti-anxiety and depression meds nationwide. But given that access to dedicated mental health professionals is so limited, many believe an integrated behavioral health approach is the only way to meet long-term workforce challenges. This integration can exist on a spectrum from coordination (e.g., including PHQ-9s in all patient intakes) and consultation all the way to system-level integration such as the model Intermountain has pursued. We explored how Intermountain has scaled this integration in detail last November.

However, there are limitations to this approach: while primary care doctors can prescribe medication for patients with immediate needs, they have fewer options for patients with low acuity conditions who need non-pharmaceutical interventions such as talk therapy. In addition, payer “behavioral health carve-outs” often mean primary care claims for mental health conditions are denied.

2. Many systems are tapping new partners as they seek to expand scale.

As discussed above, some systems are partnering with companies like Woebot Health to stretch their primary care resources with mental health apps. Similarly, Intermountain’s value-based payment subsidiary Castell announced a new partnership in August with Headspace to provide on-demand services for its health plan customers. Patients have access to on-demand modules as well as behavioral health coaching and therapy through an outcomes-based payment model. Along similar lines, Iris Telehealth has partnered with Geisinger and Allina Health to offer virtual therapy and psychiatry services.

While these tools are arguably better than nothing, they’re clearly best used to augment existing support ecosystems and shouldn’t be used as a wholesale therapist replacement. Indeed, these digital health solutions are often only meeting the needs of lower-acuity patients, not those that might need in-person support.

Therefore, some are looking for partners offering higher-acuity support. Concert Health, a virtual behavioral medical group with approximately 200 providers, has struck partnerships under a collaborative care model with Mass General Brigham, CommonSpirit Health, WellSpan Health, Mercy, and AdventHealth. Concert’s model enables systems to expand access to high-quality virtual behavioral care fully integrated with the EMR and billing platforms, although it requires a significant degree of delegation and trust.

Some systems have also tried to build up their own capabilities. Intermountain has touted savings from its in-house virtual care service, Connect Care, which includes behavioral health care. Hackensack Meridian Health has also made a concerted effort to expand de novo behavioral care access, including converting existing urgent care centers to include behavioral health supports with telepsychiatry capabilities that treated nearly 2K patients in 2022. Other systems with in-house virtual behavioral programs include UCHealth, OSF HealthCare, and Ascension.

A slide from Strategy Catalyst’s research in 2022 on virtual behavioral health programs

3. Employee needs and direct-to-employer bundles are also driving new approaches.

Many of the partnerships outlined above are being driven by systems with some degree of risk (e.g., Intermountain, Geisinger) that realize that care costs can be 3x higher for those with unaddressed mental health needs. In a similar vein, many systems are realizing that they also need solutions for their own employees. For example, after implementing its own integrated behavioral health program using the risk management platform NeuroFlow, Jefferson Health reported that 80-90% of their patients were engaging with the platform—leading to a 34% decrease in ED utilization for patients compared to a control group. But only 10-15% of Jefferson’s own employees were using the tool, despite having full access (in conjunction with Marvin, a behavioral health app for physicians). The system realized they needed a new approach for employees. Therefore, they allowed clinicians to share the resources peer-to-peer to get deeper utilization than would be possible just an email from HR.

As systems well know, employers have been skeptical of point wellness solutions in part because of low utilization—leaving them on the hook for benefits that workers aren’t even using. That’s why some systems are betting that care bundles are more attractive offering because they’re often linked to specific outcomes and offer a more comprehensive suite of services. Northwell Health’s employer bundle offering, Northwell Direct, has a DTE bundle includes a variety of mental health services (including a clinical call line, stress management sessions, coaching, and stress first aid).

4. Disruptors are setting consumer expectations with respect to technology, access, and pricing.

Consumers using apps like Talkspace and BetterHelp may grow accustomed to certain features like the ability to easily switch providers and have flexible scheduling, and they could take those expectations with them when they seek out traditional providers for more intensive needs.

Likewise, some of these services (e.g. ADHD treatment providers) are designed to make getting a prescription as fast and easy as possible. This has been taken too far in some cases. Cerebral and Done in particular have earned reputations as aggressive prescribers and have faced a lawsuit and federal investigations over allegedly unsafe prescribing practices. Given how easy they have made it to get protected medications, patients seeking these medications from in-person providers may be surprised when their doctor wants a more in-depth evaluation or recommends non-pharmaceutical treatments.

One area where these disruptors generally fall short is integration with the rest of the healthcare system, including primary care. This fragmentation can cause headaches for health systems (for example, if a new prescription is left off a patient’s record because they received it from a standalone virtual provider). Health systems looking to compete with their own virtual offerings can differentiate themselves in marketing by stressing the importance of connecting behavioral care to the rest of the care ecosystem.

5. Keep an eye on who virtual mental health providers are targeting.

After reaching the limits of growth in the D2C and commercial markets, it seems like many “disruptor” mental health companies like Talkspace and Brightside are expanding into Medicare and Medicaid. We find this surprising, given that margins are lower in these markets and we know that patients who need more comprehensive behavioral support often prefer hybrid offerings with in-person components.

For health systems, this shift in focus may make them less of a threat. Health systems are generally more concerned about volumes of younger and commercially insured consumers as they seek to build relationships with these patients. While these patients are not necessarily big spenders in the short term, creating a sense of loyalty and familiarity earlier can help systems continue to earn their business as they age.

What could be more concerning, is if all of the disruptors targeting commercial patients band together. Disruptors seem to be increasingly trying to build networks with other specialist providers to provide a full scope of support for commercial patients. In April, Talkspace launched its Behavioral Health Consortium, a new curated network of specialist providers for high-acuity care including treatment for substance use and eating disorders.

Therefore, while health system strategists might be tempted to dismiss the threat posed by disruptors in this market because they are point solutions with limited capabilities, if these companies can set up care approaches that connect their offerings into a full network, they could begin to pose a real competitive threat. For now, we think this is something that health system strategist should merely keep an eye on—these efforts are fairly limited in their current form, and we’re convinced that the most complex patients driving the bulk of system revenue will continue to need hybrid arrangements with in-person services that these companies do not offer.