< + > Beyond Therapy Bots: Evaluating AI at the Front Door of Mental Health Care

The following is a guest article by Dylan Abrams, Lead Clinical Psychologist at InStride Health

Most conversations about AI in behavioral health focus on two areas: operational efficiency and treatment delivery.

On the operational side, AI is being used to reduce documentation burden, support administrative workflows, and make care teams more efficient. On the treatment side, there is growing attention on AI tools that interact directly with patients in sensitive contexts, including chatbots that provide mental health support.

Both areas matter. But there is another use case that deserves attention: AI at the front door of care.

In behavioral health, the admissions and referral process is not just an operational workflow. It is a clinical moment. Families may be scared, overwhelmed, and trying to understand whether a program can help. Referring providers are trying to determine whether a patient may be appropriate for a specific level or model of care. The information exchanged in this stage can shape expectations, trust, referral quality, and ultimately access to the right care.

That creates a different kind of evaluation need.

An agent supporting parents or referring providers is not delivering treatment. It is not diagnosing. It is not making final clinical fit decisions. But it can still create clinical and operational risk if it overstates who is appropriate for care, minimizes acuity or safety concerns, gives unsupported information about timelines, insurance, availability, or outcomes, sounds overly promotional in moments that require clinical nuance, or fails to route urgent or ambiguous situations appropriately.

Existing mental health AI evaluation frameworks have done important work in areas like safety, crisis response, empathy, and boundaries. Those domains are essential. For AI tools used in admissions, intake, or referral support, they need to be paired with evaluation criteria that reflect the realities of that workflow.

For AI at the front door of care, those criteria include:

• Clinical Safety and Escalation: Can the system identify urgent or higher-acuity situations and route them appropriately?
• Clinical Fit Reasoning Without Final Determination: Can it help clarify fit signals and uncertainty without acting as the decision-maker?
• Operational Truthfulness: Does it avoid hallucinated or unsupported claims about process, cost, timing, care model, or outcomes?
• Referral Integrity: Does it guide parents and providers toward the right next step, especially when cases are ambiguous?
• Trust-Preserving Communication: Does it respond with warmth and clarity without sounding scripted, dismissive, or sales-oriented?
• Operational Utility: Does it improve referral quality and reduce confusion without compromising clinical standards?

These are not just product questions; they are clinical governance questions. As AI becomes more ubiquitous as a behavioral health tool, these questions are increasingly important in considering how AI can support the admissions pipeline. The goal should not be to replace clinical evaluation or automate access decisions. The goal should be to build systems that make information clearer, escalation paths safer, and decision-making more consistent while keeping humans accountable for the calls that require clinical judgment.

A therapy-support chatbot, a documentation copilot, a clinical fit agent, and a referring-provider FAQ agent should not be evaluated as if they are doing the same job. They sit in different parts of the care journey. They carry different risks. They require different human oversight. 

The opportunity in behavioral health is not simply to use more AI. It is to be precise about where AI belongs, what role it is playing, and how we know it is performing safely. For admissions and referral workflows, that means building evaluation systems that account for both clinical safety and the operational reality of helping families find the right care.

Because the front door of care is still care.