At Becker's Behavioral Health Summit in Chicago this spring, the artificial intelligence (AI) conversation had turned a corner. Panel after panel described tools they'd actually implemented into production and clinicians who were happier for it. The mood was optimistic when a year earlier it was mostly hypothetical, as speaker panels and attendees opined that AI might someday trim their overhead and improve clinical outcomes.

Five months later at the American Psychology Association (APA) convention in Washington DC, we noticed the mood around AI had fractured. AI was still at the center of many sessions and exhibitor showcases, but the questions underneath moved from what's working to where the data goes and who answers when it doesn't.

Spring in Chicago: the theoretical benefits of AI played out

The spring conference season headline was adoption. Groups that spent last year piloting AI were running it in daily practice, and the payoff they pointed to had little to do with cost. They described happier clinicians, fewer evenings lost to documentation, and how a patient's story held together better as it moved through the practice along the continuum of care. One health system described how a referral gets taken by phone, summarized, faxed, and re-keyed, where a little of the patient's story falls out at every step, like a game of telephone. Used well, AI keeps the thread intact across those handoffs and enables the staff to understand the patient's full story leading to improved communication and treatment coordination. Used carelessly, without a tested prompt or a person checking the output, it just launders the errors faster.

The tension in the room at Becker's Behavioral Health Summit was money. Several panelists said plainly that the most useful AI tools shouldn't be expected to cut costs and that several AI pilots would need to shut down once funding ran out. For a field that runs on thin margins, there is tension between what helps practices that want to improve operations and cost. It also reframes the question. If AI's return is a clinician who stays because the work got more bearable, that is worth paying for on its own terms, and it only holds if you can trust the tool delivering it.

Fall in Washington: the hidden costs and questions cropped up

AI is now in the room of every conference session or booth poster the way telehealth was a decade ago. One ethics session we attended presented that 250 state bills on AI in healthcare were introduced across 47 states in 2025, which the presenters called the biggest change since telehealth.

We found the sharpest moment at APA 2026 was a fireside panel called "The Human Side of AI," where APA CEO Arthur Evans posed questions about mental-health safeguards of AI tools to senior leaders from the largest AI companies (Google, Microsoft, OpenAI). Most of the answers were what you'd expect from a communications team: we're working on it, we care about trust, we're protecting kids. However, the useful part was one problem a panelist named plainly: Safeguards built at the model level can block legitimate work at the product level, like an EHR that needs to read suicidal-risk language to do its job. Where you put the guardrail matters as much as whether you have one.

The institutions moved too. The APA Ethics Code revision, rewritten once AI adoption took off in 2025, now centers on informed consent, keeping a human in the loop, and basic AI literacy. APA's 2026 Chatbots and Mental Health Survey found that more than a third of psychologists have patients turning to AI to act as an additional mental health professional, and that 97% worry those tools can inadvertently reinforce negative behaviors or delusional beliefs. APA has been plain that AI is not a safe or effective replacement for a qualified provider.

Three questions every EHR should answer

The APA2026 sessions that landed were about trust more than features. In "Protecting the Therapeutic Alliance While Implementing AI," clinicians said they like AI note-scribes but proofread every line and worry about where the recording goes. The APA Ethics Code revision, rewritten once AI adoption took off in 2025, now centers on informed consent, keeping a human in the loop, and basic AI literacy. The request from clinicians was consistent: they want AI built into an EHR they already trust, not one more outside tool to vet.

One session reduced the clinical due diligence to three questions every EHR should be able to answer:

  • Where does the data go?
  • Can it be deleted?
  • How was the AI model trained?

A vendor who gets cagey on any of the three has answered the question for you.

The argument we made a year ago

None of this is new ground for us. A year ago our CEO, Melissa Tran, wrote in MedCity News that the behavioral health AI boom is real and so are the risks. Her green flags were human-in-the-loop models, clearly stated limits, clinical validation, and tools designed with providers rather than just for them. Her advice was to start small, prove value, then scale, and to treat a flashy demo with healthy skepticism. Read that piece next to this year's Ethics Code revision and the APA has turned her argument into a checklist. The 13-step guide the ethics committee walked through, covering vendor leadership, clinical evidence, HIPAA, data security, and whether your data trains the model, is skepticism made practical.

It's the same discipline we applied to our own numbers. Our 2026 mid-year report checked the loudest industry predictions against verified data, and several didn't survive the contact. The "denials are surging" story, for one, didn't hold. Hype is easy. Checking is harder, and it's the work.

What insight actually means

APA built the convention around insight, and by day two the word was on every banner and in half the vendor pitches, which is usually when a word starts to lose its meaning. The sessions gave it back some. Used honestly, insight isn't a feature or a model you bolt on. It's the discipline of looking at what the data actually says, keeping a person in the loop, and checking a confident claim before you act on it.

That has a practical edge for a practice. You can't lead with insight if your numbers live in six systems that don't talk to each other, or if you can't say where your data goes and how it gets used. Insight depends on connected, trustworthy data more than on the newest model. The clinics that seemed furthest ahead were the ones that could see their own work clearly and act on it, whatever tools they used.

If you want to see how we answer them, book a demo and ask us directly.