Healthcare has always had a timing problem.
By the time a clinician notices something is off—a patient drifting from their treatment plan, a safety risk hiding in months of session notes, a care gap that’s widened since discharge—the window for intervention has often already closed. The record captured what happened, and that’s that. Close the file, shut the door. It’s a bit like spanking your dog today for something that happened on the carpet last week.
That’s the environment our tools were built for: systems that document, systems that alert, and systems that report. Useful, yes, but fundamentally backward-looking. And that just won’t cut it anymore.
Agentic AI has the power to deeply change the direction for all of healthcare, and especially for community-based care. But it requires a careful, deliberate debate—one that includes the people who’ve lived in this work and want to change it. Since I’m rarely at a loss for opinions, I suppose that’s where I come in.
What “Agentic” Actually Means (and Why It’s Different)
There’s a version of this conversation that starts with definitions and ends with buzzwords. We’re not going to go there.
Most AI tools in healthcare today are reactive. They respond to inputs, generate outputs, and send up flares. A clinician does something, the system notices and responds. A great example is ambient AI: it captures and structures what humans did, and then it’s done. (It’s genuinely useful, but there’s more that AI can do.)
Agentic AI is different. These are systems that can perceive a situation, reason across multiple data sources, plan a sequence of steps, and act without waiting to be asked. They pursue goals autonomously, and they act on what they see instead of just reporting it. IBM puts it plainly: agentic AI can accomplish a specific goal with limited supervision, unlike traditional AI models that require human intervention.
Think of it like the difference between a dashboard in your car and adaptive cruise control. One tells you what’s happening. The other responds to the environment and alerts you to potential problems. Put in clinical terms, traditional AI is like a well-organized chart. Agentic AI is like a care coordinator who’s read the whole chart, is monitoring the patient in real time, and is already making calls before you’ve had a chance to pick up the phone.
Where Are We Right Now?
When it comes to agentic AI, we’re not as far along as the hype suggests. But healthcare organizations moving cautiously are actually closer to the frontier than they realize.
Across industries, Deloitte found that while 30% of organizations are exploring agentic AI and 38% are piloting it, only 11% are actively running it in production. Gartner projects that by 2028, 15% of day-to-day work decisions will be made autonomously through agentic AI—up from essentially zero in 2024.
In the most groundbreaking news I’ll deliver today, healthcare’s adoption of agentic AI lags behind general business applications. And as we all know, community-based care lags the rest of healthcare. It sounds like a criticism, but it’s not. It reflects the complexity of the data, the weight of the regulatory environment, and the very human stakes of getting it wrong.
The easiest early wins for agentic AI have been in revenue cycle and billing, where the data is already standardized and the business case doesn’t need much of a pitch. CPT codes are easier for bots to reason out than clinical narratives (less nuance, at least on the surface.)
The gap from pilot to production is closing. But across industries, the same pattern keeps happening: a fast, dramatic win, followed by a walkback toward a hybrid model.
Klarna, for example, deployed an AI assistant in early 2024 that handled two-thirds of all customer service interactions within its first month—2.3 million conversations in 35 languages across 23 markets. Resolution times dropped from 11 minutes to less than two. JPMorgan’s COIN platform reduced legal contract review from 360,000 attorney-hours annually to seconds of computational time. DoorDash built and deployed a production voice AI in eight weeks, and Lemonade’s AI claims system settled a claim in just two seconds.
And then the other shoe dropped. Klarna acknowledged quality degradation from over-relying on full automation and moved back toward human-AI balance. DoorDash routes complex cases to human agents. Lemonade’s AI auto-approves claims but never auto-rejects—human review is required for denials.
This is a preview of where the AI world is heading. Full automation might work in specific instances, but most mature deployments eventually land on a hybrid model. In community-based care, where the complexity and stakes are considerably higher than lunch delivery disputes, having a human in the loop is more important than ever.
What Agentic AI Could Actually Unlock in Community-Based Care
In the U.S., hospitals spend approximately 17% of their total expenses on administrative and general costs, and physicians spend around 13% of their work time on administrative tasks. Here’s the part we all know: Fragmented workflows, excessive documentation, and poorly integrated clinical systems increase burnout and the likelihood of clinical error.
That’s where agentic AI can have the most immediate impact. Paraphrasing a McKinsey report: AI agents can manage many of the complex workflows that bog staff down, involving people when necessary to clear roadblocks and ensure oversight. That frees clinicians to focus on what they actually came to do.
Consider a multi-agent architecture for the claims process, for example. You’d have a dedicated agent at each step of the process, including:
- Verifying insurance and pulling relevant codes
- Calculating payments
- Generating an Explanation of Benefits (EOB)
Clinicians would review for accuracy at each key decision point. They stay in the loop where their judgment is irreplaceable, and they get their time back everywhere else.
The Community-Based Care Opportunity—And Where It Gets Very Specific
The case for agentic AI in behavioral health and home health isn’t abstract. It maps to specific failure points in the field.
Longitudinal symptom surveillance
Most EHRs are static records. They capture what happened at discrete moments in time, but they don’t reason across those moments. An agentic system could continuously analyze the full clinical record and detect patterns no clinician could catch in a 50-minute session.
This hypothetical future would function as a two-tiered system. The first tier analyzes the existing EHR for trends, insights, and outliers. An agentic system might detect prodromal signals of a manic episode days before the patient or clinician notices, autonomously schedule a check-in call, alert the care team, and, if allowed, adjust the safety plan—all without a clinician initiating a request.
The second tier adds external data: Passive smartphone sensing (keystroke dynamics, GPS mobility, sleep/wake cycles, social communication patterns), wearables tracking heart rate variability and skin conductance—what researchers call “digital phenotyping”—layered on top of the EHR data. If you’ve ever attended one of my presentations on Disruptive Innovation (and stayed awake), you’ve heard me discuss this. It changes what prevention actually means. The locus of intervention shifts from crisis response to genuine prevention.
Agentic psychotherapy delivery and augmentation
Beyond chatbots that follow scripted CBT trees, true agentic systems can conduct adaptive therapeutic interactions that evolve based on moment-to-moment linguistic, paralinguistic, and behavioral signals. These multimodal agents would track session-level therapeutic alliance in real time and modulate tone, pacing, and intervention selection accordingly. They would also implement evidence-based protocols—CBT, DBT, ACT, motivational interviewing—dynamically rather than from fixed scripts, identify ruptures in therapeutic alliance and flag them for clinician intervention.
Early predecessors like Woebot point in this direction, but they lack true autonomy, goal-directedness across sessions, or the capacity to use multiple therapeutic modalities fluidly. What’s coming is qualitatively different, and it’s a particularly significant opportunity for rural and underserved communities where access to human therapists is severely constrained.
Autonomous care coordination
Handoffs can fail. Post-discharge follow-through is inconsistent. Medication adherence monitoring is sporadic at best. Agentic AI can own these workflows end-to-end. Post-discharge, an agent can autonomously initiate outreach, assess symptom status, coordinate between prescribers and therapists, and escalate to human intervention when needed. In a stepped care model, the agent assesses when a patient in a lower-intensity service tier is deteriorating and initiates the transition to a higher level of care—completing referrals, insurance pre-authorizations, and scheduling. For Certified Community Behavioral Health Clinics (CCBHCs), where comprehensive care coordination is a core requirement, this is a direct operational unlock.
Adherence prediction is another potential area where agentic AI can help clients achieve better outcomes. Adherence, or more precisely non-adherence, is a huge problem in healthcare. A 2017 Mckinsey report summarized the problem this way:
Doctors, healthcare providers, and caregivers are all too familiar with the problem: once patients are diagnosed and put on a treatment regimen, 50 to 60 percent are likely to skip medications, follow-up appointments, and other treatment protocols. The personal and economic costs of inadequate patient adherence are enormous. An estimated 125,000 unnecessary deaths occur each year in the United States because patients don’t follow their regimens. and the problem adds an estimated $290 billion to US health expenditures annually.
While much work has been done to identify why people don’t stick with their clinical regimens and what can be done to help them, AI provides a new tool to help address this problem in real time. Non-clinical behavioral signals can predict clinical outcomes, and an agentic system that scans for dropout probability upstream could enable clinicians to intervene before a patient disengages, rather than after.
Real-time craving intervention
Substance use disorder is particularly well-suited to agentic intervention because relapse risk is often situational, time-limited, and signaled by detectable behavioral cues. An agent operating here could help identify high-risk geofence locations, push real-time coping interventions, recognize conversational or voice-based signals of craving or dysregulation and initiate brief motivational interviewing protocols, coordinate peer recovery support specialist outreach autonomously when risk thresholds are crossed, and integrate with contingency management platforms to autonomously verify sobriety and administer reinforcements. It could also go a step further and identify subtle, individualized potential triggers specific to an individual.
Agentic support for group therapy
Groups rely on interpersonal dynamics as a core part of the therapeutic process, and that’s exactly where agentic AI has the most capability to contribute beyond documentation. An agent embedded in group work could analyze group dynamics in real time—participation equity, cohesion signals, emotional tone, therapeutic factor activation—with facilitative prompts to the clinician. It could synthesize each session into secure messages flagging which members may need individual follow-up. Between sessions, it could check in with members directly to reinforce session content and flag deterioration. And over time, it could track outcomes across cohorts with autonomous effectiveness reports. Tools like Eleos’s group product capture some of this signal, and agentic systems could act on it.
Clinical decision support that actually supports decisions
Today, EHR alerts are sophisticated if-then statements—passive flags that fire when a clinician does something. Agentic systems would continuously reason over a patient’s full clinical record and proactively surface and act on clinical intelligence. For example, an agentic system might order and track PHQ-9, GAD-7, and CSSRS screenings at appropriate intervals, identify diagnostic drift before the next scheduled review, recognize polypharmacy risks and medication non-adherence patterns without waiting for a prescriber to review a med list, and generate treatment plan revision recommendations with supporting rationale for clinician review.
There’s an overarching theme I’ve noticed: the clinical record is still built on a paper metaphor. Treatment plans are static documents created at a point in time, reviewed periodically, and filed away. Agentic AI makes that document dynamic. The care record becomes a living system that evolves with the patient, flags when it’s out of alignment, and acts when it needs to.
Safety planning as a living document
Traditional safety plans are static. They’re written in a session, filed in the chart, and often left untouched until the next crisis. Agentic AI can make safety planning dynamic by:
- Continuously updating risk estimates based on validated instruments, clinical notes, and behavioral signals
- Autonomously contacting the patient when risk signals elevate
- Conducting a brief structured assessment and escalating to crisis services with a synthesized risk summary
- Coordinating means restriction counseling (e.g., contacting a designated family member with psychoeducation about firearm storage, with consent, when a patient’s suicide risk score crosses a threshold
This directly addresses the gap between crisis planning and crisis detection that accounts for many preventable deaths.
The Real Risks Leaders Need to Think Through
Agentic AI’s potential is real, but so are its risks.
An uncharted regulatory environment to figure out
Agentic systems that take clinical action—not merely support it—are likely to be classified as Software as a Medical Device (SaMD), triggering FDA oversight pathways. The rules are still being written. This is uncharted territory, so organizations deploying clinical agentic AI need legal and compliance counsel who understands this evolving landscape.
Ownership of and liability for harm
When an agent acts and harm occurs, liability attribution between the developer, the deploying organization, and the supervising clinician is unresolved. This is a governance issue every organization needs to understand before deployment. Clinicians cannot abdicate their responsibility, nor should they. If they choose to use agentic AI, a realistic risk/reward analysis needs to be done, and risk mitigation strategies need to be employed.
Algorithmic biases and ethical concerns
Training data that reflects historical disparities in behavioral health diagnosis and treatment can propagate those inequities at scale. Organizations need to know what data their systems were trained on and, in partnership with their AI vendors, build ongoing monitoring for bias into their deployment governance. Lemonade’s AI ethics advisory program and published fairness commitments are good examples that behavioral health care leaders might consider.
Read more about how Eleos helps prevent bias in its models.
Crossing into surveillance territory
Digital phenotyping (e.g., passive smartphone sensing, behavioral pattern monitoring, geofencing, and real-time risk scoring) have an ominous “big brother” feel to them that some clients and clinicians may perceive as intrusive. These technologies all raise informed consent questions, and arguably ethical questions as well. The cost/benefit analysis includes health improvement and relapse prevention versus privacy. Is it worth it? Is it intrusive? These are the questions that will need to be answered.
Solving for problems that don’t need solving
Henry Ford said it best over 100 years ago with an observation Deloitte echoed in its 2025 Tech Trends report: “Many people are busy trying to find better ways of doing things that should not have to be done at all.” He was writing about manufacturing in 1922, but he could just as easily have been describing healthcare administration in 2025. It mirrors one of my personal favorite sayings, “The secret to success is figuring out what you’re not going to do”. The real value of agentic AI is in asking which workflows needed to exist at all.
The hybrid model wins
Every mature agentic deployment maintains human escalation pathways. In behavioral health, the ethical requirement for meaningful human oversight is the way we prevent problems and maintain the trust between clinician and client. The human-in-the-loop model is still preferred compared to totally autonomous AI agents. Will that change in the future? Will there come a time when those agents are good enough to act on their own? Probably. But that is not today’s reality.
What’s Next for Community Care?
Healthcare executives and providers don’t need to wait for agentic AI to become mainstream to start preparing. They need to understand what questions to ask.
Before buying anything, ask which of your workflows are actually candidates for agent-based redesign—and whether your clinical workflows and data architecture could support that. Agentic AI requires clean, structured, longitudinal data to reason over. Organizations with fragmented records, poor note quality, or siloed systems will get less value (or the wrong value) from agentic deployment. The quality and structure of the clinical record is foundational. EHR choice matters less than EHR implementation—how well your team actually uses the system you already have. A good starting point when considering agentic AIs is the same as considering a new EHR: workflow flowcharting. Documenting how information flows through your organization—who touches it, who needs it and where it’s stored—goes a long way to understanding where any piece of technology should fit, especially agentic AI.
The organizations best positioned aren’t necessarily the largest or the best-resourced—they’re the ones with the most consistent, structured clinical data and the clearest and best documented workflows.
The right governance question is this: who’s accountable when the AI acts, and how do we make sure the human-in-the-loop design is meaningful? Rubber-stamping AI outputs without genuine clinical review defeats the purpose of oversight and creates liability without protection.
The arc of agentic AI in behavioral health is clear: from documentation (what happened) to decision support (what should happen) to autonomous coordination (making it happen) to proactive surveillance (preventing the crisis from happening at all). The constraints are regulatory, ethical, and organizational. But the organizations that navigate them thoughtfully, build the right governance frameworks, and invest in the data infrastructure to support agentic deployment are going to deliver care that looks fundamentally different from what’s possible today.
Ready to get started? Download our checklist to learn how you can understand, evaluate, and confidently adopt AI tools at your community-based care organization.
