Clinicians are ready for AI. Are the systems?

Summary

• Case managers are open to using AI to support patient management once someone explains what it does and who stays accountable.
• Remote monitoring improves outcomes only when the data goes to a named clinician who is responsible for acting on it.
• An AI agent carries out set steps such as check-ins and scheduling, and it leaves diagnosis, treatment, and discharge decisions to clinicians.
• Leaders have four decisions to make: how AI is governed, who owns the time between visits, how patients are handed off between organizations, and who gets access.

On September 17, I delivered a continuing-education webinar for the Case Management Society of America, From Symptom to Signal. The audience comprised nurses and social workers who manage transitions of care across U.S. hospitals, health plans, and community agencies, the majority of whom had little or no direct experience with AI in practice.

I continue to practice in hospital-at-home and emergency medicine, and so receive, weekly, the patient whose handover did not hold. The session was therefore not about technology for its own sake. Technology is the enabler, but the benefit has to accrue to the system around it as a whole: the patient, the caregiver, the case manager, the receiving team, and the organization that carries the risk. Judged by that standard, the session addressed the least observed interval in healthcare: the period between visits and the days following discharge.

What the session covered

The argument rested on three propositions

First, the interval between visits is where avoidable harm accumulates. Deterioration is typically identified at the emergency department or the next scheduled appointment, rather than on the day it begins. Case managers experience this as a working day spent reconstructing what happened over the weekend rather than acting on it.

Second, remote monitoring produces outcomes only when it is coupled to an accountable clinician with defined authority to act. Programs that collected patient-reported data without that operating model did not alter admissions or mortality. Programs that routed the same data to a staffed team, under explicit escalation rules, did. The technology was comparable, but the accountability structure was not.

Third, the terminology matters. A chatbot answers. A risk model predicts. An agent executes defined steps, such as structured check-ins, scheduling, transport, routing a threshold breach to a named clinician as assigned work within a response window, and holds no authority over clinical judgment. It does not diagnose, adjust treatment, or determine discharge. The single design principle that distinguishes effective programs is that escalation reaches a person, not a queue.

The session closed with the governance frameworks that already define acceptable practice for AI in clinical settings. This included the two risks that warrant explicit attention, automation bias and inequitable access, and five questions any organization should put to any vendor, League included:

• Which rules govern the system and whether the audit trail can be produced?
• Who validates the clinical content and their qualification to do so?
• What triggers a hand-off to a human? To whom and within what timeframe?
• How is performance assured for patients who do not speak English, and which reported results are published rather than self-reported?

What the room said

Clinicians are not afraid of AI

Case managers are not the constraint on adoption. Several already use AI for documentation, while others reported that patients and families are consulting consumer AI tools about their symptoms and sought guidance on how to direct that behavior safely. What the profession has lacked is a clinical account of what these systems do and where human accountability sits. Once provided, the hesitation receded.

Where hesitation persists, it is organizational. The workforce is prepared. The systems around it are not, and that is a leadership matter rather than a workforce one.

On the technology:

• “It has changed my mind about the use of AI in medical settings.”
• “Less anxiety using AI.”
• “Reminder that oversight is mine.”

On their organizations:

• “Dependent upon employer direction.”
• “Limited by system technology.”
• “Makes me realize how behind the times our agency is.”

Four decisions that rest with leadership

Read together, the responses identify four gaps. None is a product feature. Each is a decision for a board, a chief medical officer or a chief nursing officer, whether the organization is a health system carrying population risk or a health plan accountable for its members.

Decision one: Governance is a leadership act, not a procurement clause

Respondents wished to know “where my employer is in developing guardrails” and noted the absence of “case management AI standards of practice.” A profession-specific standard does not yet exist; the national frameworks do. The decision is to state which frameworks the organization adopts, who is accountable for clinical content and escalation thresholds, and how the audit trail is made visible to the clinicians whose licenses underwrite the care.

Decision two: The operating model behind the signal determines the outcome

One respondent identified the “importance of staffing models” in acting on alerts. The strategic question is not which system to procure but who owns the interval between visits, what authority they hold outside working hours, and within what period they must respond. A business case that costs the technology and omits the accountable clinician is incomplete.

Decision three: The seam between organizations is where patients are lost, and where plan and provider share an interest

Case management leaders consistently ask for warm handoffs rather than transmitted documents, visibility after discharge, protocols that receiving teams execute, and partners who adopt shared objectives. Each is a coordination problem across an organizational boundary: hospital to home, hospital to post-acute care, provider to plan. Value-based arrangements succeed or fail on that seam, and a plan and a provider observing the same discharged patient require the same assurance: that someone is watching, and that the next action is explicit.

Decision four: Access is the equity strategy, and it must be designed

Transport, language, literacy and caregiver capacity determine who is reached. Virtual care is effective and, unless deliberately designed otherwise, excludes the caregiver. A tool available only in English, or only through an application, widens the gap it was procured to close. The measure of success should be who was reached and what changed for them, not adoption counts.

Forge by League: Making the decisions enforceable

League’s position is that the four decisions above belong to the organization, and that the platform’s job is to make them enforceable rather than aspirational. Agents on Forge execute the defined steps of a care pathway: structured check-ins, scheduling, transport, medication follow-up, all under rules the organization’s own clinicians set. The agent does not hold clinical judgment; the escalation path is not optional. The guardrails are built into the platform rather than added afterwards. Reasoning is grounded in approved sources, so an agent cannot produce a clinical or coverage answer it was never given.

How Forge routes a threshold breach

• Agent runs the pathway. Structured check-in on the day it matters; asks why a dose or appointment was missed.
• Threshold set by your clinicians. Rules and content are approved by the organization, not the vendor.
• Breach becomes assigned work. Routed to a named clinician with context assembled and a response window.
• Audit trail, every time. Who was told, when, and what happened next, visible to the licensed clinician.

Put plainly: Forge was built so that League can answer the five questions above without hesitation, and so that the organizations using it can answer them to their boards.