League's Healthcare AI Agents Go Live With 99.72% Escalation Accuracy
League's AI Front Door and AI Care Companion are generally available on Forge, with a 99.72%-accurate escalation model, synthetic patient testing and interaction-level audit trails.

Why it matters
- League launched AI Front Door and AI Care Companion on November 16, 2026, with deployment possible in as little as 12 weeks.
- League's safety-critical escalation model reached 99.72% accuracy at 0.374-second latency, versus 95.36% and 1.121 seconds for Gemini 3.5 Flash Lite.
- A League and Healthcare Dive survey of 150 healthcare executives found 61% had deployed agentic AI but only 16% had reached broad deployment.
- The products draw on insights from more than 1.8 billion real healthcare interactions.
- Semantic Expansion plus the Synthetic Question Generator produced 3.9x more adequately answered questions and 42% better coverage than Gemini 3.5 Flash.
League announced on November 16, 2026 that two healthcare AI agent products, AI Front Door for payers and AI Care Companion for providers, are generally available on its Forge platform, with a deployment timeline the company puts at as little as 12 weeks. The launch lands at a moment when most healthcare organizations have deployed agentic AI in some form but few have scaled it: in a League and Healthcare Dive survey of 150 healthcare executives, 61% had deployed agentic AI, yet only 16% had reached broad deployment.
Both products pair member- or patient-facing agents with safety controls and human escalation. League says the products draw on insights from more than 1.8 billion real healthcare interactions, and they extend the Forge platform introduced on July 21, 2026. Forge supplies the underlying machinery: tools to connect agents to permitted data, set guardrails and escalation rules, test them against synthetic populations, and monitor completions, escalations and accuracy by use case. The July release built the governance system; the new products apply it to common payer and provider workflows.
What does each product actually do?
AI Front Door consolidates benefits navigation, care finding, care-gap outreach and health coaching into a single entry point, reachable through chat, text or in-app widgets. Responses about benefits, coverage and eligibility come from the plan's approved material rather than from the model's general knowledge. Its care-finding function matches members to in-network providers or services under plan rules. League frames lower call-center volume and fewer out-of-network visits as targeted outcomes — not as measured customer results.
AI Care Companion serves patients between visits, at home and after hours, alongside care and clinical teams. Health systems choose an initial population, such as people managing chronic conditions or recovering from surgery, and League configures the workflows around that group. The workflows can include:
- appointment scheduling and reminders;
- proactive symptom check-ins;
- answers to medication or treatment questions drawn from approved plans or pathways;
- protocol-based routing to specialists or on-call teams.
Designated caregivers can act on a patient's behalf, and their responses are tagged separately from the patient's own. League's stated targets for the product are better care-plan adherence and fewer avoidable emergency department visits, after-hours calls and missed follow-ups.
How does League keep the agents inside approved knowledge?
Both products confine agent responses to provided, approved knowledge instead of the model's general knowledge. Guardrails screen restricted information before it reaches the model and prohibited responses before they reach members or patients. Before release, agents must pass evaluations built on synthetic patient scenarios, edge cases and golden datasets, which serve as expert-authored answer keys. After release, the controls include drift detection, revalidation of changes and an audit record for each interaction. When a case escalates, the receiving team receives the full conversation history and relevant context.
The League Trust Center describes the platform's broader guarantees: read-only agent operation, strict session isolation, real-time detection of urgent issues and clinical misinformation, source traceability, and logging whenever guardrails activate. Customers can set escalation policies and sensitivity thresholds, but only within fixed safety boundaries. Protected health information is excluded from training pipelines; League uses synthetic data and de-identification techniques for model training instead.
These controls map directly onto what healthcare executives told League they need. In the League and Healthcare Dive survey, the leading requirements for confident launches were built-in healthcare compliance at 49%, automatic escalation to a person at 48%, and runtime controls over agent speech or actions at 46%. Of the 61% of executives who had deployed agentic AI, 70% were using it in member- or patient-facing applications — the exact setting where a wrong answer carries clinical and liability stakes.
What do the benchmarks show?
League says its hybrid architecture combines fine-tuned decision models and small language models with general-purpose large language models, with the aim of improving accuracy, reducing latency and lowering operating costs. One component, the safety-critical escalation model, identifies member and patient messages that require human support.
In League Labs' internal evaluations, that escalation model achieved 99.72% accuracy with 0.374-second latency. League compared it with Gemini 3.5 Flash Lite, which scored 95.36% accuracy and 1.121 seconds on the same task — a gap of more than four percentage points in accuracy and roughly three times the latency.
League also benchmarked two question-handling components against Gemini 3.5 Flash. Its Synthetic Question Generator, which anticipates likely questions about a user's health data, produced 2.2 times as many questions answered adequately or better, plus 11% better coverage. Adding Semantic Expansion, which rewrites long, ambiguous or shorthand-heavy questions before retrieval, raised those figures to 3.9 times as many questions answered adequately or better and 42% better coverage, at a cost of about 430 milliseconds of added latency per request.
League defines the answering measure as cases where retrieved health data is relevant and complete enough to answer fairly well or better; coverage measures how much of the necessary information is found. All results come from League's own internal evaluations, not from an independent benchmark.
Why the release matters for healthcare AI deployment
The practical significance for healthcare organizations is that oversight sits inside each interaction. Agents can be limited to approved material, tested against synthetic patients before release, handed off to humans with context, and reviewed afterward through an interaction-level audit trail. That operating model addresses, point for point, the safeguards executives cited in the survey as prerequisites for expanding member- and patient-facing AI.
The gap between the 61% of executives already running agentic AI and the 16% who have deployed it broadly shows where the market sits: pilots are common, scale is rare, and safety infrastructure is the stated bottleneck. League's bet is that confinement, evaluation and auditability — not raw model capability — are what convert pilots into production deployments.
Source: unite.ai
More from Rebecca Stone
Show full bio
Correspondent covering companies, deals and strategy in artificial intelligence and machine learning at AI In Context.
728 articles
Related articles
- Corti Brings Sovereign Healthcare AI to France With Scaleway and Dedalus
- Accenture and OpenAI Team Up to Put Agentic AI at the Core of Enterprise Operations
- Axway launches Amplify Platform to control AI costs and data risk
- Australian firm opens the books on five months of AI agent controls
- Virtana’s AI agent war room targets cross-system IT incidents