Why AI Often Fails to Deliver Value for Health Systems and What It Takes to Reach Production

Today, Cleveland Clinic published a press release detailing our partnership. It covers the first use case, automating fax referral intake, and includes quotes from Cleveland Clinic's Chief Digital Officer and Luminai's CEO. Read it here.

Mack Lu

Product Team

The pattern is familiar to health system leaders. The AI demo looks impressive. The ROI story checks out. The case for deploying the solution seems straightforward.

Then the project stalls. The pilot runs for six months without reaching production. Or it goes live for one step of the process but fails to deliver the end-to-end outcome. Sometimes, it creates more work for staff than it was meant to eliminate.

These projects rarely stall because the model cannot perform the task shown in the demo. They stall because production introduces a different level of operational complexity, usually in four areas:

• Real workflows span multiple systems and teams. An order arrives in one system. An operator interprets it and enters the information into another. It may be routed to a specific specialty or facility, then managed by a different team in a separate downstream system. The demo shows a slick automation of one step. Realizing the ROI requires the entire process to work.
• Production data is messy. Much of healthcare’s operational data is unstructured: documents with handwriting in the margins, inconsistent forms, hundred-page contracts with clauses and amendments that reference one another, duplicate patient records, ambiguous fields, and multiple sources of truth. The demo assumes clean—or at least structured—data. The gap between that assumption and the operational reality is where AI pilots stall.
• Every health system is unique. EHR configurations, service lines, facilities, internal rules, and escalation paths vary by organization. Decision logic may span several standard operating procedure documents (SOPs) or exist only in the experience of a supervisor. The solution has to adapt to each institution and workflow. Many vendors do not invest enough to capture that institutional knowledge, instead asking health systems to change their processes to accommodate a generic product that make them more like everyone else.
• The last mile is hard. Exceptions surface that no one anticipates. Issues in upstream and downstream systems disrupt the workflow. Rollouts slow because of competing priorities and stakeholder coordination gaps. Operators cannot understand why a decision was made, so they stop trusting the system. SOPs and rules change after launch, leaving the automation out of sync with the operation.

These are not edge cases. They are the core operating conditions AI must handle to deliver value in production.

Expand

The Luminai Platform

We built Luminai to address these problems. Luminai provides an AI platform for health system operations. The platform has three connected layers that carry complex workflows from unstructured inputs through decision-making and execution:

Expand

Integration Layer

The Luminai platform is interoperable by design. Our Integration Layer includes connectors for the systems commonly used across healthcare, including EHRs such as Epic; data platforms such as Snowflake and Databricks; fax, document-management, and revenue-cycle systems; and healthcare standards including HL7 and FHIR.

Luminai orchestrates work across systems rather than treating system boundaries as a reason to narrow the workflow. Where usable APIs don’t exist, including payer portals and legacy applications, our browser and computer use agents complete the work directly in the interface. This enables Luminai to operate within the real constraints of health system technology environments.

On top of the connectors sit healthcare-tuned document models that read, extract and structure the inputs that often break template-based extraction: handwriting, cramped and inconsistent forms, and documents containing multiple patients or multiple orders. No document templates required.

Intelligence Layer

The structured records from Luminai’s Integration Layer enter a health system-specific graph layer that combines clinical and administrative reference data, including ICD-10, CPT, or payer rules, with the institution's own SOPs and tribal knowledge.

The Intelligence Layer can learn and identify recurring decision patterns in historical operating data, including logic the organization has never formally documented. Those patterns are reviewed with operational leaders before being incorporated into the workflow. Validated corrections provide structured feedback that helps the system improve over time.

In one customer deployment, our custom models:

• Classify inbound documents by type and route them to the appropriate department based on departmental SOPs and historical operating data
• Separate documents containing multiple patients into individual patient records
• Turn documents containing multiple orders into discrete orders for follow-up
• Prioritize orders according to urgency and institution-specific guidance, ensuring that high-priority cases are addressed and escalated appropriately

This approach achieved accuracy comparable to human review in a fraction of the implementation time of prior AI efforts. It also outperformed generic out-of-the-box models on both performance and cost.

Two capabilities help keep that performance and accuracy durable. First, the models run on a traversable graph of institutional decision logic, with a learning loop that improves performance over time while monitoring for bias and performance drift. Second, the logic behind decisions is explainable: operators can trace an output back through the rules applied and the underlying facts in the graph. When a decision appears incorrect, teams can see why it was made and determine what needs to change.

Execution Layer

The Execution Layer runs the workflow. It updates systems of record, routes work between applications, initiates downstream processes, and coordinates the steps required to reach the intended operational outcome.

For steps that require human judgment, Luminai provides purpose-built human-in-the-loop review tools. Operators can review work in real time through an interface designed around the specific decision they need to make, with the relevant information and context already assembled.

Control is granular. Any part of the workflow can run automatically or be routed for manual review, with settings that can be configured by step, department, decision type, or confidence level. Teams typically begin with human review of every decision, then expand automation step by step as performance and trust are established.

Governed automation becomes something the health system expands deliberately rather than something it hopes will work. As workflows run, detailed reporting and alerting provide visibility into accuracy, processing time, exceptions, and other agreed-upon performance measures. Teams can verify that automations remain above defined confidence and performance thresholds and intervene when necessary.

Luminai workflows run on enterprise-grade, HIPAA-compliant infrastructure with over 99.9% reliability. Protected health information is handled under dedicated security controls, including encryption, scoped access, and comprehensive audit logging.

Luminai FDEs Extend the Platform and Build a Compounding Foundation

The Luminai platform provides the foundation to integrate across complex health system environments, connect context across systems, and capture the decision history behind each workflow. That intelligence can carry across the health system—not just within a single use case—while the Execution Layer turns it into real operational outcomes.

But durable AI transformation in health systems requires more than technology alone.

At Luminai, forward-deployed engineering is not just an implementation layer wrapped around a fixed product. Our FDEs have the tools and mandate to configure and extend every layer of the solution to accelerate value delivery within the health system’s unique environment.

They can build new integration paths, translate institution-specific procedures into decision logic, configure models, design operator review experiences, and adapt how work is executed across systems. Because they work across the full platform, they can solve for the entire workflow rather than handing gaps between product, engineering, and implementation teams back to the customer as unsolvable edge cases.

The FDE role is most intensive while the workflow is being understood, built, and brought into production. As the workflow matures, the team turns what it has learned into durable configurations, controls, monitoring, and operating processes that fit within the health system’s existing structures. The health system can operate and govern the established workflow over time, while Luminai’s FDEs remain available for material changes, platform extensions, and new workflows.

Each deployment does more than automate a workflow. It strengthens a shared foundation of integrations, institutional context, and decision logic that makes the next workflow faster to deploy and easier to operate at scale.

In Sum

Luminai’s platform captures how each health system works, connects the systems involved, and carries complex workflows through to completion and ROI realization. Combined with our FDE model, it gives health systems a way out of stalled pilots and into production, with each workflow making the next easier to deploy.