How We Built It: The Multi-Agent System Behind Complete Medical Records

The first retrieval decision is not simply identifying the provider. It is determining which action is most likely to produce the records. This realization came as part of capturing the workflow data. A non-trivial amount of time is spent on researching / learning the best or most successful pathway for a retrieval as the task is more difficult than identifying a provider.

Public and commercial provider directories can usually supply an address, specialty, NPI, or main phone number. They rarely contain the details that determine whether a records request succeeds: the medical-records department line, the monitored fax number, a separate stat pathway, the correct IVR sequence, or the contact who previously unblocked a stalled request.

Understanding the workflow data we needed to capture involved the recognition that historically, that knowledge lived with experienced operators. They searched provider websites, compared locations, called main lines, tested fax numbers, and recorded what worked. The strongest teams developed deep institutional memory, but much of it remained trapped in free-text notes or individual employees’ experience. The next operator often had to rediscover information the organization technically already knew.

Predoc built its research agent on a knowledge base of approximately 300,000–400,000 provider-research tasks. Each task generated evidence about which facility held the records, which phone or fax route worked, which attempts failed, what the provider required, and which action ultimately produced a return. Instead of leaving that knowledge in free-text notes or an experienced operator’s memory, Predoc structured it for reuse.

The results show the value of the approach. Predoc has reduced the time required to determine and prepare the retrieval path by 70% and improved first-pass retrieval success by nearly 50%. The Research Agent also identifies requests that do not require a phone call: Predoc now completes nearly 20% of retrievals through successful cold fax, with median turnaround of less than one business day.

The gains compound. Faster research reduces the direct work required before outreach. A better first decision avoids wrong calls, transfers, failed faxes, and repeated research. And the structured handoff ensures that the Voice AI Agent begins with the highest-confidence route and the context required to act on it.

This is what distinguishes the Research Agent from any existing provider directory. It does not merely surface information. It decides what should happen next, executes when it can, and gives the next capability enough context to continue the work.

When fax alone is insufficient, the workflow moves into a live interaction with the provider.

This has traditionally been a one-to-one process. One operator places one call, navigates the IVR, reaches the appropriate office, confirms patient information, explains the request, captures the facility’s requirements, and documents what should happen next.

The long tail is what makes the workflow difficult. An unfamiliar facility may require several transfers, confirmation that the patient was seen there, a referring provider, a facility-specific authorization process, or information the requester did not know would be required before the call began.

The Voice AI Agent was built from the workflow’s variability. Its proprietary harness was developed and tuned using nearly 3 years of transcribed retrieval calls. Those calls exposed the possible conversational turns, required inputs, facility-specific rules, completion criteria, failure conditions, and escalation triggers found in real provider outreach.

A foundation model provides the general conversational capability. The harness supplies the job. In this context, the harness is the workflow-specific layer of instructions, context, software, structured outputs, monitoring, and escalation controls surrounding the model. It defines what information the agent must provide, what the call must obtain, what constitutes a successful or useful partial result, and when the interaction should continue, stop, retry, or move to a human.

Without that dataset, a voice model would not reliably know what the call needed to accomplish and how to accomplish it.

Predoc estimates that the Voice AI Agent can reduce per-minute outreach costs by 75%–92% across common staffing models. The larger advantage is elasticity: calls can begin when requests are ready and run in parallel instead of accumulating in a backlog limited by the number of available human callers.

The model makes conversation possible. The proprietary data and harness turn that conversation into completed medical-record work.

In a legacy workflow, an indexer selects the packet from a queue and performs that work page by page. Manual indexation has historically required approximately two minutes per page. A 100-page packet can therefore consume more than three hours of active labor, before accounting for the time it spends waiting in a queue, moving between teams, or sitting untouched outside operating hours.

Predoc’s Indexation Agent begins processing when the records arrive. It associates the return with the patient and request, identifies document boundaries, classifies the contents, applies the required metadata, and determines which pages can proceed automatically.

Predoc has reduced the active time spent on indexation by nearly 50%. That is a deliberately narrow measure: it captures the hands-on work required to index the packet, but not the additional time eliminated when the packet no longer waits for an indexer to select it from a queue.

At the time of publication, 94.6% of medical-record pages received from providers are indexed without human intervention. People no longer need to inspect every routine page to find the minority that requires judgment. They work directly on the exceptions, allowing the same review capacity to support greater record volume without lowering the standard applied to uncertain content.

The operating model changes from requiring a person to inspect every page to directing people toward the minority of pages where judgment can change the result. The output is a patient-matched, segmented, and indexed record set ready for clinical extraction.

Predoc’s extraction layer converts document content into patient-level facts: medications, diagnoses, procedures, laboratory results, encounters, vital signs, allergies, and other relevant information. It also captures the dates, values, statuses, and source references required to interpret those facts correctly.

The extraction dataset is organized around the information that real medical records contain. Predoc processes each indexed record page by page and item by item, classifying the content into approximately 16 clinical categories and 10 non-clinical categories. Each category has its own schema defining the fields required downstream.

For a laboratory result, that schema may include the test name, value, unit, date, associated observation, patient, and source. For a clinical note, it may preserve the free text while identifying the note type, author or facility where available, and the relevant clinical information contained within it.

Each category has its own extraction schema. The system looks for the specific fields required to represent that type of clinical object downstream. The output is structured, patient-level data rather than an undifferentiated transcription of the PDF.

Ambiguity is treated as an exception, not an invitation to invent context. When the record does not contain a required attribute or the clinical meaning cannot be resolved confidently, that gap can remain explicit or move to review. The resolution can then inform the extraction schema or validation logic used on similar records.

Extraction identifies the facts. Curation determines what those facts mean together.

Raw clinical data reflects the systems and institutions that produced it, not a consistent view of the patient. The same medication, diagnosis, procedure, encounter, or laboratory result may appear repeatedly, use different terminology, carry different units, or lack the context required to support a reliable clinical interpretation.

The problem is not limited to faxed records. HIE data can arrive through standardized documents or machine-readable objects and still contain duplicates, empty entries, local naming conventions, inconsistent units, and multiple representations of the same event. Exchange standards make transmission possible. They do not automatically make the result trendable, queryable, or clinically coherent.

The reduction does not mean that two-thirds of the patient history was discarded. It means repeated, empty, incomplete, and semantically equivalent objects did not survive as separate clinical facts. Provenance-oriented categories were retained at 100%, preserving traceability while redundant clinical data was collapsed.

This is where the value of the earlier handoffs compounds. Research determines how to obtain the record. Voice AI moves the request forward. Indexation organizes the return. Extraction turns pages into facts. Curation makes those facts usable together.