The role of agentic AI in clinical and administrative healthcare work

What an AI agent is and how it differs from chat-based use

In chat-based use, the model answers a single question and the user takes the next step. In an agentic system, the AI is given a goal, for example “prepare the file for this pre-authorization request for review”. The agent breaks the goal into steps, calls one of the tools defined for it at each step, assesses the result and decides on the next step accordingly.

Tools are defined functions the agent can call, such as pulling records from HBYS, the hospital information system, converting a scanned document to text with OCR, looking up a code in a rule table or searching the literature for a source. The agent can use only the tools opened to it. An orchestration layer manages the order of the steps, the division of work when several agents are involved, permissions and logs.

This is also where agents differ from classic process automation. A script that follows fixed steps often stalls on an unexpected document format, while an agent can choose which tool to use based on the input. That flexibility also makes it necessary to limit, and record, what the agent does and with what authority.

Clinical and administrative work where agents are used in healthcare

Agents add the most value in repetitive work that requires gathering information from more than one system and checking it against rules:

• Clinical document reading means converting scanned discharge summaries, reports and test results to text, and splitting diagnoses, procedures, medications and findings into structured fields.
• Pre-authorization file preparation means bringing together the policy coverage, the remaining limit, earlier pre-authorizations and the clinical documents, and flagging missing documents before the request is sent.
• Code and rule checks compare the ICD-10 diagnosis code with the diagnosis in the discharge summary, and the procedure with the reimbursement rules. In Türkiye the main rulebook for public reimbursement is the SUT, the Healthcare Implementation Communiqué published by SGK, the Social Security Institution; Medula is SGK's pre-authorization and billing system.
• Clinical summaries and literature support mean gathering a patient's encounters into a single summary and answering a physician's question such as “What should I not miss in this patient?” together with the document and source the answer relies on.
• Stock and indicator monitoring means scanning stock items against set thresholds and audit indicators against targets every day, and listing the ones at risk.
• Preliminary review at the insurer means bringing the policy terms, claim history and clinical documents into a single view for the specialist, and comparing earlier pre-authorizations and amounts when a claim is assessed.

In all of this work the agent prepares and checks; assessment and decision belong to the physician or specialist.

The limits of agents and the division of work with a rule engine

The language models behind agents are probabilistic. They may not give the same output for the same input every time, and they can present information that is not in the source as if it were certain. This is called hallucination. In agentic setups, a value misread at one step is also carried into the following steps. In work with clinical and financial consequences, these risks are managed through the following principles:

• Decision rules run in a rule engine. Reimbursement, policy and coding rules run in an explicitly defined, versioned rule engine, and the same input gives the same result every time. The agent prepares the data, calls the rule and explains the result.
• Irreversible actions wait for human approval. The agent can read, compare and draft; an approval, a denial, writing a code to the patient record or sending an invoice happens only with the approval of an authorized person.
• Every output carries a source and rationale. What the document says is shown separately from what the agent infers, and the person deciding can open and check the cited source; the document read, the tool called, the rule triggered and the person who made the decision are all logged.

This way routine checks run in the system while accountability stays with people. AI does not replace the physician or specialist; it gives them a more complete file. We describe this in detail in the blog post dividing the work between the rule engine, AI and people.

Steps for getting started with agents

When bringing an agentic system into a hospital or an insurer, these steps limit the risks from the start:

1. Start with a single, well-defined and frequently repeated process, such as the missing-document check at pre-authorization or the SUT check before billing.
2. Define in writing which tools and data the agent can access, and limit access to the data the task requires.
3. Keep decision rules separate from the agent, in a versioned and auditable rule engine.
4. Build a human approval point into the workflow for every irreversible action.
5. Require a source and rationale in every output, and keep a log of every step.
6. Keep the data inside the institution. Health data is special category personal data under KVKK, Türkiye's personal data protection law, and under the GDPR; run the system on the institution's own servers or in an isolated cloud environment. The details are in the guide on using AI without patient data leaving the institution.
7. Define success criteria up front and start with a POC on the institution's own data. The approach is described in the guide on starting a hospital AI project with a proof of concept.

Agentic orchestration on Opinion AI's platforms

Opinion AI is a clinical decision support platform for hospitals and health insurers, built on MINA, a clinical AI adapted to Turkish. MINA's open-weight base model has been adapted to Turkish and the clinical language of 14 specialties through continued pre-training (CPT), and specialized by specialty, institution and task with LoRA adapters. The third layer is agentic orchestration over a clinical knowledge graph (GraphRAG). At run time, MINA uses agents to move through the patient's digital twin and across tools, carrying out multi-step clinical tasks; specialty billing agents prepare SGK and SUT coding specialty by specialty.

At the hospital, TIS is added as an intelligent layer on top of the existing HBYS. The AI agent MINA suggests ICD-10 codes, checks SUT compliance and analyzes discharge summaries. Data from MKYS, the Ministry of Health's inventory management system, is scanned every day against 60- and 90-day thresholds, and BH and ADSH indicators are monitored continuously. At the insurer, SIT-I opens the policy coverage and the clinical documents on the same screen at pre-authorization, and the specialist can put a clinical question to the system through Doctor Assistant and Council Mode.

On both platforms, decision rules run in the rule engine; AI explains the result and its rationale, and shows what the document says separately from its own inference. People make the decision, and every output can be audited. Opinion AI does not collect personal data; the data stays inside the institution, the institution is the data controller, and personal data is masked in line with KVKK. The Physician Ethics and Advisory Board, with physicians from 15 specialties, regularly reviews the model's clinical accuracy and ethical boundaries.

Frequently asked questions

What is agentic AI?

Agentic AI refers to AI systems that break a given goal into steps, use defined tools such as document reading, record lookups or rule checks at each step, and choose the next step based on the intermediate results. In chat-based use the model answers a single question, whereas an agentic system carries out a multi-step task from start to finish; for example, it gathers and checks the documents in a pre-authorization file and prepares the file for review.

What are AI agents used for in healthcare?

In healthcare, agents are used for work such as reading and structuring clinical documents, preparing pre-authorization files, checking ICD-10 codes and reimbursement rules, flagging inconsistencies before billing and monitoring stock and audit indicators. In this work the agent prepares and checks; assessment and decision rest with the physician or specialist.

Can an AI agent make a clinical or pre-authorization decision on its own?

Not in a responsible setup. The agent reads documents, flags gaps and shows the rationale and source behind its recommendation; irreversible actions such as an approval, a denial, writing a code to the record or sending an invoice wait for the approval of an authorized person. On Opinion AI's platforms, decision rules run in a rule engine and the decision is always made by a physician or specialist.

What is the difference between agentic AI and a rule engine?

A rule engine runs explicitly defined rules and gives the same result for the same input every time, which makes it suitable as the basis for reimbursement and policy assessments. An agent reads free-text documents, gathers information from different systems and explains the result in plain language. In a sound architecture the agent prepares the data, the rule engine assesses it and a person makes the decision.

How can the risk of errors by AI agents be reduced?

Language models can present information that is not in the source as if it were certain, and in agentic setups an error at one step can carry into the following steps. The risk is reduced by keeping decision rules in a rule engine, requiring a source and rationale in every output and checking the cited source, showing what the document says separately from what the model infers, limiting the agent's access and logging every step.

Does using AI agents take patient data outside the institution?

It does not have to. Health data is special category personal data under KVKK, Türkiye's personal data protection law, and under the GDPR, so it matters that agents run on the institution's own servers or in an isolated cloud environment and that the institution remains the data controller. Opinion AI's platforms are set up this way, and Opinion AI does not collect personal data.

To assess with us where agentic AI could help in your own processes, using your institution's own data, use the POC Request form. For the criteria to apply when choosing a solution, see the guide on criteria for choosing a hospital clinical decision support system; for the difference from general-purpose tools, see the guide on how general-purpose chatbots differ from clinical decision support systems.