Blog
October 21, 2026
par
Andrej Evtimov
Four layers in the claims stack, and where each one starts and stops.
Agentic AI is the newest layer in a claims stack that already runs a claims system, RPA and document AI. This post shows where each layer starts and stops, how the layers hand off work, and how to test any tool sold as agentic.
Agentic AI claims tools join a stack that already works
Agentic AI claims tools are the newest layer offered to claims departments. Claims leaders need a plain answer to one question. What does this layer do that the claims system, RPA bots and document AI do not?
AI use in claims is not new. The NAIC's survey of larger private passenger auto insurers covered 2021 data. Of the 193 companies that responded, 88% used, planned to use or planned to explore AI or machine learning. Claims was the most common area of use.
The label is also loose. In June 2025, Gartner predicted that over 40% of agentic AI projects will be canceled by the end of 2027. It warned of "agent washing": vendors relabeling AI assistants, RPA tools and chatbots as agentic. Gartner also estimated that only about 130 of the thousands of agentic AI vendors are real.
This post maps the four layers of a modern claims stack. It shows where each layer starts and stops, and how the layers hand off work. It also sets out eight checks for any tool that calls itself agentic.
Four layers of claims automation software
Claims automation software comes in four layers, and most claims operations run several of them today. Each layer solves a different problem. Trouble starts when one layer is asked to do another layer's job.
The claims system of record
The claims system is the official record of the claim. It holds policy and coverage data, reserves, payments, diaries and notes. It enforces authority limits and keeps the audit trail that regulators and auditors read.
It is built for structured data: fields, codes and transactions. Most claims systems store a long medical file, for example one of 400 pages, as attachments. The system keeps the pages but is not built to read them.
Robotic process automation (RPA)
RPA uses software "bots" to repeat the steps a person takes on a screen. Gartner defines RPA as software that automates tasks using scripts that emulate human interaction with the application UI. UI means user interface: the screens and buttons a person clicks.
In claims, RPA keys data from one system into another, opens diaries, sends standard letters and files attachments. It is fast and predictable when the inputs are clean. It breaks when a screen changes or a document does not match the rule. It does not understand what it moves.
Document AI (intelligent document processing)
Document AI, often called intelligent document processing (IDP), turns documents into data. In a claim file, it sorts pages into types such as medical records, bills and police reports. It also pulls fields like dates of service, provider names, billing codes and totals.
Document AI works one page or one document at a time. It reports what a page says. It does not explain what the whole file means.
Agentic AI
An AI agent works toward a goal rather than following a fixed script. The OECD describes AI agents as systems that perceive and act on their environment with some autonomy, using tools to reach specific goals. It uses "agentic AI" for systems in which several agents work together, split up tasks and pursue complex goals.
In a bodily injury (BI) file, a goal might be to build the treatment history and flag conflicts with the claimed injury. The agent plans the steps and reads across hundreds of pages. It checks one record against another and returns its findings. In a well-designed claims tool, each finding links to the page it came from. The adjuster reviews that evidence and decides.
Where each layer starts and stops
The table sets the four layers side by side. The last two columns show where each layer stops and who takes over.
- Claims system of record. Works on: Structured fields and transactions. Best at: Reserves, payments, authority limits, audit trail. Stops at: Reading the documents in the file. Hands off to: The adjuster, and other layers through APIs.
- RPA. Works on: Screens and structured data. Best at: Repeating fixed, rule-based steps. Stops at: Exceptions, changed screens, anything needing judgment. Hands off to: The claims system, or an exception queue.
- Document AI (IDP). Works on: Single documents and pages. Best at: Sorting pages and pulling out fields. Stops at: Meaning across documents. Hands off to: Agentic AI or the claims system.
- Agentic AI. Works on: The whole claim file. Best at: Multi-step checks across documents, with source-linked findings. Stops at: Decisions on coverage, liability, reserves and settlement. Hands off to: The adjuster, who decides and records it in the claims system.
Each layer passes work to another layer or to a person. No layer takes over the adjuster's decision.
How the hand-offs work on a bodily injury file
Take a hypothetical commercial auto BI claim. Six months after the loss, a demand package arrives with 600 pages of medical records, bills and imaging reports. The work moves through the layers in five steps.
- Intake. The claims system logs the demand, starts the diary and records the response deadline. An RPA bot may file the attachments and update a status field.
- Sorting and extraction. Document AI sorts the pages, removes duplicates and puts them in date order. It pulls out dates of service, providers, diagnosis codes and billed amounts.
- Checks across the file. Agentic AI takes that output and works toward set goals. It builds the treatment timeline and compares it with pre-loss records. It flags gaps in care, conflicts between providers and diagnoses that predate the loss.
- Review and decision. The adjuster opens each flag and checks the cited page. The adjuster then sets or adjusts the reserve, requests missing records or refers the file to counsel.
- Record. The decision and its reasons go back into the claims system. That keeps one official record of what was decided and why.
Two design choices make these hand-offs hold up. First, agentic AI should read the document AI output and also go back to the original pages. That way, an extraction error does not turn into a finding. Second, results should return to the claims system through APIs, not by copying a user's clicks. An API (application programming interface) is a direct, coded link between two systems. Screen-level work is RPA's job, and it fails when the screen changes.
A finding is only useful if the adjuster can check it quickly. A separate guide covers how to check a claims AI answer against the source page.
Insurance claims AI still leaves the decision with the adjuster
Agentic does not mean autonomous decisions. In 2026 interviews with 25 organizations, the OECD found that they generally constrain agentic systems rather than grant unrestricted autonomy. Claims is a strong case for that approach. Coverage, liability, reserves and settlement offers carry legal and financial consequences. They belong to a trained professional.
The NAIC bulletin sets governance expectations for insurers' AI use. The NAIC adopted its model bulletin on the use of AI systems by insurers in December 2023. It expects a written AI systems program and oversight of AI bought from third parties. Its control factors include "the extent to which humans are involved in the final decision-making process." As of 31 August 2026, 26 jurisdictions had adopted the bulletin. Each state's own text governs.
In practice, a sound insurance claims AI setup draws three lines:
- The agent can read, sort, compare and draft.
- The agent cannot pay, deny, set a reserve or write to a claimant on its own.
- Every finding the adjuster relies on can be traced to a page in the file.
How to tell if a tool is actually agentic
Gartner's warning about agent washing applies to claims. A workflow tool with a chat window is not an agent. Neither is a summarizer that reads one document at a time. Claims teams can run eight checks on any vendor, in a live demonstration on one of their own closed files.
- Goal or script? Given a goal such as "list every prior treatment for the lower back," does it plan the steps? Or does it follow a flow set up in advance?
- Whole file or one document? For a question that needs three different documents, does it combine them or answer from one?
- Cross-checks. Does it compare records and flag conflicts, such as a pain score that does not match the visit notes?
- Source links. Does every finding link to the exact page, and can a reviewer open it in one click?
- Honest gaps. Asked about a record that is not in the file, does it say the record is missing? Or does it produce an answer anyway?
- Visible steps. Can a reviewer see the steps the agent took and the documents it read?
- Set limits. What can it do without a person, and what needs approval? Can it change data in the claims system?
- Hand-offs. Does it receive documents and return results through APIs, screen automation or manual upload?
A tool that does poorly on checks 1 to 3 is probably document AI or RPA with a new label. That tool can still be useful, and it should be bought and priced as what it is. Teams weighing an in-house build can read what carriers learn when they build claims AI in-house.
AI claims processing works best in layers
AI claims processing is not one product but a set of layers, each with a defined job and a defined stopping point.
Tools built for bodily injury claim file review, such as amaise, apply this model to injury files. Specialized agents each run one task, such as reading medical records, building the treatment timeline, finding gaps or spotting contradictions. The agentic AI platform for claims links each output back to its source page, and the professional makes the call.
For claims leaders, the test for agentic AI claims tools comes down to three questions. Where does the tool start? Where does it stop? Who decides? A tool that leaves the decision with the adjuster, and ties every finding to a page, fits the stack.
Common questions
Does agentic AI replace the claims system or RPA?
No. The claims system remains the official record, and RPA still runs fixed steps on structured data. Agentic AI adds a layer that works across the whole claim file.
Is document AI the same as agentic AI?
No. Document AI turns single pages into data. Agentic AI uses that data to work toward a goal across the whole file.
Does agentic AI make claim decisions?
No. Coverage, liability, reserve and settlement decisions stay with the adjuster. Claims guidelines and tool settings should state which actions need human approval.
Key takeaways
- A claims stack has four layers: the claims system of record, RPA, document AI and agentic AI.
- RPA repeats rule-based steps on screens and structured data.
- Document AI sorts pages and pulls out fields, one document at a time.
- Agentic AI works toward a goal across the whole file and links each finding to its source page.
- Decisions on coverage, liability, reserves and settlement stay with the adjuster.
