Turning multimodal first discover of loss (FNOL) proof into tagged, decision-ready consumption so adjusters begin with context as an alternative of uncooked artifacts.
Guide FNOL processing consumes important professional time on repetitive duties as a result of unstructured, multimodal proof should be interpreted by portals designed for human interplay. Pictures captured within the discipline, walkaround movies, scanned paperwork, and dictated or recorded notes all enter the system at consumption, the place choices immediately affect declare cycle time, downstream accuracy, and buyer expertise.
Throughout insurance coverage strains, this second is deceptively advanced. FNOL consumption is commonly described as “simply opening a declare,” however in apply, it’s the place giant volumes of unstructured information should be interpreted, validated, and correlated earlier than any significant choices can start.
The problem is critical: claims professionals spend extreme time on repetitive consumption validation. Navigating portals, verifying proof completeness, and decoding artifacts earlier than making use of their experience to higher-value choices takes appreciable time. Trade observations recommend that consumption validation can devour a considerable share of an adjuster’s time throughout preliminary declare processing, with typical submissions requiring significant display work earlier than evaluation can start. Throughout quantity spikes from catastrophic occasions or seasonal surges, these delays compound, creating backlogs that sluggish declare decision and influence buyer expertise.
On this publish, we display how a hands-free FNOL consumption system combines brokers constructed with the Strands Brokers SDK for area reasoning with Amazon Bedrock AgentCore Browser Device for dwell portal interplay. This strategy preserves human experience whereas eradicating repetitive display work.
The answer combines two complementary capabilities:
Strands Brokers is an open supply SDK that takes a model-driven strategy to constructing generative AI brokers. On this structure, the brokers (constructed with Strands Brokers) apply insurance-specific enterprise guidelines, corresponding to proof interpretation, cross-modal correlation, and declare complexity evaluation utilizing basis fashions (FMs) served by Amazon Bedrock.
Browser reasoning is carried out by Amazon Nova Act, a consumer SDK that interprets natural-language directions (for instance, “open the following unprocessed declare” or “set off picture evaluation”) and interprets them into grounded UI actions. Amazon Bedrock AgentCore Browser software gives the managed, remoted Chrome session that Nova Act connects to for executing these actions. AgentCore Browser Device additionally gives session recording and dwell view capabilities for observability.
On this workflow, Nova Act drives the consumption course of by reasoning about what’s seen on display by the AgentCore Browser session, whereas the Strands-based brokers carry out area reasoning within the background. Nova Act determines when proof should be analyzed and orchestrates portal interactions, and the area brokers decide what the proof means by making use of the identical area logic a human reviewer would use.
The result’s automation of handbook display work whereas preserving human oversight and auditability. Claims professionals obtain context-rich, pre-analyzed submissions prepared for judgment reasonably than validation. Tagged proof turns into a sturdy operational asset, supporting higher routing, sample evaluation, and steady workflow refinement throughout the claims lifecycle.
The workflow is illustrated utilizing actual browser automation recordings captured immediately from the system in motion.
The chance: Optimizing claims consumption to amplify human experience
Throughout insurance coverage strains (auto, property and casualty, life, well being, and specialty), declare consumption marks the second when unstructured info first enters the system. Pictures, movies, scanned paperwork, and recorded notes arrive collectively, usually incomplete, inconsistently labeled, and infrequently standardized.
Claims professionals deliver deep area data to this second. They know what usable proof appears to be like like, what is usually lacking, how artifacts relate to 1 one other, and which indicators matter for protection, severity, and subsequent steps. But in the present day, a lot of that experience is utilized by sluggish, handbook portal work clicking by screens and visually inspecting artifacts one after the other. Earlier than significant evaluation can start, reviewers should reply foundational questions that rely closely on expertise. These embody whether or not required artifacts are current, whether or not pictures and movies are usable and related, whether or not audio notes include materials observations, and whether or not the submission is enough to proceed at once.
Answering these questions requires painstaking display work. A typical FNOL submission can embody dozens of artifacts unfold throughout a number of views, requiring reviewers to find proof, open and interpret every merchandise, correlate indicators throughout modalities, evaluate findings in opposition to coverage thresholds, and seize summaries for audit continuity.
These steps are important, however they’re additionally repetitive and mechanical. They require consideration reasonably than judgment. Consequently, expert adjusters and examiners spend a disproportionate period of time validating consumption completeness earlier than they will apply their experience to higher-value choices.
This problem exists in on a regular basis claims processing and turns into extra pronounced throughout quantity spikes from disaster occasions, seasonal auto claims, or surges in well being and life claims exercise. As workloads enhance, backlogs develop, proof assessment turns into rushed or inconsistent, and human judgment is utilized later than it must be.
The problem isn’t a lack of knowledge or expertise. It’s that area data is being exercised too late within the course of, after time has already been spent on repetitive consumption validation.
Why encoding area data adjustments the panorama
Declare consumption accelerates when vital resolution logic is captured in structured guidelines and utilized constantly at ingestion time, reasonably than relying solely on particular person expertise and instinct.
Skilled reviewers intuitively know which picture angles are required for various declare sorts, when video can substitute for lacking photos, which combos of artifacts sign larger complexity, and which gaps are more likely to stall downstream processing.
Agentic generative AI makes it potential to encode this working data into enterprise guidelines and reasoning instruments that may be utilized constantly as proof enters the system.
By combining Strands Brokers with Nova Act and the AgentCore Browser Device, mechanical consumption work like navigating portals, opening claims, and triggering evaluation is separated from area reasoning. Nova Act advances the workflow by the Browser Device session, whereas Strands Brokers apply professional logic to interpret, tag, and correlate proof.
When proof is tagged at ingestion, lacking or inadequate artifacts are detected early, relevance turns into express reasonably than implicit, and claims may be triaged primarily based on what’s current. Human reviewers start with context as an alternative of ranging from scratch.
Why automated proof tagging issues – now and later
Automated tagging accelerates the present declare by guaranteeing consumption completeness and readability earlier than downstream steps start. Reviewers spend much less time confirming fundamentals and extra time making use of judgment the place it issues.
Over time, constantly tagged proof turns into a sturdy information asset. As a result of tags are generated by codified area reasoning, not one time interpretation, insurers can do the next:
- Enhance routing and prioritization
- Scale back rework attributable to incomplete submissions
- Determine patterns that result in delays or escalations
- Refine consumption guidelines as new situations emerge, with out altering compliance boundaries or resolution authority
As tagged proof accumulates, unstructured artifacts are now not remoted recordsdata. Photos, movies, and audio develop into searchable, analyzable indicators that help new workflows, corresponding to proactive outreach when widespread gaps are detected, pre-staging claims for specialised groups, and shortening cycle instances for comparable future claims.
Most significantly, tagging permits area experience to be utilized as soon as at ingestion and reused all through the lifecycle, reasonably than rediscovered repeatedly at totally different phases.
That is the shift agentic automation permits: shifting experience upstream, enriching downstream programs with structured indicators, and enabling quicker, extra constant decision, with out eradicating people from the loop.
To display how this shift may be carried out with out modifying current portals, the next part walks by an agentic FNOL consumption structure that mixes browser-level automation with reasoning-driven brokers.
Resolution overview: Agentic consumption with out portal adjustments
This prototype demonstrates how FNOL consumption may be automated end-to-end utilizing agentic reasoning and browser-level interplay. In manufacturing, the identical browser automation strategy would work in opposition to current portals with out modification, as a result of the Nova Act consumer SDK interacts with the UI as a human would.
The prototype is constructed to reflect a sensible manufacturing surroundings. The FNOL portal and backend companies run as a containerized software on AWS, whereas agent-driven browser automation interacts with the dwell portal precisely as a human reviewer would. This separation permits area reasoning and UI management to evolve independently, whereas preserving auditability and operational security.
At a excessive stage, the answer assumes a working familiarity with how fashionable, agentic programs are deployed on AWS. This consists of using FMs for reasoning, containerized companies for software runtime, and event-driven storage for state and proof. No prior expertise with conventional robotic course of automation (RPA) instruments is required. The automation described right here depends on reasoning over UI state reasonably than replaying pre-recorded scripts or hard-coded flows.
AWS account and permissions
You want entry to an AWS account with permissions to deploy and handle the assets utilized by the answer, together with AWS Cloud Growth Package (AWS CDK), Amazon Elastic Container Service (Amazon ECS) on AWS Fargate, Amazon Easy Storage Service (Amazon S3), Amazon DynamoDB, Elastic Load Balancing (Software Load Balancer), Amazon CloudFront, and AWS Id and Entry Administration (IAM) roles and insurance policies.
The deployment assumes that AWS credentials are configured domestically utilizing a typical improvement setup with the AWS Command Line Interface (AWS CLI).
Runtime surroundings and deployment mannequin
The FNOL consumption person interface and backend companies, together with proof evaluation and declare complexity analysis carried out utilizing Strands Brokers, are packaged as Docker containers and deployed on Amazon ECS with AWS Fargate. Infrastructure is provisioned utilizing AWS CDK, which builds container photos and creates the required compute, storage, and networking assets as a part of a single deployment workflow.
Unstructured proof artifacts corresponding to photos, movies, and transcripts are saved in Amazon S3. Declare metadata, proof references, and agent-generated evaluation outputs are endured in Amazon DynamoDB. This permits brokers to retrieve, correlate, and motive over proof all through consumption.
Browser automation in apply
Agent-driven browser automation is executed from a separate management surroundings, corresponding to a workstation or automation host, and connects to the deployed FNOL software by an AgentCore Browser session. This displays how browser automation is usually operated in real-world environments. Nova Act, the consumer SDK liable for browser reasoning, connects to the managed Chrome session supplied by AgentCore Browser Device by Chrome DevTools Protocol (CDP) over WebSocket. The automation layer observes and interacts with the dwell portal by this managed browser, whereas backend companies stay hosted and remoted.
By conserving browser management exterior to the applying runtime, the system maintains clear operational boundaries. Brokers see precisely what a human reviewer would see on display, make choices primarily based on present UI state, and act intentionally with out requiring direct entry to portal internals or software code.
Deployment workflow and setup
The complete deployment workflow, together with infrastructure provisioning, container deployment, non-compulsory information era, and browser automation setup is automated by scripts and configuration recordsdata supplied within the accompanying GitHub repository.
Structure overview
At a excessive stage, the structure consists of the next complementary layers:
- Browser interplay. Nova Act connects to an AgentCore Browser Device session by Chrome DevTools Protocol (CDP) over WebSocket, reasoning in regards to the FNOL portal’s UI state and appearing intentionally on what’s seen.
- Area reasoning. Two brokers are constructed with the Strands Brokers SDK: an Proof Analyzer agent that interprets and tags multimodal proof, and a Claims Complexity Analyzer agent that assesses declare complexity.
- Execution observability. Screenshots, prompts, reasoning, and UI state transitions are captured robotically at every step, producing a reviewable audit path with out extra instrumentation.
- Infrastructure and persistence. Amazon ECS on AWS Fargate runs the applying, Amazon S3 shops proof artifacts, Amazon DynamoDB maintains declare state and evaluation outputs, and Amazon CloudWatch gives operational visibility.
The next diagram reveals how the totally different elements match collectively to automate FNOL consumption:
The structure is deliberately layered to separate portal interplay, area reasoning, execution observability, and infrastructure issues, whereas preserving a single, end-to-end FNOL consumption workflow. Agent-driven browser automation operates on the high of the stack, interacting with the FNOL portal precisely as a human reviewer would. Area-specific reasoning is utilized independently by Strands Brokers, whereas AWS infrastructure gives the managed basis for execution, persistence, and operational visibility.
Browser interplay with Nova Act and AgentCore Browser Device
Nova Act is liable for observing and interacting with the FNOL portal’s person interface, with out embedding any area logic or decision-making. Operating inside an AgentCore Browser Device session and connecting to the browser utilizing Chrome DevTools Protocol (CDP), Nova Act causes in regards to the present UI state in actual time. It navigates declare queues, identifies unprocessed proof sections, invokes Analyze Photos, Analyze Movies, and Analyze Audio actions, interacts with modal dialogs, and scrolls solely when essential to keep away from unintended UI adjustments. This strategy permits automation to behave like a cautious human reviewer: observing what’s seen on display, deciding which motion is acceptable, and appearing intentionally primarily based on present state reasonably than replaying predefined steps or brittle scripts.
Execution observability and auditability
As a result of AgentCore Browser executes actions by a managed browser session, each interplay is observable and traceable by design. Because the automation runs, actions may be noticed dwell by the Chrome DevTools Protocol (CDP) session, offering real-time visibility into how the agent interacts with the FNOL portal.
At every resolution level, screenshots are captured robotically, whereas prompts, choices, and UI state transitions are recorded as structured metadata. Collectively, these artifacts type a whole execution path that makes the agent’s habits clear and reviewable. It’s at all times potential to find out what the agent noticed on display, why a particular motion was taken, which proof was processed, and what conclusions have been derived because of this.
This produces a pure audit path with out requiring extra instrumentation or customized logging. That is a vital functionality in regulated insurance coverage environments the place explainability, traceability, and operational accountability are as vital as automation itself.
Capturing screenshots throughout agent execution
On this prototype, browser automation is configured with a session-specific logging listing. Because the agent executes every act() step, Nova Act captures the seen browser state and persists screenshots alongside step metadata corresponding to prompts, timestamps, and motion identifiers.
These artifacts help each operational troubleshooting (by revealing precisely what the agent noticed when encountering sudden UI states) and audit or post-run assessment, with out counting on steady display recordings. Every execution produces an remoted, timestamped folder containing screenshots and logs. This makes runs reproducible, inspectable, and clearly attributable to a particular session.
Downstream processing and storage on AWS
After proof has been analyzed and tagged, AWS companies present the sturdy basis required to persist outcomes, preserve declare state, and help operational visibility all through the consumption workflow.
The 2 Strands-based brokers deal with all reasoning-driven processing independently of the person interface. The Proof Analyzer agent performs multimodal proof evaluation throughout photos, movies, and transcripts with structured metadata tagging, and the Claims Complexity Analyzer agent evaluates declare complexity utilizing specialised instruments. Analyzed artifacts, summaries, and tagging outputs are saved in Amazon S3, whereas declare state, proof references, and agent-generated outcomes are maintained in Amazon DynamoDB to protect a whole, queryable report of what was noticed and inferred.
Operational logs, metrics, and execution traces are captured by Amazon CloudWatch, offering visibility into system habits and supporting monitoring, troubleshooting, and audit necessities. Collectively, these elements rework uncooked FNOL submissions into structured, decision-ready inputs at ingestion time, earlier than claims are routed or escalated for added assessment. This makes certain that downstream processes obtain constant, context-rich, and traceable info from the beginning.
With the structure in place, the rest of this publish follows the system because it operates in apply. The subsequent sections stroll by a single FNOL consumption sequence because it unfolds within the dwell portal, beginning on the FNOL queue and progressing by proof evaluation and complexity classification. Every step is illustrated utilizing precise browser automation recordings and screenshots captured throughout execution, exhibiting how agentic automation and area reasoning work collectively in actual time.
From FNOL queue to say choice
The workflow begins on the FNOL queue. Nova Act observes the dwell portal state by the AgentCore Browser Device session, identifies the following declare prepared for processing primarily based on seen standing indicators, and navigates into the declare element view. As a result of the choice is grounded in present UI state reasonably than pre-recorded scripts or hard-coded selectors, the identical logic handles new queue layouts, reordered columns, or altering row counts with out modification.
The next screenshot reveals the queue precisely as Nova Act noticed it throughout a consultant run. The automation causes over seen UI components, corresponding to declare rows and standing indicators, to find out the following eligible motion. The screenshot is robotically captured as a part of the execution log and serves as each a troubleshooting artifact and an audit report.
With the declare chosen and its element view loaded, the workflow strikes from queue administration into consumption processing. At this stage, the element view has completed rendering and the obtainable proof sections are recognized, establishing the following part of structured proof evaluation.
Proof evaluation: Structured assessment throughout modalities
Proof assessment is the place FNOL consumption sometimes slows down. The execution recording highlights this by exhibiting three distinct actions, Analyze Photos, Analyze Movies, and Analyze Audio, every comparable to a separate assessment path. This mirrors how a human examiner evaluates proof one modality at a time reasonably than treating all artifacts uniformly.
At this stage, obligations divide cleanly. Nova Act manages UI management circulate, figuring out when proof is prepared, invoking the suitable evaluation motion, and ready for completion. Strands Brokers run server-side, making use of insurance-specific reasoning with codified enterprise guidelines that replicate how a human reviewer would interpret every artifact. This separation is intentional. UI orchestration determines when and the place evaluation ought to happen, whereas area reasoning determines what the proof means. The outcome mirrors how human examiners work, first figuring out obtainable proof, then decoding it, whereas permitting every step to execute at machine velocity and with full traceability.
The next screenshot reveals the proof part as Nova Act observes it earlier than evaluation begins. Controls corresponding to Analyze Photos, Analyze Movies, and Analyze Audio are detected primarily based on present UI state reasonably than hard-coded selectors.
Analyze photos: Tagging visible proof
When Analyze Photos is invoked, every submitted picture is evaluated independently. Strands Brokers apply insurance-specific enterprise guidelines that replicate how skilled reviewers interpret visible proof. This consists of figuring out what the picture depicts (for instance, automobile injury, roof floor, siding, inside, or medical documentation), assessing whether or not the attitude is acceptable for the declare kind, confirming readability and value, and flagging injury indicators. Somewhat than leaving this interpretation implicit, every picture is tagged with structured attributes that make human judgment express, constant, and reusable all through the declare lifecycle.
The next screenshot captures the UI state as picture evaluation is initiated, with the picture set and evaluation controls seen to the automation and recorded as a part of the execution hint. After evaluation is triggered, the modal opens and presents the submitted photos whereas Strands Brokers consider every one within the background. The screenshot reveals this in-progress state, with picture proof seen on display as analysis is utilized.
Analyze movies: Treating movement as first-class proof
When Analyze Movies is invoked, video submissions are evaluated as proof, not as opaque attachments. Strands Brokers assess what every video captures: whether or not it provides info past nonetheless photos, dietary supplements lacking pictures, or corroborates or contradicts different artifacts. Video-derived indicators are then tagged and normalized to take part immediately in downstream reasoning alongside photos and paperwork, reasonably than being handled as a secondary or handbook assessment step.
The next screenshot captures the portal state as video evaluation is triggered, preserving UI context so video analysis stays absolutely traceable and auditable.
When Analyze Audio is invoked, the system processes audio proof by corresponding name transcripts reasonably than uncooked audio immediately. For every audio recording, a corresponding textual content transcript is retrieved. The Proof Analyzer Strands Agent then analyzes the transcript textual content to extract materials observations and factual statements, reported injury or circumstances, and contextual particulars that complement visible or document-based proof.
Transcribed indicators are then correlated with picture and video tags, mirroring how a human reviewer cross-references spoken context with visible proof throughout consumption. The next screenshot captures the portal state as audio evaluation is triggered, with the transcript seen to the automation.
Taken collectively, the Analyze Photos, Analyze Movies, and Analyze Audio steps produce a layered, reviewable report of consumption. AgentCore Browser Device captures screenshots and session exercise at every step. Nova Act information the immediate, reasoning, and motion taken in response to the present UI state. Strands Brokers persist the structured tags and classification reasoning produced for every artifact.
The result’s a whole audit path of what was on display, why the agent acted, and the way the proof was interpreted, all of which is generated as a byproduct of execution reasonably than by extra instrumentation.
Why this step issues
By the top of proof evaluation, each picture, video, and audio artifact has been evaluated and tagged. No proof stays unclassified, and high quality, relevance, and completeness are made express reasonably than inferred.
This transforms uncooked FNOL submissions into structured, decision-ready inputs earlier than downstream routing, escalation, or handbook assessment happens. This units the stage for complexity evaluation of the submitted declare within the subsequent step.
Complexity evaluation: From tagged proof to triage choices
With proof absolutely tagged, the agent evaluates every declare holistically. As a substitute of counting on static consumption fields alone, Strands Brokers mix declare metadata with evidence-derived indicators noticed throughout consumption to evaluate complexity utilizing guidelines already acquainted to insurance coverage operations. These embody severity indicators throughout modalities, proof completeness and inner consistency, and coverage thresholds that decide escalation or routing.
As a result of this evaluation is grounded in what was submitted and noticed, complexity classification displays the true state of the declare reasonably than assumptions made at submission time. Claims are categorized as Easy or Complicated. Easy claims are auto resolved, whereas Complicated claims are routed to “Wants Evaluation” standing with structured notes generated robotically explaining why the declare was flagged. These notes present quick and actionable context for downstream customers.
These screenshots seize the portal state after complexity evaluation has accomplished. Proof-derived indicators are surfaced alongside structured notes, making the rationale for classification clear, reviewable, and auditable.
Why this issues for insurance coverage carriers
By deferring human involvement till complexity has been assessed, experience is utilized on the proper second, on interpretation, judgment, and determination reasonably than consumption validation. Easy claims progress with out pointless friction, whereas advanced instances are surfaced early with context already in place. This reduces queue contamination, prevents late-stage escalation, and improves predictability throughout each steady-state operations and quantity spikes.
Human-in-the-loop, not human-in-the-weeds
This method doesn’t take away individuals from the method. It adjustments the place they have interaction. With proof already analyzed and tagged, adjusters start their work with context as an alternative of uncooked artifacts. Evaluation replaces reprocessing. Corrections develop into suggestions reasonably than rework.
Over time, this suggestions improves enterprise guidelines incrementally as patterns emerge throughout claims. FNOL consumption evolves from a bottleneck right into a studying system, one which repeatedly refines how proof is interpreted, routed, and acted upon with out growing operational burden.
Why this issues for insurance coverage carriers
This strategy essentially adjustments the place and when experience is utilized within the claims lifecycle. By structuring multimodal proof at ingestion time, carriers cut back consumption dealing with time by robotically assessing completeness and relevance. Claims transfer quicker, particularly throughout quantity surges, as a result of fewer submissions stall downstream ready for validation.
Proof interpretation turns into extra constant and fewer depending on particular person reviewer expertise. Gaps are recognized early, lowering downstream corrections and rework. Equally importantly, adjusters expertise much less cognitive fatigue and might deal with choices reasonably than validation. These capabilities work with out changing current programs or disrupting established workflows. The automation works with the portals carriers already depend on.
Past FNOL: The worth of tagged unstructured proof
Whereas FNOL is the entry level, the worth of structured, tagged proof extends throughout the complete claims lifecycle. After unstructured artifacts are constantly interpreted and tagged at ingestion, they cease behaving like static attachments and start functioning as operational indicators.
Claims may be routed primarily based on what the proof really reveals reasonably than counting on coarse consumption fields or handbook triage. Downstream workflows arrive pre-populated with context, together with injury indicators, completeness indicators, and corroborating proof. This reduces friction at each handoff and minimizes the necessity for re-validation. As patterns emerge throughout claims, proof assortment steerage improves organically, serving to carriers establish widespread gaps and modify consumption expectations earlier than these gaps create downstream delays.
Over time, historic claims develop into analyzable primarily based on what was really submitted and noticed, not solely how claims have been labeled at consumption. This permits deeper operational perception into cycle-time drivers, escalation patterns, and proof high quality throughout areas, perils, and declare sorts. Tagged proof turns unstructured recordsdata into reusable, queryable information that helps higher choices with out altering compliance boundaries, resolution authority, or core programs.
The outcome isn’t solely quicker FNOL processing, however a basis for extra adaptive, evidence-driven claims operations.
Conclusion
On this publish, we confirmed how a hands-free FNOL consumption system combines Strands Brokers with Amazon Bedrock AgentCore Browser Device and Amazon Nova Act to construction multimodal proof in the intervening time it enters the system. FNOL shifts from a validation bottleneck to an acceleration level. Claims progress with context already established. Routing choices are knowledgeable by what was really submitted. Escalations happen earlier and extra predictably. Easy instances transfer ahead with out pointless dealing with.
This shift doesn’t rely upon changing portals, rewriting current programs, or altering resolution authority. It comes from making consumption interpretation express (how proof is evaluated, which indicators are significant, and the way gaps have an effect on downstream processing) and making use of that interpretation constantly at ingestion time. What was beforehand re-derived by repeated handbook assessment turns into structured, sturdy, and reusable.
The result is just not automation for its personal sake, however a more practical use of judgment. Interpretation occurs as soon as. Proof is tagged in a approach that persists. These indicators journey with the declare as an alternative of being rediscovered at every stage. FNOL consumption improves by clearer indicators and higher circulate, permitting downstream processes to start out with context reasonably than uncertainty.
To discover this strategy in your personal surroundings, deploy the prototype from the GitHub repository, and be taught extra within the Amazon Bedrock AgentCore documentation, the Amazon Nova Act documentation, and the Strands Brokers documentation.
Concerning the writer
