How Agentic AI Changes the Claims Operating Model
9 min read
Part 2 of 2: when the AI does the work and people own the calls, the job, the queue, and the governance all change
Every wave of automation in claims sparks the same worry, that the adjuster’s role disappears and control goes with it. Agentic AI raises the opposite prospect. When the AI does the routine work on every claim and people own the judgment and authority calls, adjusters spend their time on the decisions that matter and the busywork falls away. That shift redesigns the claims operating model: the adjuster’s role, the shape of the queue, ownership of the whole file, and the controls around AI decisions all change. Running the new capability through the old workflow, or without production-grade governance, captures little of the value and adds real risk.
Part 1 of this series (Agentic AI vs STP: Why STP Rate Stalls Near 10%) reframed the metric and introduced three autonomy levels: assist, recommend, and execute. This blog post follows that distinction through to roles, queues, capacity, and governance, because an AI that executes actions on a claim is a different operational animal from one that only drafts and recommends.
What you will learn:
- The core shift: the human moves from operator to authority
- Why autonomy level, not claim type, sets impact, reversibility, and required authority
- Why a pure decision queue needs persistent case ownership behind it
- Why the human exception queue can become the new bottleneck, and how to size it
- The governance and control structure agentic claims automation actually requires
- Where the claimant fits, and when a human is non-negotiable
The core shift: human adjuster moves from operator to authority
Today the adjuster is the default operator and automation is the exception. Task-level automation flips it: the AI does the routine work, and the human is the authority on the calls that carry stakes.
Today the adjuster does the work, tools assist, and anything the tools cannot handle stays on the desk, which is most of it. Task-level automation reverses that relationship. The AI handles the routine work – assist, recommend, and low-stakes execute – on every claim, and the human becomes the authority on judgment and regulated actions. That is a different job and a different process, and the software has to be built around both.
| Dimension | Old workflow – human-led | New workflow – AI-driven |
|---|---|---|
| Unit of human work | A caseload of claims you own end to end | Judgment and authority decisions, plus ownership of defined claim segments |
| Process shape | Linear pipeline; each step waits for a person to push it | Event-driven; the AI advances files and pulls a person in on a trigger |
| What sits in the queue | Claims waiting for an adjuster | Decisions waiting for judgment, inside claims that still have an owner |
| Skill premium | Throughput, administrative accuracy, process knowledge | Judgment, negotiation, complex investigation, oversight of the AI |
| Management and QA | Claims per adjuster; sample closed files after the fact | Automation coverage, exception rate, override outcomes, live monitoring, audit trail |
| Capacity model | Headcount times claims per day | Automation coverage plus human capacity for exceptions, sized for peak |
Autonomy level, not claim type, sets the risk
What determines risk is whether the AI assists, recommends, or executes, weighed against impact if wrong, reversibility, confidence, and required authority. The operating model is built on that grid, not on claim categories.
Part 1 of the blog introduced the three levels. In the operating model they become the control surface. An assist task (draft a rationale) can run almost everywhere, because a person reviews before anything changes. An execute task (issue a payment, set a reserve, send a denial) needs tight conditions, because it changes the claim and some actions are hard to reverse.
| Autonomy level | Example | Impact if wrong | Reversibility | Where it belongs |
|---|---|---|---|---|
| Assist | Draft reserve rationale, summarize a file | Low | Full | Broadly, with light review |
| Recommend | Propose a coverage position with evidence | Medium | Reversible before action | Where confidence and data quality are high |
| Execute | Issue payment, set reserve, send denial | High | Often hard to reverse | Only low-stakes, reversible actions; high-stakes stay manual |
Set autonomy per task using this grid, tuned by confidence and data quality, and the same claim can mix all three levels: execute the intake, recommend the reconciliation, assist the rationale, and leave the coverage call to a person. This is what “human in the loop” means in practice, and it is configuration, not a slogan.
Who owns the claim when the adjuster role moves to oversight
Automated progression turns the caseload into a decision queue routed to specialists. That only works with persistent ownership of complex and vulnerable files behind it, or accountability and strategy fragments.
With the AI advancing files, the adjuster’s work shifts from grinding 40 files to resolving the judgment calls surfaced across the book. Route those by decision type: liability calls to the liability specialist, coverage disputes to the coverage expert. Manage decision quality and turnaround rather than claims-closed-per-adjuster.
A pure decision queue has failure modes worth naming.
- Specialists can see isolated questions without the claim’s strategy, so decisions become locally correct but globally inconsistent.
- Accountability fragments when no one owns the outcome.
- Policyholders lose a stable point of contact, and files bounce between specialists.
- The carrier still owns the outcome legally and operationally, whatever the AI did.
So the workable model is hybrid: automated progression and decision-based routing on top of persistent case ownership for defined segments. Simple, high-frequency claims can run largely on automated progression with pooled exception handling. Complex, litigated, or vulnerable-claimant files keep a named owner who holds the strategy and the relationship, with the AI doing the assembly and surfacing the decisions. Decide which segments need an owner up front, because that choice shapes routing, staffing, and accountability.
The human exception queue can become the new bottleneck
Exception handling does not scale automatically. Bursts of correlated decisions after a catastrophe, an outage, or a model change can swamp the human queue, so size it for peak, not average.
The appeal of automated progression is that the machine absorbs volume. The catch is that every stop feeds a human queue, and that queue has its own capacity limit. A simple way to size it:
Human demand = claim volume × exceptions per claim × minutes per exception.
Average demand is the easy case. The hard case is correlated bursts: a catastrophe that spikes volume, a data-source outage that raises exceptions per claim, a rule or model change that briefly raises both, or model drift that quietly increases escalations. Any of these can push the decision queue past capacity at once.
Plan for it. Staff and cross-train for peak rather than average, prioritize surfaced decisions by severity and deadline, age and escalate items that wait too long, and define a fallback mode for when the queue exceeds capacity, whether that means widening autonomy on low-stakes work, pulling in cross-trained staff, or deliberately slowing intake. A decision queue with no capacity plan turns a resilient process into a fragile one.
AI claims processing also has its constraints
An AI operator drops the human constraints of shifts and sequential handoffs, and takes on machine and ecosystem ones instead: third-party data, API limits, model latency and cost, contradictory inputs, permissions, and regulatory waiting periods.
Redesigning around an operator that works continuously and in parallel is the right instinct. The honest version is that this trades human constraints for a different set. Agentic systems still depend on external data and third parties, API rate limits and system availability, model latency and cost, incomplete or contradictory inputs, security permissions, mandated waiting periods, and how fast claimants and providers respond. Some of the old idle time between steps disappears; some of it moves to waiting on an API, a document, or a statutory clock. Designing the process around the AI is still the right move; assuming the AI has no constraints at all is how timelines slip.
Why governing AI in claims takes more than just audit trail
Executing actions on real claims under regulation needs production-grade controls: validation before deployment, drift monitoring, segregation of duties, incident response, appeal paths, and review of decisions humans did not override.
An audit log is table stakes. Running AI that recommends and executes on real claims, under market conduct rules, needs a control set closer to model risk management than to workflow configuration:
- Predeployment validation against acceptance thresholds, per claim type and autonomy level
- Model and prompt version control, with change management for vendor and model updates
- Drift and outcome monitoring, not just uptime, watching for quality decay over time
- Access control and segregation of duties between who builds, who authorizes, and who assures
- Data retention, privacy, and permissions on what the AI can read and store
- Bias and disparate-impact testing on decisions and recommendations
- Incident response, rollback, kill switches, and a manual fallback mode
- Appeal and reconsideration paths for claimants, and for adjusters who disagree with the AI
- Random review of decisions, including the ones humans did not override
That last point matters most. Override rate is an ambiguous signal: a low rate can mean the recommendations are excellent, or it can mean people are rubber-stamping. Only sampling the non-overridden decisions tells you which. Build that review in from the start, because a regulator will ask how you know your oversight is real.
One automation owner is not enough
Enterprise deployment needs separated roles: an operational owner, independent validation, platform ownership, and executive accountability. Combining rule-setting, monitoring, and assurance in one seat removes the check.
Automation owner is necessary but not sufficient. Concentrating rule-setting, monitoring, and assurance in one role removes the separation of duties a regulated process needs.
A workable structure separates four things:
- Operational owner: sets autonomy thresholds and runs the day-to-day, in claims operations
- Independent validation and challenge: tests models and outcomes without owning them, in model risk or a second line
- Technology and platform ownership: runs the system, integrations, and security
- Executive accountability: a named owner of claim outcomes and regulatory exposure
Claims, legal, compliance, model risk, security, data governance, technology, and workforce leadership all have a stake. The point is not a heavier committee; it is that the person tuning autonomy should not also be the only person grading its results.
Where the claimant fits
The operating model is not only internal. Decide when a claimant is offered a human regardless of task type, how vulnerable claimants are identified, and which conversations stay voice-to-voice.
An efficiency-led redesign can optimize the internal workflow and forget the person on the other end. Answer a few questions explicitly.
- When is a claimant offered a human regardless of what the task grid says, for example after a serious injury or a bereavement?
- How are vulnerable claimants identified and routed to a person early?
- Which communications stay voice-to-voice rather than automated?
- How is continuity maintained when several specialists touch one file?
- What happens when a claimant challenges an AI-prepared conclusion, and how do they reach a human who can reconsider it?
EY notes that carriers are rebalancing digital and voice interaction rather than assuming digital is always better, and the claims most likely to need a voice are the complex, high-severity ones this series is about.
“Clive gave us measurable gains on paper, but more importantly, it unlocked momentum. We’re seeing faster responses, more streamlined workflows, and a clear path to scaling operations without scaling costs. This also allows our team to dedicate more time to complex issues, driving better outcomes for our claimants.”
Marc Mercer, Director of European Claims, INSHUR
How to redesign the operating model, not just install a tool
A practical sequence for the transition, once you have made the task-level shift from Part 1 of the article.
Step | Why it matters | |
|---|---|---|
Classify tasks by autonomy level and risk (assist, recommend, execute against impact, reversibility, confidence) | Sets where the AI can act and where a human authorizes | |
Decide which claim segments keep a persistent owner | Prevents fragmented strategy and accountability on complex and vulnerable files | |
Redesign the queue around decisions, routed by type, with SLAs | Matches judgment to expertise and measures what now matters | |
Size the human exception queue for peak, with fallback modes | Keeps the process resilient to catastrophe, outage, and drift bursts | |
Stand up separated roles: operational owner, independent validation, platform, executive accountability | Preserves segregation of duties in a regulated process | |
Build governance: validation, drift monitoring, incident response, appeal paths, review of non-overrides | Makes oversight real and defensible in a market conduct exam | |
Define claimant rules: when a human is offered, how vulnerable claimants are routed | Protects experience and continuity, not just efficiency |
When you sequence it right, the AI’s constant work and the adjuster’s focused judgment support each other, inside controls that can hold up to an audit. Bolt the AI onto the old process instead, and you just get an expensive tool fighting the workflow it was added to.
Key takeaways
- Task-level automation is an operating-model change. The human moves from operator to authority.
- Autonomy level, not claim type, sets the risk. Build the model on the assist/recommend/execute grid, weighed by impact, reversibility, confidence, and authority.
- The caseload becomes a decision queue, but complex and vulnerable files still need a persistent owner, or strategy and accountability fragment.
- The human exception queue can become the bottleneck. Size it for peak using volume times exceptions times minutes, with fallback modes for correlated bursts.
- Governance is larger than an audit trail: validation, drift monitoring, segregation of duties, incident response, appeal paths, and review of decisions humans did not override.
- Separate the roles (operational owner, independent validation, platform, executive accountability) and decide the claimant rules up front.
How the Agentic Claims Operating Model Works in Five Sigma
The carriers that get the most from agentic automation will redesign the process around the AI doing the work and people owning the calls, with persistent ownership of hard files, an exception queue sized for peak, and governance that holds up in an exam.
Clive AI, Five Sigma’s agentic AI claims expert, is built for that operating model. It runs natively inside the Five Sigma claims management platform or on top of a carrier’s existing system and works the tasks across the lifecycle under the insurer’s SOPs, at the autonomy level you set per task and claim type, advancing files and logging each action on a shared live claim record. When it reaches a judgment or authority call it stops and hands off with the context assembled. It escalates when data is missing, when confidence falls below your threshold, when sources conflict, when an action needs approval, or when an action fails. Because autonomy is configuration rather than custom engineering, who decides what, and when a human steps in, stays in your control.
Planning your agentic AI strategy?
See what your claims operation could look like when AI handles the routine work and your team stays in control.
Frequently asked questions
Does agentic AI replace adjusters?
No. It does the routine work so adjusters spend their time on judgment: coverage disputes, negotiation, complex liability, and ownership of the hardest files. The job shifts from operating the process to owning the decisions and supervising the AI.
What is the biggest implementation risk with agentic claims automation?
Running the new capability through the old workflow, or without production-grade controls. The value comes from redesigning the queue, engineering the handoff, sizing exception capacity, and building governance, not from installing an agent on top of yesterday’s process.
How do you keep the human exception queue from becoming a bottleneck?
Size it for peak, not average: claim volume times exceptions per claim times minutes per exception. Then prioritize by severity and deadline, cross-train for bursts after catastrophes or outages, and define a fallback mode for when the queue exceeds capacity.
What governance does agentic claims automation need?
More than an audit log: predeployment validation, model and prompt version control, drift and outcome monitoring, segregation of duties, bias testing, incident response and rollback, claimant appeal paths, and random review of decisions humans did not override.
Why is override rate an unreliable metric on its own?
A low override rate can mean the AI’s recommendations are excellent, or that people are rubber-stamping them. Only sampling the decisions humans did not override tells you which, so pair override rate with random review.
When should a claimant always get a human?
Regardless of the task grid, offer a human for serious injury, bereavement, identified vulnerability, or when a claimant challenges an AI-prepared conclusion. Decide these rules up front and route vulnerable claimants to a person early.
Michael Krikheli
Co-Founder & CTO, Five Sigma
Michael Krikheli co-founded Five Sigma to bring AI-native automation to P&C claims past rule-based automation into agentic AI that handles the routine work on every claim, with people engaged where judgment matters.
Related resources
- Part 1 of this series: Agentic AI vs STP: Why STP Rate Stalls Near 10%
- Blog: From AI-assisted to agentic AI claims: where the adjuster’s role moves
- Blog: AI-assisted vs agentic AI claims management: what actually changes
- Clive AI solution overview: Clive AI agents and use cases
