AI Claims Decisioning Framework 2026: What US Claims Leaders Told Us About How Decisions Actually Get Made

I had the same conversation with US claims leaders. It started the same way every time. I would ask: “How are claims decisions actually made at your organization today?”

Not the conference keynote answer. Not the analyst presentation answer. The honest one.

What I heard didn’t match the dominant industry narrative. So we ran a structured survey to check whether my conversations were a pattern or just noise. Twenty respondents - Chief Claims Officers, VPs of Claims, Heads of Claims, COOs, Claims Technology Leads, and Senior Claims Adjusters at carriers from under $250M to over $20B in direct written premium. Personal lines, commercial, specialty, multi-line. We called the result The 2026 Claims Decisioning Pulse.

The findings reshaped how I talk to carriers about AI claims decisioning. Five signals stood out. The most important one isn’t about technology at all - it’s about a regulatory framework that most claims AI vendors aren’t discussing with their carriers.

If you’re a VP of Claims or CCO at a US P&C carrier planning Claims AI deployments for 2026 or 2027, the framework I’ll walk through below changes how you should evaluate vendors, sequence capabilities, and structure your ROI model. It also changes what “fully autonomous claims decisioning” actually means in the US market - which, as it turns out, isn’t what the slide decks suggest.

Download the full 17-page Pulse 2026 report (free, no paywall)

What 20 US claims leaders told us about today’s AI claims decisioning reality

I’ll start with the headline finding, because it sets up everything else.

Of the 20 respondents in our sample, 17 still make meaningful claims decisions manually or with workflow support only. Eleven described their process as “mostly manual - an adjuster reviews every claim, a supervisor approves, the system records the outcome.” Six described “workflow-automated - the system routes and tracks, but humans make every meaningful decision.” Three respondents - all Claims Technology / Digital Claims Leads at personal lines and multi-line carriers - described an AI-assisted decisioning model where AI scores or flags claims while a human still decides each one. A subtle pattern worth flagging: the leading edge of AI-assisted decisioning in our sample is being driven by claims technology teams, not by C-suite mandate.

Zero respondents reported autonomous claims decisioning at scale.

Read alongside Deloitte’s 2024 survey of 200 US insurance executives, which found that scaling generative AI from pilot to production remains the single biggest gap between intent and reality in the industry, this seems to confirm the conventional story: carriers are behind on AI maturity, and the work ahead is mostly about pushing pilots into production [1].

That story is partially right. But it misses the more important reason these numbers look the way they do, and that reason is regulatory, not technological.

In my experience working with VPs of Claims at US P&C carriers, the “AI maturity gap” framing leads to bad deployment decisions. It tells carriers they’re behind, so they should move faster. It tells boards to fund autonomous decisioning pilots. It tells vendor evaluation committees to score capabilities that touch the decision itself. All three of these are wrong in the US market, for reasons we’ll unpack below.

The regulatory framework most vendors don’t mention

The NAIC Model Bulletin on the Use of Artificial Intelligence Systems by Insurers, adopted in December 2023, directs state regulators to assess - among other factors - “the extent to which humans are involved in the final decision-making process” when evaluating a carrier’s AI program [2]. As of early 2026, 24 US states have adopted the bulletin or substantively similar guidance, covering more than half the country [3].

That phrase - extent of human involvement in final decision-making - is the foundation of the regulatory framework I want to walk through. Let me translate what it means in operational terms.

Decisions adverse to the consumer - denial of coverage, reduction of payout amount, claim closure without payment, fraud-flagged outcomes - sit under heightened regulatory scrutiny in the US market. Counsel advising carriers on AI regulation puts it plainly: every AI-enabled decision, especially an adverse one, will be examined for bias, interpretability, and procedural fairness. Regulators expect human oversight on those decisions [4].

State-level laws are reinforcing this further. California Health & Safety Code restricts health care service plans from relying solely on automated tools in claims decisions. Florida enacted legislation in 2025 requiring that a qualified human professional review and make the final determination when AI is used to evaluate insurance claims. Colorado’s SB 24-205 (the Colorado AI Act, effective February 2026) requires consumer disclosure and human oversight for high-risk AI systems including claims. New York, Connecticut, Pennsylvania, Texas, Virginia, Georgia, and Nebraska have followed with similar frameworks [5].

So when you read that 17 of 20 respondents in our sample still make decisions manually or workflow-only, the right interpretation isn’t “they’re behind.” It’s “they’re aligned with where the regulation is pushing.” Adverse decisions require human-in-the-loop review by design, not as a maturity stage carriers will grow out of.

This changes the strategic question for VPs of Claims. The question isn’t “when will we achieve autonomous claims decisioning.” It’s “what does AI do before the decision, and how cleanly is that contribution documented when an examiner asks?”

The STP bifurcation: where AI works and where it doesn’t

Once you accept the regulatory framework above, straight-through processing (STP) in US claims has to be understood as a bifurcated discipline. There are two distinct paths through your claims pipeline, and AI’s role differs sharply between them.

The approve path

A routine claim - one that clears all coverage rules, documentation checks, and fraud signals - can be paid out and closed without a human touch. This is where STP works end-to-end. This is also where the measurable ROI of Claims AI lives.

I worked with a Northeast specialty carrier whose VP Claims targeted a 30-40% STP rate on the approve path in 18-24 months. The economics on that path are straightforward: every claim that runs through STP saves adjuster time, reduces cycle time variance, and improves customer satisfaction by closing claims faster. The J.D. Power 2025 U.S. Property Claims Satisfaction Study found that customer satisfaction for claims closed within 10 days scored 762 out of 1,000, dropping 167 points to 595 when repairs took more than 31 days [6]. STP on the approve path is what drives those numbers in the right direction.

The adverse path

A denied claim, a reduced payout, a fraud-flagged outcome - these route to an adjuster or supervisor by design. AI can advise, score, and flag along this path, but a human must approve the adverse outcome.

I’d require any Claims AI deployment to log every AI contribution on this path with explainable reasoning - what inputs the model received, what features influenced the output, where the human reviewer overrode (if anywhere), and what the final outcome was. This isn’t a future audit requirement. It’s what the NAIC AI Systems Evaluation Tool pilot examinations are checking right now in 12 US states [7].

Why this bifurcation matters for vendor evaluation

In my experience working with Heads of Claims evaluating Claims AI vendors, the question I’d recommend asking first isn’t about model accuracy or integration timelines. It’s this: “What’s your STP rate breakdown between the approve path and the adverse path?”

If your vendor pitches you a unified 60% STP rate without that breakdown, one of two things is happening. Either they’re measuring approve-path-only STP and presenting it as total, in which case the number is misleading but not dishonest. Or they’re including adverse decisions in the STP rate, in which case their product either isn’t built for US adverse-action requirements or isn’t being deployed with regulatory compliance in mind.

I’ve worked with carriers who took unified STP projections at face value and built business cases around them. When the regulatory reality hit during deployment, the achievable STP rate was 30-50% lower than projected. That gap is the difference between a Claims AI program with positive ROI and one that gets cancelled in year two.

What the Pulse 2026 data shows about where AI actually lives

Let me come back to the research. When we asked respondents to mark every area where AI or material automation is in production at their company today, the responses clustered on the periphery of the claims process - not at the decision itself.

The most-deployed capabilities were:

Seven of the 20 respondents told us they have none of the listed AI capabilities in production today. Three of those seven run carriers with more than $1B in DWP, two operate above $5B, one above $20B. AI maturity in US claims isn’t gated by company size; it’s gated by which capabilities a carrier has decided to deploy first.

Of the thirteen respondents running some AI, what’s striking is that the five most-deployed capabilities - document understanding, fraud signals at FNOL, FNOL intake, triage, and customer communications - are exactly the places where AI accelerates the adjuster before the decision (or follows it), without taking the adverse-action call away from a human. Reserve recommendation and decisioning lag, because that’s where regulatory scrutiny is sharpest.

This isn’t an accident. The carriers running AI in production have figured out the bifurcation framework intuitively, even if they haven’t articulated it. They’re deploying AI on the data-gathering and triage work that consumes adjuster capacity, while keeping adjusters at the decision point where regulators require human involvement.

The data-collection tax: where adjuster time actually goes

Here’s the finding from Pulse 2026 that reshaped how I talk about Claims AI ROI.

We asked respondents to identify the biggest gap between how claims decisions get made today and how they’d want them made. The answers were striking. Twelve of 20 respondents - close to two-thirds of the sample - flagged adjuster time spent gathering data versus evaluating the claim as their largest operational gap. That beat real-time visibility (9 of 20), cycle time (8 of 20), decision consistency (8 of 20), audit trail effort (7 of 20), reserve accuracy (6 of 20), catastrophe-event scalability (5 of 20), and catching fraud before payment (5 of 20).

I call this the data-collection tax. Adjusters spend their day pulling policy details from the policy admin system, checking ISO ClaimSearch for prior claims, downloading documents from intake, reviewing third-party reports, cross-referencing fraud signals, checking telematics data on auto claims, and reading email threads with claimants. Then they make a decision in the time that’s left.

In my experience working with VPs of Claims, the bottleneck in US claims operations isn’t adjuster judgment. It’s the work that has to happen before judgment can be applied. Cycle time is the metric the CEO asks about, but the data-collection tax is what causes it.

I worked with a Midwest regional auto carrier whose Head of Claims tracked adjuster time allocation for a week before launching a Claims AI evaluation. The shadow study found adjusters spent roughly 62% of their day on data gathering and 18% on actual claim evaluation - the remainder went to communication and administrative tasks. That ratio is what made the business case for document understanding and FNOL intake automation defensible to the CFO. Without the time-allocation baseline, the business case would have been hand-waving.

This reframes the ROI conversation. The right Claims AI investment isn’t an attempt to automate the decision. It’s an attempt to give the adjuster their afternoon back. Document understanding, FNOL intake automation, fraud signal scoring at first notice - these capabilities attack the data-collection tax. They also happen to be deployable without touching the regulatory danger zone around adverse decisions.

A former VP of Claims at a small multi-line carrier with both P&C and life lines put it well in the open-ended section of our survey:

“Approximately 80% of tasks are straightforward, while the remaining 20% consist of complex, one-off cases. Technology is utilized at multiple stages of the process; however, existing tools typically support only selected parts rather than the end-to-end workflow.”

That 20% is where adjusters earn their salary. That 20% is where carrier reputation lives. That 20% is also where DOI exam findings tend to come from. The implication: automate the boring 80% end-to-end so adjusters have time for the 20% only humans can do well.

Data quality leads the blocker list - and ROI uncertainty isn’t far behind

Here’s where the framework gets uncomfortable for vendor pitch decks.

When we asked respondents to name the single biggest blocker stopping them from closing the gap between today’s process and their target state, the answers shifted the conversation away from where vendors usually focus.

Data quality and fragmented systems came in first - seven of 20 respondents identified it as their single biggest blocker. ROI uncertainty came in second at five of 20, followed by internal IT capacity at four. Two new blocker categories appeared in our sample that hadn’t surfaced in early conversations: vendor selection (one respondent: “no clear leader for our use case”) and regulatory governance (one respondent: NAIC AI Model Bulletin compliance and DOI exam exposure).

The right interpretation isn’t that ROI doesn’t matter. It’s that data quality and ROI are the two binding constraints, and they bind together. A VP of Claims who can’t show the CFO a defensible LAE-reduction model with a 24-month payback won’t get a Claims AI project funded. But the reason that model isn’t defensible is usually the same reason the deployment will struggle in production: data is fragmented, quality is inconsistent, and the ROI projection rests on data foundations the carrier doesn’t yet have.

I’d argue this is the most important finding in the report for board-level audiences. The dominant industry narrative for three years has been “the technology is ready, the obstacles are integration.” This sample says something different: the technology is ready, the obstacles are data, governance, and economics - in that order.

A VP of Claims who can’t show the CFO a 24-month payback model with defensible LAE reduction assumptions doesn’t get the project funded - no matter how good the vendor’s tech stack is. And here’s the kicker: business cases built on vendor-projected STP rates that include adverse decisions are projecting fictional numbers under US regulatory reality. When finance discovers the gap, the project dies.

In my experience working with Heads of Claims, the way out of this is to sequence Claims AI deployments around capabilities where value is measurable in 90 days, not 24 months, AND where the underlying data is most structured. Document understanding on first-party auto is usually the right entry point. You can time the adjuster productivity improvement directly. You can show finance the LAE delta. You can build the measurement infrastructure alongside the deployment and use that data to fund the next capability.

I’d require any Claims AI program in 2026 to have a measurement plan in place before go-live, not after.

The audit-trail exposure no one is preparing for

The fifth signal from Pulse 2026 is the one that should keep claims executives awake at night.

We asked: if a state DOI examiner asked for a complete decision audit trail for the AI or automation already in production at your company, how confident are you that you could produce it within 5 business days?

Of the 14 respondents exposed to the audit-trail standard - those with AI in production or who otherwise engaged with the question rather than marking N/A - only three said they could produce a complete audit trail with high confidence. That’s about 21%. The other 11 anticipated significant manual reconstruction or expected to struggle to meet the deadline.

This is real exposure. The NAIC AI Systems Evaluation Tool - designed to standardize how state examiners assess insurer AI compliance during market conduct examinations - entered pilot examinations in 12 US states in Q1 2026. By the end of 2026, the framework will expand. The first published findings from those exams are expected in 2027.

Carriers running AI in production today are 12-24 months away from being asked to produce documentation that, in roughly four-fifths of the cases in our sample, doesn’t exist in queryable form yet.

I’ve worked with carriers who underestimated this exposure. The day the regulator asks for documentation, “we’ll figure it out” is the most expensive answer in the building. Retrofitting audit trail capability after a market conduct finding costs 5-10x what it costs to build it in from day one.

I worked with a Southeast P&C carrier whose VP Claims learned this the hard way during a routine market conduct examination. The carrier had deployed an AI triage capability 14 months earlier without explicit audit trail logging at the decision level. When the DOI examiner requested documentation for 25 claims spanning the deployment period, the team spent six weeks reconstructing decision rationales from disconnected systems. The remediation work cost roughly four times what it would have cost to build audit trail capability into the original deployment. The lesson: audit-readiness is not a phase you add later. It’s a design constraint from day one.

Three questions I’d require any claims executive to ask of their AI deployment this quarter:

1. For every AI-supported decision in production, can we produce a complete log of inputs, model outputs, human overrides, and final outcome - in queryable format, not as a manual reconstruction?

2. For decisions adverse to the claimant, can we produce documentation showing exactly where a human reviewed and approved?

3. If a DOI examiner asks for all of the above for 100 claims across our NAIC-adopted states, can we hit a 5-day SLA without pulling adjusters off active work?

If any answer is “we’d struggle,” audit trail infrastructure is your next Claims AI investment - not another capability.

What this means for sequencing your 2026 Claims AI deployment

Pulling the framework together, here’s how I’d recommend Heads of Claims sequence Claims AI deployments based on what Pulse 2026 surfaced.

Phase 1 (Months 1-6): Deploy where measurable in 90 days.

Start with document understanding on first-party auto claims, or FNOL intake automation, or fraud signals at first notice. These capabilities accelerate the adjuster before the decision, so they don’t touch the regulatory danger zone. They’re also where ROI is fastest to measure. Build the measurement infrastructure alongside the capability.

Phase 2 (Months 4-12): Expand pre-decision automation.

Add triage and routing, then customer communication automation, then vendor orchestration. Each layer builds on data infrastructure from Phase 1. Each layer continues to give adjusters time back without changing who makes the decision.

Phase 3 (Months 9-18): Reserve recommendation with full audit trail.

Reserve recommendation is the first capability that touches the decision. Deploy it only when (a) you have measurement infrastructure proving value from Phases 1-2, and (b) you have audit trail capability in place from day one. Treat reserve recommendation as advice to the adjuster, not as automation of the decision.

Phase 4 (Months 18+): Adverse path documentation.

If and when you deploy AI on the adverse path, the deployment is primarily about documenting human-in-the-loop review, not about automating the call. AI scores the claim. AI flags the indicators. The adjuster reviews. The system logs everything. The audit trail produces on demand.

This isn’t the sequence vendor decks pitch. It’s the sequence the regulatory framework and the ROI math actually allow.

What we got wrong and what’s coming next

One honest note. The 2026 Claims Decisioning Pulse is a qualitative pulse check, not a statistically representative benchmark of the US insurance market. Twenty respondents (19 carrier-side claims leaders plus one technology vendor observer) over a three-week window from May 8 to May 29, 2026. That’s not enough to make industry-wide claims, and we’re explicit about that in the report.

What it is good for: surfacing directional signals from senior practitioners that we can triangulate with public industry data from J.D. Power, NAIC, McKinsey, and Deloitte. Where our signals diverge from public data, we say so. Where they align, the pattern is worth taking seriously.

We’re planning a Wave 2 of the research for Q3 2026 with a larger sample. If you’d like to participate or receive the next report, our research team is collecting interest through the report download form below.

Download the full Pulse 2026 report

The full 15-page report includes:

• All five signals with detailed data tables and analysis

• Verbatim quotes from claims leaders on what they wish vendors understood

• Triangulation with J.D. Power 2025, NAIC Model Bulletin, McKinsey, and Deloitte data

• Methodology section with full disclosure of sample composition and limitations

• Five takeaways for 2026 Claims AI planning

The report is free, no paywall beyond a short form gate. We don’t share emails with third parties. The form goes to one person on our research team.

→ Download The 2026 Claims Decisioning Pulse

If the framework above maps to questions you’re working through for your 2026 Claims AI roadmap, I’d be happy to walk through the implications for your specific operation. We offer 45-minute working sessions for VPs of Claims and CCOs at US P&C carriers - no slides, no pitch deck, just a working conversation about which capabilities would deliver measurable value in your first 90 days and how to structure the audit trail from day one.

→ Book a 45-minute Claims AI assessment with me

‍

Sources and citations

[1] Deloitte Insights. “Scaling gen AI in insurance.” December 2025. Based on Deloitte survey of 200 U.S. insurance executives (100 L&A, 100 P&C) conducted in June 2024.

[2] National Association of Insurance Commissioners. “Model Bulletin: Use of Artificial Intelligence Systems by Insurers.” Adopted December 4, 2023.

[3] National Association of Insurance Commissioners. “Implementation of NAIC Model Bulletin: Use of Artificial Intelligence Systems by Insurers.” Updated 2025-2026.

[4] Buchanan Ingersoll & Rooney PC. “When Algorithms Underwrite: Insurance Regulators Demanding Explainable AI Systems.” October 2025.

[5] Baker Tilly. “The regulatory implications of AI and ML for the insurance industry.” August 2025.

[6] J.D. Power. “2025 U.S. Property Claims Satisfaction Study.” March 18, 2025.

[7] National Association of Insurance Commissioners. “Artificial Intelligence and State Insurance Regulation.” Issue brief, March 2026.