Workers Comp AI Crosses From Claim Summaries to Next-Best Actions

A model that tells an adjuster which file to escalate today changes when money moves on that file, and money moving on a different schedule is a calendar-period effect in a loss triangle. Deloitte put 70% to 80% of workers compensation claims processing in the range of high automation (Deloitte 2020, cited in WCRI, June 2025). Reserving methods assume the handling process behind the diagonals held still.

The claim that workers comp AI has crossed from documentation into decisions arrived this month from three vendors. Insurance Journal reported on September 7 a Sollers Consulting white paper written with Guidewire and CLARA Analytics that lists predictive triage, fraud detection and next-best-action recommendations alongside claim summarization as live use cases.

Key Takeaways

• 70% to 80% of claims processing falls in Deloitte's high-automation range, and that population is concentrated in the low-complexity claims whose closure timing sets the first two diagonals of a workers comp triangle.
• 25 jurisdictions plus four more had adopted the NAIC model bulletin or their own AI guidance by August 6, 2026, none of which produces a claim-level key an actuary can join to a transaction record.
• A 5.0% average cut in approved bureau loss cost level took effect across NCCI states in 2026, selected off development data that already contains whatever handling changes carriers made in 2024 and 2025.
• A 91 calendar-year combined ratio for private carriers in 2025, up from 86 in 2024 on net written premium of $41.6 billion, leaves less room for a development pattern that turns out to be partly a workflow artifact.
• 77% of insurers reported being in some stage of AI adoption in 2024, up from 61% in 2023, so adoption is itself a moving calendar-year variable rather than a fixed state of the book.

What the Paper Puts on the Record

The authors are Jeffery Kaczynski of Sollers Consulting, Oliver Winkenbach of Guidewire and Mubbin Rabbani of CLARA Analytics: the integrator, the core platform and the claims-AI vendor. They supply the workflow. None of them reserves the book that the workflow produces.

Rabbani's line in the launch release is the one with actuarial content: "AI-driven claims intelligence can help carriers identify escalation risks earlier, improve reserve accuracy, and support faster intervention" (Sollers Consulting, August 2026). Carrier Management reported the paper on August 19 under the same orchestration framing.

The paper itself sits behind a registration form, so its described contents here come from those two trade reports and the authors' own release rather than from the document. What can be checked independently is whether the frontier it describes is real, and the best evidence is not a vendor's.

WCRI's June 2025 study, built on 34 interviews across 20 organizations including ten insurers and third-party administrators and ten state agencies, found the same use cases and described one of them in reserving terms. On settlement-timing prompts, the report is explicit: AI can identify when a claim is ready for settlement and nudge the adjuster to pursue it, and "this use can help close claims earlier, reducing the open claim inventory and associated reserves" (WCRI, June 2025).

Those two sentences describe different things. Improved reserve accuracy is a statement about estimation error against a fixed underlying process. Earlier closure and lower open inventory is a change to the process itself.

How a Recommendation Lands in the Triangle

The governance burden splits cleanly at the point where a model output becomes an input to a handling decision. A summarizer that condenses a 100,000-word claim file changes how long an adjuster reads. A triage score that assigns a nurse case manager in week two rather than week ten changes when medical payments start, when the case reserve is set at its realistic level, and how the claim looks at 12 months.

Ultimate severity can be entirely unchanged and the diagonals will still move. Automated preauthorization that approves a guideline-compliant request in seconds instead of days pulls paid medical forward. Litigation-risk scoring that triggers early outreach shifts attorney involvement, which is one of the strongest drivers of workers comp claim duration.

Use-case tier	What the model touches	What moves in the data
Summarization and intake	Adjuster reading time, document handling	Expense timing; little effect on indemnity or medical diagonals
Predictive triage and escalation	Case manager assignment, claim routing	Case reserve adequacy at early maturities; reported-to-paid lag
Automated utilization review	Treatment authorization speed	Paid medical emergence pulled forward within the accident year
Fraud and investigation flags	Which claims get investigated, and when	Denial and reopen patterns; late-emerging disputed claims
Next-best-action and settlement prompts	Offer timing, closure decisions	Closure rates by maturity; open inventory; the shape of the tail

An age-to-age factor cannot distinguish a claim that developed differently from a claim that was handled differently. The standard corrections exist, and Berquist-Sherman adjustments for shifts in case reserve adequacy and closure rates are the right family of tools, but they need a dated change-point and an identifiable affected population.

An AI rollout supplies neither by default. It is phased by adjuster team, claim type and state; the recommendation acceptance rate moves week to week as adjusters learn what to override; and the vendor ships model updates on its own release cadence. That is a gradual, non-uniform, partially unobserved intervention spread across several calendar periods.

The measurable version is a small set of series that a carrier either logs or does not: recommendation acceptance and override rates by adjuster cohort, reserve-change timing distributions before and after deployment, reopen rates, litigation conversion, and adverse-outcome exceptions. Each is a candidate covariate for a change-point test on the carrier's own triangles.

The pricing side is not waiting for that test. NCCI's approved bureau loss cost level fell 5.0% for 2026 across its states, weighted by effective date and premium, extending an unbroken run of annual decreases since 2014. Those selections come off industrywide development data, so the exposure is not confined to whichever carrier deployed first.

The Metadata a Reserve Review Cannot Join

Every control listed above depends on one thing that usually does not exist: a model version and a workflow version stamped on the claim transaction, with effective dates, so a reserving actuary can partition a triangle by which regime handled which claim.

That history typically lives in the vendor's platform rather than the carrier's data warehouse, and it changes when the vendor ships. A carrier can know it deployed a triage model in 2025 without being able to say which of its 2025 claims received a recommendation, which adjuster accepted it, and which version produced it.

Regulation is not filling the gap. Twenty-five jurisdictions had adopted the NAIC model bulletin on the use of artificial intelligence systems by insurers as of August 6, 2026, with California, Colorado, New York and Texas issuing insurance-specific guidance instead (NAIC, August 2026). Those bulletins require documented AI programs, model inventories and testing, written for market-conduct supervision of consumer outcomes. An inventory of models is not a join key to a claim.

The feedback loop closes on itself from there. Models trained on prior adjuster decisions reproduce prior handling practice, and WCRI flags the representativeness trap directly: higher rates of legal representation in one industry or employer type do not imply similar rates once the mix of employers changes. Widespread adoption then alters the development data used to recalibrate the next model generation.

The cushion available to absorb a mistake here is finite and shrinking. NCCI put the industry reserve position at $14 billion redundant at year-end 2025, roughly 12% of carried reserves and down from $16 billion a year earlier (NCCI, May 2026). That redundancy is measured against development patterns that now record workflow policy alongside claim severity, and a workflow reversal, whether a vendor model update or a state restricting AI in claim decisions, moves the diagonals again with no accident-year event behind it.