Five Observations from the CLM Workers' Compensation Conference

I moderated a panel at the CLM Workers' Compensation Conference last week titled How AI Is Transforming the Workers' Compensation Field, alongside Angela Cabado, Director of Risk Management at Marriott Vacations Worldwide, and Katie Grove, Attorney at MGC Law. The session ran long, and the conversation continued well after the room emptied.

The substance of what I heard in that room, from carriers, TPA`s, defense counsel, and risk managers, did not match the pace of the industry coverage. The questions were sharper. The concerns were closer to operational ground, and change management. And across those conversations, the same themes surfaced repeatedly, largely regardless of company size or maturity. The issues that drew the most attention were not the ones the trade press has been emphasizing.

Five observations are worth recording.

1. The plaintiff bar is already pricing AI into negotiations. The defense side is not.

This is the observation I would put in front of every claims leader who reads this, and it is the one that drew the most visible reaction in the room.

Embroker's 2024 Legal Risk Index Report, based on a spring 2024 survey of 200 US legal professionals, found that roughly 78 percent of law firms had not adopted AI tools at the time of the survey, citing privacy, security, and misuse concerns. That headline obscures the asymmetry that matters: adoption is not evenly distributed across the bar. Several participants described the same pattern from the field, more sophisticated AI-assisted demand preparation emerging from plaintiff firms, particularly around AI medical chronology analysis and inconsistency detection. The plaintiff side is incentive-aligned around leverage and demand velocity, and AI accelerates exactly that work.

This is accountability inversion in its most expensive form. Accountability inversion is what happens when the party that bears the least operational burden gains the greatest informational advantage, and the party carrying the cost is the one least equipped to see it. The defense bar and the carriers they represent are absorbing the cost of an information asymmetry they did not create and are not yet measuring. By the time it surfaces in case outcomes, in settlement values, trial exposure, and the velocity of demand cycles, the gap will have compounded.

Two questions belong on every claims leader's desk this quarter. First: what are our outside defense counsel using, and how do we know? Second: where in our own workflow are we still operating on the assumption that document volume and document insight are the same thing?

2. 40 - 60% of adjuster time is a silent failure the industry is not measuring.

Wisedocs analysis, published with the March 2026 launch of our Claims Decision Intelligence platform, places this figure at between forty and sixty percent of claims professionals' working hours spent organizing, reading, and searching documents before they make a single decision. The number is well rehearsed in vendor decks. What is less rehearsed is what that time displaces.

It is not throughput that suffers. Throughput metrics look fine. Adjuster’s close files. Cycle times come in within standard. The failure is silent because the metrics that would surface it are not the metrics the industry tracks.

What suffers is reserve accuracy. What suffers is the timing of settlement decisions. What suffers is return-to-work coordination, which has direct and quantifiable cost implications and direct human implications for the injured worker. What suffers is the moment of judgment that determines whether a file becomes a litigation exposure or a managed resolution.

Angela Cabado made the point with the clarity that only comes from operating a claims program at scale: the issue is not that adjusters lack judgment. The issue is that the structure of the work spends their judgment on tasks that do not require it. The hours before the decision are eating the decision.

That is the AI opportunity in workers' compensation. Not replacing judgment. Returning the hours that precede it.

3. The summary is the claim. Treat it accordingly.

Katie Grove brought the sharpest framing to the panel, and it is the framing the defense bar needs to internalize quickly.

When an AI medical chronology, summary, or reserve recommendation enters a claims file, it becomes operative. Subsequent decisions are made against it. Reserves are set against it. Settlement positions are anchored to it. By the time it reaches litigation, it is not an artifact of the workflow. It is the claim.

This has two consequences that are not yet evenly understood across the industry.

First, every AI-assisted output must be traceable to its source documentation. Not as a compliance gesture, but as a precondition for the output being usable at all. If a finding in a summary cannot be linked to the page, the paragraph, and the document it came from, the finding is not defensible, and the file built on top of it is not defensible either. Regulation is already moving in this direction. California's SB 574, the first US statute to govern how attorneys and arbitrators use generative AI, codifies the underlying principle: the human professional remains accountable for AI output, must independently verify it, and cannot delegate judgment to the machine. The obligation it places on lawyers, to personally verify every citation an AI tool produces, is the same obligation every claims organization should already be placing on itself. Other states are watching. Opposing counsel is already there.

Second, the question to ask AI vendors is not whether their tools are accurate. Accuracy is necessary and not sufficient. The question is whether the tool produces outputs that can be defended in deposition, audited by a regulator, and reconstructed by a successor adjuster two years later. If the answer to any of those is uncertain, the tool is not yet a fit for the workflow.

4. Trust in AI outputs scales with architecture, not with model performance.

Wisedocs partnered with ALM's Property Casualty360 on a 2025 survey of more than sixty-five claims professionals across the industry, released in August at the Workers' Compensation Institute's Annual Conference. 16% of respondents reported medium or high trust in AI-generated outputs operating without human oversight. That number rose to 60%, nearly four times higher, when experienced professionals remained involved in reviewing and validating those outputs.

The instinct to read this as a finding about AI limitations is wrong. It is a finding about what adoption actually requires.

The organizations gaining real traction with AI in claims are not the ones that handed over the keys. They are the ones that built workflows in which AI accelerates the work, structures the information, and surfaces the patterns, while experienced claims professionals remain central to the decisions that carry consequence. The architecture is what earns the trust. The model performance is downstream of the architecture.

This is also where the industry's preferred metrics start to fail. Trust does not show up in throughput. Throughput will improve whether or not the workflow is defensible. The question is whether the outputs being produced at that throughput can be relied on, audited, and stood behind. That is an outcome question, and it requires outcome measures.

5. The workforce transition rewards judgment. Volume alone is no longer enough.

The retirement wave in the adjuster workforce is real. The institutional knowledge loss is real. The World Economic Forum's projection of claims adjusters, examiners, and investigators among the fastest-declining occupations through 2030 is real.

What is not yet widely accepted is the shape of what comes next.

Volume still matters. Caseloads still have to be managed and staffing economics are still real. But volume alone is becoming an insufficient measure of value. The claims professional whose contribution is measured only in files closed per week is exposed. The claims professional whose contribution is measured in reserve accuracy, return-to-work outcomes, litigation avoidance, and the quality of the decisions made on the most complex files is positioned. The organizations that retrain, retain, and elevate the second category will outperform the organizations that try to preserve the first.

This is the closing observation, and it is the one that most directly contradicts how the industry has been talking about AI. The risk is not that AI replaces adjusters. The risk is that organizations continue measuring the wrong things long enough that the right things stop being protected. By the time the metric catches up to the outcome, the firms that optimized for throughput will have lost the ground that matters.

What this argument has to account for

None of this works if the technology is treated as self-sufficient, and the honest version of this case has to name where it breaks. AI claims processing automation carries real failure modes. Models hallucinate, asserting findings that are not in the record. Optical character recognition degrades on the handwritten notes, faxed records, and poor scans that fill real claims files, and a AI medical summary built on corrupted extraction inherits every error silently. Automation bias is its own hazard: the better a tool looks, the more tempting it becomes to stop checking it, which is precisely when the worst errors enter the file unexamined. And governance in most organizations is immature relative to the speed of adoption, with policies for AI use trailing the actual use by quarters or years.

These are not reasons to wait. They are the reasons the argument of this piece holds. Each failure mode is a silent failure, invisible to throughput metrics and detectable only through traceability, human review, and outcome measurement. The organizations that name these risks and build for them will pull ahead of the ones that either deny them or use them as an excuse for inaction.

It is also worth saying that the industry is not a single curve. Large national carriers, regional carriers, TPA`s, and self-insured employers are moving at meaningfully different speeds, with different data estates, different governance maturity, and different tolerance for risk. The observations here apply across all of them, but the starting line is not the same, and the right next step for a national carrier is not the right next step for a self-insured employer.

The bottom line

The workers' compensation industry has spent decades optimizing the speed of claims processing automation. AI is now exposing the difference between processing work faster and making better claim decisions. Those are not the same thing. Over the next five years, the organizations that internalize that distinction earliest, and rebuild their metrics, their workflows, and their defensibility standards around it, will separate themselves from the ones that mistook motion for progress.

That is the conversation worth continuing. If your organization is working through any of these questions, I would welcome the exchange. Book a call with our team and we'd be happy to show you what this looks like in practice.