How VA, SSA, and Private Insurers Use AI for Disability Claims

“AI for disability claims” now covers three operating realities that should not be treated as one market. At the VA, AI and automation are tied to a measurable production surge: average processing time fell from 141.5 days to 80.7 days, while the disability compensation and pension backlog declined 73% from its peak, according to GAO’s 2026 review.[1] At the same time, the VA was operating through workforce pressure, including the reported loss of about 2,700 examiners during the same period.[2] That is not a small efficiency story. It is a capacity story, an accuracy story, and a governance story happening at once.

The SSA looks slower, but the more precise reading is that it is more constrained: aging infrastructure, a large disability backlog, long hearing waits, and limited public documentation around AI-supported tools make its trajectory hard to compare directly with the VA. Private disability insurers, by contrast, appear commercially ahead, with broader use of AI-enabled segmentation, document review, and workflow routing, but their strongest performance claims are often vendor or carrier case-study claims rather than independently audited public metrics.

The VA Shows What Scale Looks Like Before Governance Catches Up

The VA is the clearest place to see the AI impact on disability claims because it has enough volume, public reporting, and oversight to expose both the operational gain and the mess underneath it. The average processing-time drop from 141.5 to 80.7 days is meaningful inside a claims operation. It means fewer files sitting untouched, fewer stale evidence packets, fewer veterans waiting through another monthly cycle without a decision. A 73% backlog reduction is also not a dashboard vanity metric when the queue itself has been a policy failure for years.[1]

But the VA numbers also show why speed alone is a poor proxy for success. GAO reported that the VA’s issue-based accuracy rate was 94.02%, below the agency’s 98% target, a target the VA had not met for more than a decade. GAO also cited an 84% claims-packet accuracy figure, meaning about one in six claim decisions had at least one error under that measure.[1] In a production environment, those two accuracy figures describe different kinds of pain. Issue-based accuracy can look tolerable while packet-level errors still push work downstream to correction, appeal, rework, or claimant frustration.

That is where AI changes the labor map rather than simply removing labor. A missing-document flag can save an examiner from hunting through a packet. An extraction tool can pre-fill evidence. A rating recommendation can push a case toward faster disposition. But if the extracted fact is wrong, or if the model gives confidence without an audit trail, the work does not disappear. It moves to the examiner who must reconcile the machine output, the reviewer who must catch the error, the appeals unit that inherits the dispute, or the claimant who has to challenge the result.

The VA’s modernization push includes major spending as well as operational tools. In 2024, the VA awarded IBM a $485 million contract for an Automated Decision Support system intended to help process benefits claims.[3] The agency has also described using automation and AI to support record claims processing.[4] Those investments sit beside controversy over the Smart Ratings Recommendation tool, which was paused after concerns about how automated recommendations might affect disability decisions.[4] The important distinction is that a paused tool is not evidence that all automation failed. It is evidence that decision support becomes politically and operationally sensitive once it approaches the rating judgment itself.

The closer AI gets to the compensability decision, the more the evidentiary burden changes. A dashboard that shows aged inventory is one thing. A tool that extracts a blood-pressure reading or recommends a rating is another. The VA Office of Inspector General found problems in automated hypertension evidence extraction, with reporting that described a 27% error rate in that extraction process.[5] Hypertension may sound like a narrow condition category, but the operational lesson is broader: when the machine reads medical evidence incorrectly, downstream quality controls must be strong enough to catch the error before it becomes a benefit decision.

GAO’s review is blunt about the governance side. It identified 43 recommendations related to VA modernization and AI adoption, with 28 implemented as of the report, and it found that the department still lacked a comprehensive AI governance strategy.[1] That matters because AI governance is not a policy binder sitting next to the model. It determines who can approve a tool, what documentation examiners see, how performance is monitored after deployment, what happens when accuracy slips, and whether appeals data is treated as an early warning system rather than a separate problem.

The VA case is therefore not a clean success story or a failure story. It is the most useful public case because the gains and weaknesses are both visible. A claims organization can reduce average processing time and still miss its accuracy target. It can cut a backlog and still create new pressure on quality review. It can invest hundreds of millions of dollars in decision support and still need a better way to explain where human judgment begins, where model output ends, and who is accountable when the two conflict.

SSA Is the Capacity Warning, Not Just the Slow Track

The SSA is harder to write about because the available evidence does not support the same level of operational detail. Its AI-related tools, including HeaRT, IMAGEN, and Quick Disability Determinations, are discussed in agency and secondary materials, but the public record does not provide an independent audit trail comparable to the VA’s GAO and OIG scrutiny. That gap should not be filled with speculation. The safer conclusion is narrower: SSA’s AI posture is shaped by severe institutional constraints and incomplete public documentation.

Those constraints are substantial. Reporting on SSA’s AI push cited an average 233-day wait for a hearing, an initial denial rate of about 60%, and a disability backlog of about 1.4 million cases.[6] The agency has also faced staffing reductions, with the research record citing 7,000 jobs cut, while working through modernization efforts that included consolidating 54 legacy systems into one.[7] For a claims operation, that combination changes the meaning of “AI adoption.” A model may be technically useful, but it still has to live inside intake systems, evidence workflows, staffing realities, hearing queues, and policy review processes that determine whether its output can be acted on safely.

The National Academy of Social Insurance Task Force’s Phase One report, published in April 2025, framed AI and emerging technology in disability benefits around principles and safeguards rather than a claim that SSA already has a mature AI operating model.[8] That is the right level of caution. In the SSA environment, the primary AI question is not whether a document classifier or prioritization tool can improve a local workflow. It is whether the agency has enough technical, staffing, and governance capacity to evaluate the tool, document its use, monitor errors, and keep claimants from becoming the place where unresolved system risk shows up.

The SSA also shows why disability claims AI cannot be evaluated only at the point of initial decision. A high initial denial rate means many claimants experience the system through reconsideration, hearings, or appeals. If an AI-supported tool changes which claims are fast-tracked, which evidence is highlighted, or which cases receive additional review, the effect may not be visible in the first processing metric. It may appear later as hearing delay, remand, claimant attrition, or staff rework. Without stronger public documentation, those effects remain difficult to measure.

Private Insurers Look More Mature Commercially, Less Transparent Operationally

Private disability insurers are not waiting for a single public-sector modernization program. They are buying, integrating, and testing commercial AI across claim segmentation, document review, correspondence analysis, medical summarization, fraud signals, and workflow routing. Industry compliance coverage in 2026 described widespread use or integration of AI in insurance claim handling, including a 77% adoption or integration figure.[9] EvolutionIQ has also been described as serving a majority of U.S. disability carriers, a market-penetration claim that signals how quickly specialized disability-claim AI vendors have moved into the carrier workflow.[9]

The commercial promise is straightforward: identify which claims need immediate human attention, which can move through standard workflows, which require missing documentation, and which may become costly or complex. Tata Consultancy Services has published a disability-insurance AI case study claiming more than 95% accuracy and a 40% cost reduction.[10] Those numbers are worth noticing, but they should not be lined up against VA accuracy metrics as if they were the same species of evidence. The TCS figures are self-reported case-study claims, not a public audit by an inspector general or a federal accountability office.

That asymmetry is not a minor footnote. If a private insurer says a model is accurate, the claims professional needs to know accurate at what: document classification, benefit-period prediction, likelihood of return to work, fraud referral, missing-evidence detection, denial recommendation, or payment calculation. A model can be highly accurate at sorting claims into buckets and still create unacceptable risk if the bucket changes the level of human review. A cost-reduction metric can reflect fewer touches per file, but it does not by itself show whether appeals increased, whether vulnerable claimants were disproportionately affected, or whether examiners had enough information to challenge the model.

Reinsurers and industry commentators have been more cautious than the promotional material sometimes suggests. RGA has described disability-claim segmentation and ethical AI as active industry topics, while also warning that AI can be both a boon and a bane in disability insurance.[11][12] The caution is practical: mid-size carriers may face cost and implementation barriers, and all carriers face ethical, legal, and explainability questions when AI affects claim handling. The private market may have more mature tools, but maturity in procurement is not the same as maturity in accountability.

Regulators are now trying to close part of that gap. The NAIC adopted its Model Bulletin on the use of artificial intelligence systems by insurers in December 2023, and the research record indicates that more than 24 states had adopted it or related guidance by 2026.[13] The bulletin is principles-based; it does not create one uniform national audit regime. Separately, at least six states have enacted or advanced human-in-the-loop laws or requirements affecting health or disability-related insurance decisions, including Florida, Arizona, Colorado, California, Illinois, and Texas, according to the research record. The legal direction is visible even if the details remain fluid.

Litigation is also starting to pull AI claim tools into discovery. In Lokken v. UnitedHealth, a March 2026 ruling allowed discovery into the insurer’s use of AI-related denial tools.[14] That ruling is not a final judgment that the tools were unlawful, and it should not be treated as one. Its importance is procedural and strategic: plaintiffs are asking for the machinery behind the decision, and courts may allow them to look. For insurers, that turns model documentation, vendor governance, human-review records, and escalation rules into litigation assets or liabilities.

The private-insurer track therefore has a different risk profile from the VA. It may move faster commercially, and it may have stronger incentives to automate repetitive claim work. But unless carriers can show how AI outputs are audited, overridden, monitored, and explained, their most impressive efficiency claims will remain vulnerable to regulatory skepticism and discovery pressure.

The Metrics Do Not Travel Cleanly Across Systems

Comparing the VA, SSA, and private insurers is useful only if the evidence is labeled honestly. A 42% VA processing-time reduction is an agency-level operational result under public oversight.[1] A private vendor’s 40% cost-reduction claim is a commercial case-study result.[10] A 233-day SSA hearing wait describes system capacity, not AI performance.[6] These numbers can sit in the same article, but they cannot bear the same interpretive load.

Accuracy is the most tempting metric to misuse. The VA’s 94.02% issue-based accuracy rate and 84% claims-packet accuracy rate are imperfect, internally generated measures, and critics have questioned the limits of the agency’s self-review process.[1][5] Still, those figures are exposed to public challenge. Private-sector accuracy claims may be technically sound within a particular implementation, but without independent review, sample details, error definitions, and claimant-impact analysis, they cannot answer the same policy question.

The same problem applies to speed. Faster processing can mean better evidence routing, fewer idle days, and reduced claimant hardship. It can also mean that uncertainty is being pushed into post-decision review. GAO’s testimony described a “Whac-a-Mole” dynamic in which backlog reduction can shift pressure elsewhere in the system, including appeals.[1] Anyone who has watched a claims queue long enough knows the pattern: a metric turns green in one unit, and another unit quietly absorbs the exception work.

That is why governance has to be operational, not abstract. A useful AI governance framework should specify who owns model approval, who monitors drift, who sees the audit trail, how human overrides are captured, how appeal outcomes feed back into model evaluation, and when a tool must be paused. Health systems and insurers building these structures can borrow from broader clinical AI oversight disciplines, including governance committee models such as a clinical AI governance committee charter, but disability claims require their own attention to benefit rights, evidence standards, appeal pathways, and claimant vulnerability.

What Policy and Investment Decisions Should Take From the Three Tracks

The first implication is that buyers and policymakers should stop asking whether AI “works” in disability claims as a general question. The VA evidence supports a narrower and more useful answer: AI and automation can help move large disability-claim inventories faster, but speed gains can coexist with persistent accuracy shortfalls and incomplete governance.[1] That is enough to justify continued investment in some tools, and enough to require stronger audit controls before expanding others.

The second implication is that SSA modernization cannot be solved by dropping AI onto fragile infrastructure. If hearing waits, staffing reductions, legacy-system consolidation, and limited tool documentation define the environment, then AI investment has to include measurement capacity, staff training, data quality, claimant safeguards, and public reporting. Otherwise, the agency may add technical complexity without improving institutional capacity.

The third implication is for private insurers and their vendors. Commercial disability-claim AI will face less patience for black-box performance claims as state laws, NAIC-aligned expectations, and discovery fights mature. A carrier that can document human review, model limits, override behavior, adverse-impact monitoring, and appeal feedback will be in a stronger position than one that can only cite accuracy and cost savings from a vendor deck.

AI is already reshaping disability claims, but its apparent success depends on which system is being measured, which metric is being privileged, and who has authority to audit the result. The VA shows measurable acceleration with unresolved accuracy and governance problems. SSA shows how institutional capacity can limit AI’s usefulness before performance can even be judged. Private insurers show commercial maturity under rising accountability pressure. Those are three tracks, not one verdict.

References

1. GAO-26-109137 (VA Disability Benefits: Opportunities and Challenges to Modernizing Technology and Adopting AI) — GAO.
2. VA Pushes AI for Disability Claims Amid Workforce Cuts — Legis1.
3. VA Uses Automation, AI to Process Record Benefits Claims — GovCIO Media & Research.
4. Automating The Backlog: AI And The Future Of Military Claims — Military.com.
5. VA Using AI, Mandatory Overtime to Process Claims Backlog — The War Horse.
6. SSA's Push to Incorporate AI Into Decision-Making — London Disability.
7. NASI Task Force Issues Report on AI at SSA — Empire Justice Center.
8. Task Force on Artificial Intelligence, Emerging Technology, and Disability Benefits: Phase One Report — Digital Government Hub, April 2025.
9. Navigating AI and Claim Handling in 2026 — Enlyte.
10. Adopting AI Technologies to Transform Disability Insurance Claims — TCS.
11. Hot Topics in Disability Claims: Segmentation and Ethical AI — RGA.
12. The Dual Role of AI in Disability Insurance: A boon and a bane — RGA.
13. Use of Artificial Intelligence Systems by Insurers — NAIC, December 2023.
14. Court Allows Discovery Into Insurer's Use of AI to Deny Claims — Hunton Insurance Recovery Blog.