Domain Atlas / Caseworker documentation & copilots
VA claims automation (automated survivor-benefit decisions)
In a review issued April 30, 2026 (report 25-00153-47), the Department of Veterans Affairs Office of Inspector General found that at least 8,000 of an estimated 8,100 automated Dependency and Indemnity Compensation (survivor-benefit) granting decisions issued from September 2023 through August 2024 - nearly all - contained at least one legal or procedural deficiency, such as incomplete evidence summaries and omitted favorable findings, with most rating decisions listing only the death certificate as evidence. The OIG separately found that at least 2 percent of the decisions (at least 190) carried monetary-impact legal errors totaling at least 2.7 million dollars (2,727,764 dollars in questioned costs); the roughly 98 percent figure is the share with any legal or procedural defect, not the monetary-error rate. The system, phased in beginning May 2020, extracts data from scanned documents and applies predefined encoded rules to grant service-connected death claims end to end with no human involvement when the rules are met; the OIG describes it as rules-based automation and document extraction, not machine learning, and its figures are outcome statistics from a statistical sample rather than a per-interaction rate.[3]
What happened
Beginning in May 2020, the Veterans Benefits Administration (VBA) phased in end-to-end automation of veterans' service-connected death claims — Dependency and Indemnity Compensation (DIC). Built by VBA's Pension and Fiduciary Service with VA's Office of Information and Technology, the system extracts data from scanned documents such as death certificates and claims applications and applies predefined encoded rules to generate rating decisions, awards, and notification letters, with no human involvement when all automation rules are met. It is deterministic rules-based automation and document extraction — the Office of Inspector General does not describe it as machine learning, and it should not be read as ML scoring. Because the rules are predefined, a rule defect recurs systematically on every matching claim rather than scattering as noise.
An OIG review issued April 30, 2026 (report 25-00153-47), covering September 2023 through August 2024, found that at least 8,000 of an estimated 8,100 automated DIC granting decisions — nearly all — contained at least one legal or procedural deficiency, including incomplete evidence summaries and omitted favorable findings that are statutory requirements; most rating decisions listed only the death certificate as evidence. It is worth separating the two figures the coverage collapses: the roughly 98% "deficiency" count is decisions with at least one legal or procedural defect, while the quantified monetary-error floor is narrower — at least 2% of decisions (at least 190) carried monetary-impact legal errors totaling at least $2.7 million ($2,727,764 in questioned costs), including granting death benefits for conditions unrelated to service without supporting medical evidence and wrong effective dates. Press coverage (Public Radio East, June 2026) framed it as an automation "glitch" that "ruined 98%" of claims; Task & Purpose (May 2026) carried a $22,692 example overpayment and a Disabled American Veterans representative's qualified support for automation only with robust human oversight (the widely quoted ~2,000-claims-per-day figure is that representative's hypothetical framing; VA's own claim, cited in the report, is more than 1,000 DIC payments or adjustments per day). The correction record is the case's spine. In April 2020 a VBA analyst reported through VA's internal defect-tracking system that automated DIC decisions listed only the death certificate as evidence; the Pension and Fiduciary Service closed the defect without action, and the same deficiency was central to the 2026 findings. The external OIG, not the internal channel, forced correction: VA removed the long-form guidance from its manual in March 2025, immediately after the OIG's preliminary briefing — roughly five years after the internal ticket, and the OIG's full public report did not land until 2026, roughly six years after it. The OIG also found the quality-review checklist for automated DIC claims was less rigorous than the review traditional claims receive and did not focus on legal and procedural requirements, and that VA's PACT Act section 701(b) modernization plan to Congress did not fully explain that VBA grants service-connected death claims end to end without human intervention. The program expanded despite persisting errors: the VA Secretary announced expanded DIC automation in May 2025; 20 additional automated decisions from September–October 2025 showed similar errors as of November 2025; recommendation 1 (strengthen the automation) remained open, with VBA concurring only in part and disputing part of the report's methodology.
The same pipeline carried copilot-grade components and parallel human failures that the OIG documented separately. In VBA's Automated Benefits Delivery hypertension automation (report 22-02936-175, September 2023), 27% of reviewed automated claims (16 of 60, December 2021–September 2022) had inaccurate determinations, because contractor-produced automated summary sheets lacked comprehensive blood-pressure information and rating specialists, under unclear guidance on "predominant" blood pressure, failed to resolve inaccurate extracted readings — and ABD leaders told the OIG they were unaware of the decision errors until the review. A Special Monthly Compensation calculator inside the Veterans Benefits Management System for Rating produced wrong results for complex scenarios (limb loss, blindness, aid and attendance): monthly underpayments of $132 to $4,170 and a $373 overpayment in tested scenarios; VBA and VA IT could not determine the cause or how long the errors had occurred, and after a November 2023 allegation VA disabled the calculator in October 2024 and reverted to a correctly functioning legacy calculator (the review was scenario-based and did not identify affected veterans). An April 2025 review of PACT Act effective dates (24-01153-52) found incorrect effective dates on about 31,400 of 131,000 first-year claims (24%), with roughly 26,100 causing at least $6.8 million in improper payments and a $20.4 million three-year projection, and two automated date tools found unreliable. In the surrounding human workflow, an OIG review of PACT presumptive denials (December 2024) estimated 45% of about 19,200 denied claims contained at least one error — human processor errors under confusing guidance, not automation errors — with unwarranted exams in an estimated 6,900 denials costing about $1.4 million and two veterans improperly denied and underpaid about $56,700; a September 2025 review found a senior representative in Philadelphia authorized about 85,300 claims across FY2022–2024 (about 19 times the national average) at 4.7 minutes per claim versus a roughly 21-minute norm, with about 84% of a statistical sample of her January–June 2024 authorizations containing at least one error and an estimated $2.2 million in improper payments from that window; and a May 2026 review found nearly 10,000 unwarranted or poorly justified overrides of VBMS-R software safeguards in six months, with more than $67,000 in erroneous payments and no override quality-review process.
The sociotechnical reading
Almost every automation case in this Atlas turns on a correction step that is missing, overwhelmed, or pushed onto the least-resourced actor. VA is the case where several correction channels existed, all of them internal, and every one failed — while the only channel that ever changed behavior was external and slow. The contrast with MiDAS is exact and instructive: there the corrective force was courts and litigation acting on a punitive fraud system; here it is a government's own Inspector General acting on a benefit-delivering one, where the errors run both ways (survivors under- and over-paid, effective dates early and late), and where the internal quality check was not absent but calibrated weaker for automated claims than for the manual review it replaced. The check was dialed down at exactly the point the human was removed — an April 2020 defect ticket flagged the core deficiency and was closed without action, a quality checklist was documented as less rigorous for automated claims, and a production-incentivized authorizer cleared claims in minutes. None of it caught what a watchdog eventually did, six years later.
Two dynamics make this shape its own. The first is a rate mismatch: the institution expanded automated throughput (a May 2025 announcement) while known errors persisted (confirmed November 2025), and its only working correction arrives on a multi-year cadence (2020 defect to 2026 report). When production is continuous and correction is periodic, the governance question stops being "does a correction channel exist?" — several did — and becomes "does the rate of correction keep pace with the rate of automated production?" On this deployment it did not, so the error stock grew between audits. The second is quieter and lives in the record itself. Because the automated grants wrote thin evidence summaries — many listing only the death certificate, omitting statutorily required favorable findings — each error lowered the odds that the next reviewer, appeal, or audit could catch it. Errors written into memory reduce the system's own capacity to detect them; this is the Lab's memory-contamination stressor made literal in a benefits record, and it is why a system can look like it is being reviewed while quietly becoming harder to review.
That reframes the levers. It is not a better rule set: predefined rules reproduce any remaining defect identically on every matching claim, so a model upgrade is precisely the move that buys least, and a clone of the same rules checks nothing — the real second look has to be a genuinely different, legal-and-procedural human review. The load-bearing controls sit on the write pathway (a determination should not become a grant, a monthly payment, and a binding letter with no one in between), on the record-side reconciliation that stands in for the audit the interior lacked (reconcile an actioned payment against its source before it becomes a live obligation), and on an internal review rhythm that does not wait for a watchdog to arrive. The honest boundary is that none of this measures the survivors and veterans on the other end: the improper payments, the underpayments, and the legally deficient notification letters are documented outside any diagram like this one, and the tools here are rules-based automation and document extraction, not machine learning — the headline error rates are outcome statistics from statistical samples, never the tool's per-interaction rate.
The concepts used in this reading are defined in the Field Guide; the governance responses live in the Practice Library. The model organization for this case can be stress-tested in the PAN Lab.