A VA watchdog found errors in nearly all 8,100 automated survivors' death-benefit decisions it reviewed, with $2.7M in improper payments

The problem

The Department of Veterans Affairs pays Dependency and Indemnity Compensation (DIC) to the survivors of veterans whose deaths are service-connected, and it processes some of those claims with an automated system rather than a human adjudicator. The VA Office of Inspector General reviewed whether that automated process correctly granted survivors’ entitlement over the 12 months from September 2023 through August 2024 (source). It is worth being precise about the technology: this is process automation, not artificial intelligence. As FedWeek put it, the system is “not specifically characterized as artificial intelligence,” and it “extracts data from scanned documents and applies predefined rules to generate decisions” (source).

What was built

The VBA’s Pension and Fiduciary Service runs an automated pipeline that reads scanned claim documents and applies predefined rules to produce a rating decision and the notification letter that goes to the survivor, for a benefit the VA has processed this way since 2020 (source). Some claims were, in the OIG’s words, “completely processed from beginning to end by automation” (source). The promise of that design is speed and consistency; the OIG set out to test whether the output was actually correct.

The outcome

It was not. Across a review of some 8,100 automated decisions and notices, the OIG found the deficiency rate was close to total. In the watchdog’s own summary, “At least 8,000 decisions or letters omitted favorable findings and had issues such as incomplete evidence summaries and incorrect formatting” (source). FedWeek, reviewing the same report, wrote that “more than 8,000 ‘contained at least one legal or procedural deficiency’” (source). A disability attorney who read the audit, Claire Hillan Sosa of the Deuterman Law Group, put the arithmetic bluntly to WFAE: the OIG “included a review of 8,100 cases of these automated DIC decisions and found that in 8,000 of them there were errors. So, this is essentially every case” (source).

The money followed. The OIG found that “At least 2 percent of decisions had legal errors, resulting in an estimated $2.7 million in improper payments,” a figure the report records as questioned costs of $2,727,764 (source). FedWeek attributed those overpayments to the system “not verifying the eligibility of a claimant and granting service-connected death benefits” without the required proof (source). WFAE recorded a concrete instance: “the system automatically paid out more than $22,000 to a family after attributing a veteran’s death to a service-connected condition without checking for any required” medical evidence (source).

Crucially, the OIG located the fault in the automation itself, not in a one-off glitch: “Deficiencies stemmed from flaws in automation rules and processes that, if not addressed, will recur because the system applies predefined rules without human involvement” (source). The audit also found the safety net was thinner where it was needed most, reporting that “the quality review process for automated claims was less rigorous than for traditional claims,” and that “a review of additional cases through November 2025 confirmed similar errors” (source).

The weakest load-bearing detail here is not a figure but a framing: the “98 percent” and the round “8,100 reviewed / 8,000 in error” come from the press and from an attorney reading the audit, whereas the OIG’s own careful wording is “at least 8,000 decisions or letters.” This story reports each figure as its source states it and does not launder the press characterization into the primary; the OIG document is the anchor, the press supplies the color (source).

Why this matters for buyers

This is a verified negative from an unusually strong posture: the auditor is the government’s own inspector general, examining the government’s own automation, with the numbers published on the record. The lesson is the founding thesis of this site, and it does not depend on the system being “AI.” A rule-based automation that “applies predefined rules without human involvement” is only as good as those rules, and here they were wrong often enough that a watchdog found a deficiency in essentially every case it sampled (source). The controls that would have caught it are unglamorous and specific: a quality-review process for automated output at least as rigorous as for human-handled claims, and an eligibility check the automation cannot skip. The audit found the first was weaker for automated claims, and the second was the exact step the system omitted when it paid a family $22,000 it should not have (source). Buyers of automation should assume the vendor’s error rate is whatever their own audit measures, not whatever the design promised.

How this was verified

Method: the VA OIG report summary for “Review of Automated Decisions for Veterans’ Service-Connected Death Claims” (report 25-00153-47, published April 30, 2026) was fetched live this session and archived to the Wayback Machine (local copy sources/vaoig-report-page.html; the full report PDF saved as sources/vaoig-25-00153-47_final.pdf). It stands as the Tier-1 primary for the deficiency finding, the 2 percent legal-error rate, the $2,727,764 questioned costs, and the automation-rules causation, all quoted verbatim. Those figures are independently corroborated by FedWeek (May 12, 2026) and by WFAE, the NPR member station in Charlotte (June 19, 2026), which also supplies the $22,000 individual instance and the attorney’s “essentially every case” characterization. Every source was copied into sources/. Date verified: 2026-08-24. This story’s status is checking; it is a cautionary process-automation audit record and will never carry a green client-outcome badge.

Related case files

• The Dutch Tax Authority’s algorithmic risk-profiling drew a €2.75M regulator fine after it wrongly flagged families - the same failure mode in another government’s benefits automation: an automated decision layer applied to a vulnerable population, later found unlawful and harmful by an independent authority.
• The UK DWP’s machine-learning fraud model, checked by the National Audit Office - the contrast case: a government benefits-automation deployment whose numbers were put through an independent audit, which is exactly the scrutiny that surfaced the VA’s failures here.
• US Treasury’s machine-learning system recovered $1 billion in check fraud - the positive counterpart in the same buyer (the US federal government): automation that delivered a measured, audited outcome rather than an unaudited one.
• Zillow’s algorithmic iBuying business was wound down after a ~$304M writedown - the private-sector version of the same lesson: an automated-decision system whose flaws only became legible once the losses were counted.

Sources

1. Tier 1 (primary). U.S. Department of Veterans Affairs, Office of Inspector General, “Review of Automated Decisions for Veterans’ Service-Connected Death Claims,” report 25-00153-47, published April 30, 2026. States the review period (September 2023 through August 2024), “At least 8,000 decisions or letters omitted favorable findings,” “At least 2 percent of decisions had legal errors, resulting in an estimated $2.7 million in improper payments,” questioned costs of $2,727,764, the automation-rules causation, and that “the quality review process for automated claims was less rigorous than for traditional claims.” https://www.vaoig.gov/reports/review/review-automated-decisions-veterans-service-connected-death-claims (archived http://web.archive.org/web/20260824083147/https://www.vaoig.gov/reports/review/review-automated-decisions-veterans-service-connected-death-claims ; local sources/vaoig-report-page.html, full report sources/vaoig-25-00153-47_final.pdf).
2. Tier 2 (strong secondary, independent trade press). FedWeek, “Audit: Automated Benefits Decisions at VA Rife With Errors, Overpayments,” May 12, 2026. Corroborates the ~8,100 reviewed and more than 8,000 with “at least one legal or procedural deficiency,” the 2 percent / $2.7 million overpayments, the eligibility-verification failure, and that the system is “not specifically characterized as artificial intelligence” but “extracts data from scanned documents and applies predefined rules to generate decisions.” https://www.fedweek.com/federal-managers-daily-report/audit-automated-benefits-decisions-at-va-rife-with-errors-overpayments/ (Wayback save returned an upstream error this session; local copy sources/fedweek.html).
3. Tier 2 (strong secondary, independent public radio). WFAE (NPR, Charlotte), “Audit finds VA automation glitch ruined 98% of veteran survivors’ benefits claims,” June 19, 2026. Carries the “8,100 cases … 8,000 of them there were errors. So, this is essentially every case” characterization from disability attorney Claire Hillan Sosa, and the more-than-$22,000 individual overpayment instance. https://www.wfae.org/2026-06-19/audit-finds-va-automation-glitch-ruined-98-of-veteran-survivors-benefits-claims (archived http://web.archive.org/web/20260824083305/https://www.wfae.org/2026-06-19/audit-finds-va-automation-glitch-ruined-98-of-veteran-survivors-benefits-claims ; local copy sources/wfae-npr.html).