evidoria

← Back to browse

Good practice Imported

CBP's RelativityOne — DHS Turns to AI to Tame a Record 1-Million-Request FOIA Backlog

United States of America · Washington, D.C. · See the United States of America profile · See the Washington, D.C. profile

Evidence: Observational / pre–post Top 83% 40/100 · Ask Evidence Copilot about this practice

Facing a record 1M+ annual FOIA requests, DHS's Customs and Border Protection uses the RelativityOne AI platform to learn from reviewers and flag relevant documents, while civil-liberties groups warn AI still isn't ready to judge legal redaction nuance.

1,000,000+
FOIA requests received (FY2025) (FY2025)
~60%
Share of all federal FOIA requests (FY2025)
221,068
Pending requests at start of FY2025 (FY2025 start)
245,572
Pending requests at end of FY2025 (FY2025 end)
16%
Backlog as share of requests received (FY2025)
CBP's RelativityOne — DHS Turns to AI to Tame a Record 1-Million-Request FOIA Backlog

Details

Maturity
Established
Promoter
U.S. Customs and Border Protection (Department of Homeland Security)
Period
2024–2026
Keywords
public records, FOIA compliance, border & immigration enforcement, government transparency

Context

In fiscal year 2025, the U.S. Department of Homeland Security received more than 1 million Freedom of Information Act (FOIA) requests - about 60% of all FOIA requests filed against the entire federal government - and its Customs and Border Protection component processes the bulk of that volume using RelativityOne, an AI-enabled document-review platform approved for FOIA use by DHS's Privacy Office in 2024.

Objectives

RelativityOne is built to learn from human reviewers' redaction and relevance decisions and then independently tag subsequent documents as relevant or irrelevant, surfacing records with a confidence score and supporting citations for a human FOIA analyst to accept, modify or reject.

Activities

DHS started fiscal year 2025 with 221,068 pending FOIA requests and, despite receiving and processing roughly 1 million more, ended the year with 245,572 pending - a backlog equal to 16% of requests received - and the department has said it plans to add further automation to improve efficiency and reduce administrative redundancies.

Results

DHS describes the year-end backlog as having been maintained rather than reduced even as caseload surged, so no attributable reduction in processing time or backlog has yet been demonstrated.

Conclusions

Civil-liberties observers remain cautious: the ACLU has said AI might help expedite finding documents but that redactions involve legal nuance AI isn't ready for, and the Cato Institute has noted that DHS's disclosures omit detail on which systems make independent coding decisions versus merely assisting a human reviewer.

Implementation
Indicative cost
High (€500k–€5M) — Commercial platform (RelativityOne) licensed for CBP's FOIA division; DHS has not disclosed contract costs.
Time to results
Medium (1–3 years) — Approved by DHS's Privacy Office in 2024 and in active use processing FY2025's caseload of over 1 million requests; DHS has signalled further automation is planned but has not published a timeline.
Staffing & skills
CBP FOIA analysts, DHS Privacy Office reviewers

Conditions for success

  • Formal privacy review (Privacy Threshold Analysis) before approval for FOIA use
  • Human analysts retain final decision authority over AI-flagged relevance and redaction tags
  • The system is trained continuously on human reviewers' actual redaction and relevance decisions

Common failure modes

  • Backlog can persist even with AI assistance if incoming request volume outpaces processing capacity
  • Lack of public disclosure on which coding decisions are made independently by AI versus a human reviewer undermines transparency

Commonly funded by

National / regional programmes

Indicative funding routes for practices of this type — always check each programme's current calls and eligibility rules.

Do you run this practice? Claim it — verified implementers get a public contact pathway and can propose corrections.

Data sources

Where this practice's information was retrieved from, and when.

Attachments

Similar practices you may find useful