DHS AI Use Case Inventory, 2022–2026

Sourcing

All 954 rows come directly from six DHS-published spreadsheets — not from a secondary summary or a prior derivative document. Each vintage was parsed independently and cross-checked against its own row and column structure before anything was normalized:

• 2022 baseline:24_1213_ocio_2022-dhs-ai-use-case-inventory.xlsx
• Aug 2024:2024_08_12_ocio_dhs-inventory-of-ai-use-cases.xlsx
• Dec 2024:24_1216_ocio_2024-dhs-ai-use-case-inventory_1.xlsx
• Jul 2025:25_07_01_ocio_2024-dhs-ai-use-case-inventory July_Revision.xlsx
• Jan 2026:26_0128_mgmt_2025-DHS-AI-Use-Case-Inventory.xlsx
• Jun 2026:26_0612_mgmt_cx_2025-june-update-dhs-ai-use-case-inventory.xlsx

Row counts

Each vintage's row count was independently verified against the raw file — 21 (2022), 67 (Aug 2024), 197 (Dec 2024), 198 (Jul 2025), 236 (Jan 2026), and 235 (Jun 2026), for 954 total. Two things worth knowing about that count: the Jul 2025 file's raw sheet includes two trailing footnote rows after the last real entry (a naive non-blank-cell count would land on 200, not 198 — we excluded both). Three vintages also include a handful of pilot-stage IDs formatted "DHS-P#" rather than the usual "DHS-###" numbering, which are included as real entries.

Four template generations, one schema

DHS has changed its inventory template three times since 2022, and the fields simply aren't the same across vintages:

• 2022 and Aug 2024 predate any high-impact framework entirely — those rows are tagged "not assessed" on the designation chart and table, not "not high-impact." That distinction matters: it means DHS hadn't yet made a determination, not that it made a favorable one.
• Dec 2024 and Jul 2025 (OMB M-24-10) used separate Rights-Impacting and Safety-Impacting yes/no fields, each with its own "presumed category" field populated when a presumption was made but the use case doesn't (yet) carry a determination either way. We map a "yes" on either field to high-impact, a populated presumed category with a "no" answer to downgraded, and a plain "no" to not high-impact. This mapping is our own analytical overlay, not a category DHS itself published under those two templates — treat the Dec 2024 / Jul 2025 designation numbers as a reasonable bridge to the current framework, not a DHS-stated fact.
• Jan 2026 and Jun 2026 (OMB M-25-21) introduced the single "Is the AI use case high-impact?" field with lettered options — we read a) as high-impact, b) as presumed-high-impact-but-downgraded, and c)/d) as not high-impact. DHS's own raw data has minor entry noise here (e.g., "presumed high-impact but determined not high-impact" with and without a comma) that we normalized on the leading letter, not the trailing text.

Reading the growth chart

The rise from 67 to 197 use cases between Aug 2024 and Dec 2024 is substantially a reporting-scope change — the M-24-10 template required far more fields, including systems DHS had not previously disclosed — rather than four months of new AI deployment. Treat the chart as a record of what DHS has disclosed at each point, not a literal deployment timeline.

Freshness note

This dataset reflects DHS's June 12, 2026 posting (26_0612). Checked directly against DHS's AI Use Case Inventory Library on September 15, 2026: the newest file listed is still that June 2026 revision — nothing dated July 2026 or later has been posted. The library page's own "last updated" stamp can move (e.g., to reflect an administrative review) without a new inventory file accompanying it, which is what produced an earlier, unconfirmed signal of a July 2026 update. This page will be re-checked before any figures are carried forward without review.

What we didn't do

We didn't deduplicate use cases across vintages into one row per system — the table shows every vintage's listing of a use case as its own row, so you can see exactly what DHS said at each point in time, including where a use case was renamed, retired, or reappeared under a new description. We also didn't attempt to resolve every ID-continuity question by hand — DHS itself doesn't always carry a consistent name for the same system across vintages (a handful of 2022 entries, for instance, are matched to their later DHS-ID only by summary text, not by a shared identifier in DHS's own data).