Loading page…
Peter Hatch outlined requirements for AI-assisted casework, while some agency uses avoid a high-impact designation because people review their outputs.
Listen to this story
At the November 12, 2026, GDIT Emerge event, HSI intelligence official Peter Hatch described a procurement standard shaped by criminal-case discovery: investigators must be able to audit AI outputs, trace them to original documents, and document AI use in reports. HSI already has nearly a dozen AI applications deployed, so the requirement applies to current workflows as well as future purchases. For vendors, the practical test is whether tools support source-level validation and human accountability—not just whether they produce useful answers.
Deployed uses include summarizing and categorizing tips, classifying investigative media, and redacting audio and video.
About five additional applications are in pre-deployment, including tools for lead generation, translated-text analysis, digital-record review, and maritime data.
Hatch highlighted identity resolution across travel forms, social media, and public data as one of AI’s most impactful uses.
Homeland Security Investigations needs AI that investigators can check against original documents—not answers they must take on trust. Peter Hatch, a senior HSI intelligence official, outlined that requirement at a November 12, 2026 event in Washington. AI inputs and outputs can become part of the case material subject to disclosure in court, according to FedScoop’s coverage.
Hatch is principal assistant director for HSI’s Office of Intelligence. Speaking at GDIT Emerge, he said the agency has more AI guardrails than other Department of Homeland Security units. HSI is Immigration and Customs Enforcement’s investigative arm, handling federal criminal cases involving smuggling, cybercrime, financial crime, human trafficking and other offenses.
Hatch tied those requirements to the legal scrutiny that comes with building criminal cases. HSI’s work is subject to discovery—the process of disclosing case material—including emails, notes and what investigators put into and receive from AI systems. Material used to build a case must be able to withstand examination in court, he said.
That shapes what HSI asks suppliers before acquiring a tool: Can investigators audit its outputs, and can they get to the underlying source documents? Hatch said the agency needs workflows that let people validate results throughout the process, with humans retaining responsibility for the work.
Investigators also have to document when they use AI to build a case and disclose that use in reports. The requirement covers the use of the technology itself, alongside checking whether its output is correct.
Every step of the way, we have to know and validate that it was correct.
Peter Hatch, principal assistant director for HSI’s Office of Intelligence
These requirements apply to an agency already using the technology. HSI has nearly a dozen deployed AI use cases, according to FedScoop’s analysis of DHS’s latest AI inventory. The listed applications span information arriving at the agency and media used in investigations or legal proceedings:
Hatch singled out identity resolution—working out which records refer to the person investigators are seeking—as one of AI’s most impactful uses. HSI has travel-form information, social media and large amounts of publicly available data, he said. He described the challenge as finding “100 needles in a haystack” and determining which one is the actual person being sought.
Around five additional use cases are in pre-deployment stages, meaning they are still being developed or acquired. Those planned applications include generating leads, analyzing translated text, reviewing digital records and combining fragmented maritime data. They are distinct from the tools already listed as deployed.
Some applications HSI is pursuing carry the designation “presumed high-impact but determined not high-impact.” The agency’s justification is that their outputs support decisions rather than serve as the principal basis for them, and people review those outputs before any action is taken.
That distinction has drawn scrutiny. Experts cited by FedScoop have questioned the high bar for classifying uses as high-impact and raised concerns that these applications do not have to meet the same standards. Human review is therefore part of both HSI’s safeguard and its justification for the classification.
For suppliers, Hatch’s request was not simply for more models or tools. He said HSI wants to use what is already available better, embedding it in auditable workflows that cite source material and keep people in charge. AI can make a good agent faster, he said; it cannot turn a bad criminal analyst into a great one.
Loading discussion...
Join the conversation
Explain whether a person checking the output changes the underlying risk.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.