This is Part 2 of a 3-part series on internal investigations in the age of AI. Join Michael for a webinar on September 8, 2026 on this topic — Register HERE.

AI tools promise to transform internal investigations — faster document review, instant summaries of witness interviews, pattern detection across millions of communications. The promise is real. So are the risks. Companies rushing to deploy AI in their investigation functions are discovering, sometimes painfully, that these tools fail in ways traditional investigation methods never did. Here are the five failure modes that should keep every general counsel and chief compliance officer up at night.

Failure Mode 1: The Hallucinated Summary

Generative AI tools confidently produce summaries that contain statements no witness ever made. An AI-generated summary of a witness interview that attributes a fabricated admission to an employee — or omits the exculpatory statement that actually mattered — can poison an entire investigation. If disciplinary action or a disclosure decision rests on a hallucinated summary, the company faces wrongful termination exposure, a corrupted investigation record, and a credibility problem with regulators. The rule must be absolute: no AI summary enters the investigation file without human verification against the source material.

Failure Mode 2: The Missed Hot Document

AI-assisted document review can dramatically reduce cost and time — and it can also miss the smoking gun. Models trained on general language patterns may not recognize coded language, industry euphemisms, or the significance of a document that matters only in context. When DOJ later finds the document your AI review missed, “the algorithm didn’t flag it” is not a defense. Validation protocols, statistically sound sampling, and human review of borderline categories are not optional extras — they are the difference between a defensible review and a negligent one.

Failure Mode 3: The Privilege Giveaway

This may be the most dangerous failure mode of all. When investigators feed privileged interview memoranda, legal analysis, or work product into a third-party AI tool, they may be disclosing privileged material to an outside party — with potentially devastating waiver consequences. Key questions: Does the vendor retain or train on your inputs? Who at the vendor can access the data? Does your engagement structure preserve the confidentiality that privilege requires? Companies need to answer these questions before the first document is uploaded, not after opposing counsel serves a motion to compel.

Failure Mode 4: The Discoverable Prompt Trail

Every prompt an investigator types, and every output the AI generates, is potentially discoverable. An investigator’s casually worded prompt — “find me evidence that Smith knew about the payments” — can be characterized by opposing counsel as evidence of a predetermined conclusion. AI interactions are part of the investigation record. Treat them that way: establish prompt discipline, retention protocols, and clear guidance on what investigators should and should not put into an AI system.

Failure Mode 5: The Overreliance Spiral

The most insidious failure is cultural. As AI tools get better, investigators get comfortable — then complacent. Judgment atrophies. The investigator who once read every key document now skims the AI summary. DOJ’s guidance specifically asks what baseline of human decision-making is used to assess AI. If the honest answer is “not much,” your investigation function has a problem no technology can fix.

Compliance Action Items

– Prohibit use of unapproved, consumer-grade AI tools for any investigation task; maintain an approved-tool list with documented security and confidentiality terms.

– Require human verification of all AI-generated summaries against source materials before they enter the investigation file.

– Build validation and sampling protocols into every AI-assisted document review, and document the methodology.

– Establish privilege protection procedures governing what materials may be processed by AI tools and under what contractual safeguards.

– Adopt prompt discipline and retention policies treating AI inputs and outputs as part of the investigation record.

AI tools used to conduct investigations are only half the story. In Part 3, we turn to the other half — what happens when AI itself becomes the subject of the investigation.