AI Tools in Workplace Investigations: What They Can Do, Where They Break Down, and What Investigators Are Responsible For

Investigators at mid-to-large employers are increasingly using AI tools to handle parts of the investigation workflow: drafting interview outlines, generating first-pass transcriptions, summarizing witness statements, cross-referencing policies. Some vendors are now marketing tools that claim to assess credibility through linguistic pattern analysis. The practical and legal questions these applications raise are not hypothetical — they're showing up in discovery files and EEOC charges now.

Understanding where AI tools are useful, where they introduce risk, and what documentation obligations follow is becoming a core competency for anyone running workplace investigations.

Where AI Tools Break Down

Large language models generate text probabilistically. A structural consequence of that design is hallucination — outputs that are fluent and plausible but factually incorrect or unsupported by the underlying source material. When an investigator uses an LLM to identify inconsistencies across a set of witness statements, the model may surface patterns that don't exist in the source text, or miss ones that do. In the logistics layer of an investigation, errors like that are easy to catch. When the same output appears in a credibility section of an investigation report, it enters the litigation file uncorrected.

The sharpest version of this problem involves credibility assessment. Several vendors are marketing tools that claim to detect deception through analysis of written or spoken language patterns. The scientific basis for these claims is thin. The National Academies of Sciences' 2003 review of deception detection research found that even polygraph evidence — far more structured than language pattern analysis — lacks sufficient reliability for high-stakes decisions. Applying an AI-generated credibility score to a harassment complainant's account, then using that score as a basis for closing an investigation, creates a significant evidentiary problem when the file is reviewed.

Disparate impact is a second risk that receives less attention. AI models trained on historical workplace communication data may encode patterns that disadvantage employees whose writing style, English fluency, or cultural expression diverge from the training distribution. Under Griggs v. Duke Power Co., 401 U.S. 424 (1971), employment practices that produce disparate impact — regardless of discriminatory intent — require a business necessity justification. An AI tool that systematically discounts written statements from non-native English speakers could trigger exactly this analysis. The EEOC's 2023 technical assistance document on AI and employment discrimination makes clear that employers remain responsible for discriminatory outcomes produced by algorithmic tools, including those supplied by third-party vendors.

Where AI Tools Genuinely Help

For tasks that don't require human judgment, AI tools provide real efficiency. An investigator handling a retaliation claim can use an LLM to generate a comprehensive interview question framework — covering temporal proximity, comparator treatment, and pretext indicators — then edit it against the specific facts of the matter. The AI produces scaffolding; the investigator makes every substantive decision about what to ask and how.

First-pass transcription and summarization of recorded interviews is another legitimate application, provided investigators treat AI output as a draft requiring verification. SHRM's guidance on workplace investigation documentation reflects a longstanding practice standard: investigation records must accurately reflect what witnesses actually said. Any AI-generated summary needs to be reviewed against the source recording before it enters the file.

Policy cross-referencing is a third useful application. Feeding a company's code of conduct, anti-harassment policy, and applicable HR procedures into a retrieval system and asking it to identify which provisions are implicated by a given fact pattern is straightforward document retrieval — not a judgment call — and it reduces the risk of missing a relevant policy provision.

Documentation and Disclosure

Investigation files don't stay inside HR. Plaintiff's counsel, the EEOC, labor arbitrators, and courts review them. If an AI tool contributed to a finding — summarizing witness accounts, flagging an inconsistency, structuring a credibility analysis — and that contribution is not reflected anywhere in the file, investigators face a documentation problem when the file is scrutinized.

The NLRB General Counsel's 2023 memorandum on AI and the NLRA signals that regulators are actively examining how AI outputs affect employee rights determinations. Reasonable practice now is to treat AI-generated content the way any secondary source should be treated: document where it was used, what it produced, and what human review was applied before the output entered the file.

A Practical Distinction

A workable gatekeeping question for any AI application in an investigation is whether the task involves judgment or logistics. Judgment calls — credibility determinations, cause findings, policy violation conclusions, interview sequencing based on witness dynamics — belong to the human investigator. Logistics tasks — question drafting, scheduling, document organization, first-pass transcription — are where AI tools provide value without introducing the reliability and bias problems described above.

A second useful test is discoverability: whether the investigator could explain the AI output and how it was used to opposing counsel. If the answer is no, the tool shouldn't be in that step of the workflow.

The regulatory landscape is developing quickly. The EEOC's technical assistance, Illinois and New York state-level AI bias statutes, and the EU AI Act's provisions on high-risk employment applications all reflect a consistent direction: employers who cannot document how AI was used, what it produced, and how human reviewers evaluated that output before it influenced a decision will face difficulty when those decisions are challenged.

Marshal's investigation workflow logs every action, document version, and finding in a tamper-evident audit chain — exactly the kind of human-controlled record that makes AI-assisted investigations defensible when the file gets scrutinized.