The Human Oversight Imperative: How AI Should (and Shouldn't) Support HR Investigations

The pressure on HR investigation teams in 2026 is coming from two directions at once: Employees are raising concerns through more reporting channels than ever, and a single complaint can trigger thousands of emails, messages, and documents that someone has to read, organize, and defensibly document. At the same time, federal enforcement has intensified.

According to the U.S. Equal Employment Opportunity Commission (EEOC), a record $528 million was recovered during fiscal 2025 as a result of the agency’s pre-litigation enforcement activities (i.e., mediation, conciliation, and pre-cause settlement agreements), apparently the highest amount ever recovered in the agency’s 60-year history. In most cases,  how an employer investigates and documents the complaint prior to an EEOC inquiry or lawsuit determines how well the matter will resolve. A robust, well-reasoned investigation can help resolve an issue early, while a poorly documented one can expose an organization to thousands or even millions of dollars in liability.

With HR departments being asked to lead more investigations involving increasingly complex evidence in this regulatory climate, CHROs at companies of all sizes are evaluating how artificial intelligence can help. And importantly, they care deeply about safeguards around the use of AI so that human judgment stays paramount.

Clearbrief founder and CEO Jacqueline Schafer — a member of the 2026 SHRM Labs WorkplaceTech Accelerator, Legalweek's 2025 Innovator of the Year, and a specialist in AI for investigative writing — believes the answer is a matter of structure: Let AI handle carefully capturing the factual findings, but keep areas of the report involving decision-making in human hands.

Where AI Earns Its Keep

Before drawing any lines, it is worth clarifying what AI is good at doing, because the argument for using it is quite strong. In investigations, AI is effective because it helps put structure around the chaos of massive amounts of emails, notes, and other evidence that needs to be described in a good report.

Clearbrief helps researchers at the point where synthesis is happening. AI ingests vast amounts of electronic data, email chains, and their attachments, and it outputs usable materials like hyperlinked timelines, witness interview guides that link to specific allegations and policies, and preliminary reporting where the conclusions are all linked to the underlying evidence. The investigator is still reading everything and writing the report, but the organization and citation of thousands of pages is done by AI.

"AI doesn't replace investigators, but acts as a force multiplier," Schafer said. "It helps with things like creating witness interview outlines for the HR leader based on the documentation, organizing and summarizing large sets of files, and generating the first draft of the report structure. But the way the AI output is designed makes all the difference. It needs to be built with guardrails that prevent bias and over-summarizing the facts.”

Schafer's experience leading Clearbrief to serve hundreds of the largest global law firms, as well as consulting with experienced HR leaders and investigators, helped her team home in on how to structure the AI “under the hood” so that it can avoid “hallucinations” or mistakes. But the clearest safeguard is to make it easy for the writer to verify and check every finding in a report thanks to hyperlinks to the exact pages. Schafer built Clearbrief to operate inside Word so that checking its output requires less time. The most defensible documentation connects each statement to the evidence that substantiates it, making it easier for supervisors to audit and far more defendable if subjected to scrutiny by a regulator, an opposing attorney, or an internal compliance reviewer.

Using It Safely: What to Build In Before You Rely on It

The value only holds if the tool is built responsibly, and in an investigation, "responsibly" has a specific meaning. Two safeguards matter most.

The first is grounding. Every output the tool produces — whether that’s a timeline, a summary, or a draft finding — should tie back to a specific piece of source evidence: a linked document, a page number, a time stamp. When the connection between a claim and its source is visible, the investigator can verify the tool's work rather than simply trust it. That design choice is what separates a tool that supports defensible work from one that introduces risk.

The second safeguard is data handling. Investigations involve some of the most sensitive information an organization holds, so the questions to ask a vendor are direct: Can the data you put into the tool be used to train a large language model? What are the security and privacy agreements around its use? Vague answers here are a reason to slow down.

Get those two things right, and AI becomes a genuine asset to an investigation. Get them wrong, and the same tool becomes a liability. The rest of the decision is about where, even with those safeguards in place, AI still shouldn't be making the call.

Where Humans Stay in Charge

Schafer is direct about where AI has no business operating: "We don't want AI tools taking the first pass at doing the HR work of human judgment and credibility determinations, such as decisions about the outcome of an investigation. It creates a risk that we will be influenced by the AI's suggestion.”

That principle carves out three areas of work in particular:

• Determining whether misconduct occurred or whether discipline is warranted. Both are determinations that rest on judgment, and both carry considerable legal, interpersonal, and ethical weight. It simply isn't possible to fully rely on AI to read people's intentions or discern their context, and those elements are often central to either determination. It is appropriate for AI to identify the handbook provisions and facts that relate to an issue being investigated, but not to decide the outcome.
• Assessing credibility. Whether someone is telling the truth can't be reduced to a formula. It depends on tone of voice, omissions, contradictions that come up over time, and contextual factors that don't map neatly onto patterns.
• Conducting the interview itself. AI can help draft questions, prepare reminders for the interviewer, or structure an interview plan, but it shouldn't be deployed to speak with witnesses or employees directly. Testimony is shaped by who's in the room. People need to know they are being heard in the moment when they provide testimony, and they need to feel like they are being heard by another person.

These are the human-centered boundaries that Clearbrief is built around. Clearbrief will draft the bones of the report to capture the key evidence in neutral wording, and the tool is designed to highlight in bright colors the places in a report where a human has to weigh in, instead of filling in the gaps with generated conclusions.

"We built Clearbrief so that our investigation reports don't reach a conclusion," Schafer said. "They flag for the writer where human judgment is required."

Keeping Hallucinations Out of the Report

This is exactly where the grounding safeguard earns its keep, because the risk it guards against is not hypothetical. The legal profession offers a preview of what happens when unverified AI output makes it into high-stakes written work.

As of late July 2026, researcher Damien Charlotin's database of AI hallucinations in court filings has logged more than 1,700 cases, including over 160 in employment law alone, where lawyers, often using legal-specific AI tools, filed briefs citing cases and factual evidence that didn't exist or misstating the meaning of cited sources.

Sanctions have climbed accordingly. In April, a federal magistrate judge in Oregon imposed a total of $110,000 in fines and fees on two attorneys and dismissed their case after their filings contained more than two dozen fabricated citations, one of the steepest penalties recorded to date.

The recurring lesson from these cases is that courts have placed the burden of verifying AI-generated output on the person who signs the document. Using an AI program, even a dedicated legal one, doesn't shift that burden.

The same logic applies to HR. Suppose an investigator submits a report concluding that a manager engaged in a pattern of inappropriate emails, citing seven specific examples. What happens if two of the cited emails were misidentified by the AI tool used to summarize and analyze them: wrong sender, wrong thread, or incorrectly presenting something the model invented as a direct quote?

If those findings turn out to be false on key points, the consequences don't fall on the tool. A firing built on that report can become a wrongful-termination claim, and the thread that unravels it becomes the opposing counsel's best exhibit. The investigator signed the report, and the organization owns the outcome.

This is why grounding matters so much. A tool whose every claim links back to its source for systematic verification lets the investigator catch that misattributed email before it ever reaches a finding. Clearbrief also has patented scoring for every linked reference that flags the writer to double-check accuracy based on how well the statement is supported by the linked text.

"We want to make it easier for the person writing the report that leveraged AI to also have full confidence in sharing it with colleagues," Schafer said.

What Human Oversight Looks Like in Practice

Effective oversight goes beyond a final check at the end; it needs to be woven through the process. Three guardrails make that a reality:

• Each element of each output should be cited back to a source, whether that’s a linked document, a page number, or a time stamp. If a draft doesn't let an investigator click a passage and land on the underlying record, the tool isn't built for responsible work.
• The human retains ownership over relevance, context, and meaning. AI can bring up and retrieve evidence, but it's the investigator's job to decide whether that evidence actually matters to the question at hand.
• Oversight is auditable. The AI's role at each stage should be transparent, with a clear record of where a human reviewed and approved the work.

A Buyer's Lens for HR Leaders

Before signing on with any AI investigation tool, HR leaders should be able to answer six questions about it:

1. Where does the tool's role end, and where does human judgment take over? A vendor that can't clearly describe the boundaries of their product is a cause for concern.
2. Can each conclusion in the output be traced back to the original source evidence? This is critical for investigative work.
3. What strategies exist to prevent AI hallucinations? Effective tools ground their outputs in source evidence, require human verification, and clearly flag uncertainty in their results.
4. Can data used with the tool be used to train a large language model, and what are the security and privacy agreements around its use? Inconsistent or unclear answers warrant caution.
5. Is a human required to review the tool's suggestions before anyone acts on them? A human-in-the-loop requirement is a marker of responsible AI; automated decision-making opens the door to liability.

Black-box outputs, vague data policies, and any vendor’s claim to replace human judgment are reasons to walk away.

What's at Stake

Investigations have always been about showing your work. Today's enforcement climate rewards HR teams that can document, case by case, exactly what decisions were made, by whom, and on what evidence. AI is the newest and most powerful tool for doing that — and the easiest to apply badly.

The teams that hold up under what's coming will be the ones who treat their AI choices the way they'd treat any key witness: interrogating its trustworthiness, documenting its outputs, and ensuring a human is the one to put their findings on the record.

Learn more about Clearbrief.