Overview

You're investigating a workplace conduct allegation. You have interview notes from the complaining party, the respondent, and three witnesses. You have emails, Slack conversations, and meeting notes. You have 47 pages of raw material and no clear narrative.

You think: "I'll ask AI to draft the investigation findings."

Stop. Do not do this.

An AI-generated investigation summary can, and likely will:
- Misidentify culpability, placing blame on the innocent party
- Inadvertently admit company liability through careless language
- Violate confidentiality by naming sources or revealing sensitive medical/personal information
- Create exposure for discrimination claims by making assumptions based on protected characteristics
- Generate findings that contradict each other or your actual evidence
- Use language that, in discovery, looks like bias or predetermined conclusions

Investigation documentation is legal evidence. If a case goes to court, opposing counsel will examine every sentence. "This looks like AI boilerplate" is not a defense. "This isn't what our evidence actually supports" is worse.

But AI has a real role in investigation work. Not drafting findings. Not making judgments. Organizing. Analyzing. Helping you think clearly about what you actually have.

This lesson teaches you the exact boundaries of AI usefulness in investigations. You'll learn what AI can do without legal risk. You'll learn what AI absolutely cannot do. You'll learn a workflow that uses AI where it helps and keeps human judgment and legal review where it belongs.

Why This Matters for HR Professionals

Investigations are the highest-stakes HR work you do. They touch employee livelihoods, company liability, legal exposure, reputational risk. They're also emotionally heavy: someone's alleging harm, someone's being accused, witnesses are uncomfortable.

The stakes mean everything about an investigation becomes evidence. Your notes. Your process. Your reasoning. How you interviewed people. Who you interviewed. How you weighted evidence. All of it can be discovered and used against you.

Poor investigations create cascading liability:
- Company liability: Didn't investigate properly, didn't fix the problem, allowed misconduct to continue (negligent retention, hostile environment).
- Individual liability: The accused person can sue the company for defamation or wrongful termination if the investigation was unfair or inaccurate.
- Reputational liability: Other employees lose faith in HR when they see an investigation handled poorly.

Most investigations don't need external lawyers. Many situations are straightforward: someone violated a clear policy, you have evidence, you can make a fair finding. But investigation documentation must be legally defensible.

This is why AI as "drafter" is dangerous. But AI as "organizer" is valuable. Organizing 50 pages of notes into a clear timeline saves time and helps you see what you actually have versus what you're assuming.

Important: If your investigation involves allegations of discrimination, harassment, sexual misconduct, or retaliation, involve employment counsel before finalizing anything. These are legally protected categories. AI should never touch the substance. AI can organize materials, but a lawyer should draft findings.

What AI Can Do (Safely)

AI can help with organization:
- Timeline building from interview notes
- Fact extraction (what was said, what was documented, what's disputed)
- Contradiction identification (where accounts differ)
- Evidence categorization (what's email vs. chat vs. interview)
- Witness list organization

AI can help with thinking:
- Sanity-checking your logic ("Based on this evidence, does this finding hold up?")
- Spotting gaps ("What other evidence should I gather to settle this contradiction?")
- Identifying assumptions ("Am I assuming intent without evidence?")

AI cannot safely do:
- Draft investigation findings (that needs human judgment)
- Assess witness credibility (that's your call based on demeanor, consistency, bias)
- Make causality claims (was X caused by Y, or coincidence?)
- Recommend discipline (that's policy interpretation + judgment)
- Determine legal implications (that's legal's job)

The Investigation Process: Where AI Fits

Phase 1: Scope and Planning (No AI)

You receive a complaint. First step: define exactly what you're investigating.

Scope is a tightly defined question. Not "Is this manager a bad manager?" but "Did the manager refuse to provide training to this employee, and if so, was it discriminatory?"

Define scope before you interview anyone. This prevents scope creep and keeps the investigation manageable.

AI doesn't help here. This is judgment.

Phase 2: Evidence Gathering (No AI, but organize as you go)

Conduct interviews. Document everything:
- Who you interviewed
- When
- What they said (or transcription)
- What documents they reference

Gather documents:
- Emails
- Chat records
- Performance evaluations
- Training records
- Anything relevant to the allegation

Save everything in one place. Organized. Dated.

AI doesn't help during gathering, but get organized as you go. You'll need this.

Phase 3: Organization (AI helps here)

Once you've gathered materials, you have a pile of raw notes and documents.

Ask AI: "Organize these investigation materials. I'll paste interview notes and document summaries. Create a timeline of events, showing what happened and who said what. Also show where accounts match and where they contradict."

Example input:

[Interview with Maria, 3/15: "On March 1st, I asked my manager David for permission to attend the training course on quarterly planning. He said 'no, we don't have budget.' I was upset because other people in the team took that course."]

[Email from David to Maria, 3/1: "I've approved your request for the quarterly planning training. Enrollment is open through March 15."]

[Interview with witness Marcus, 3/16: "I heard David tell Maria he wouldn't let her take the training. I was surprised because he'd approved similar training for other team members before."]

AI produces:

Timeline and Account Summary:

March 1 - Training Request and Decision
- Maria's account: Requested training. David denied for budget reasons.
- David's email: Approved Maria's training request.
- Contradiction: Maria recalls denial. Email shows approval.
- Witness account: Marcus heard David deny the request.
- Analysis: Email is contemporary record. Interviews are retrospective. Email is stronger evidence than recall.

Training Access Pattern
- Marcus reports David previously approved training for other team members.
- No direct evidence of other approvals/denials gathered yet.
- Possible next step: Review training records to verify.

This organization helps. It's neutral. It shows where evidence supports what. It identifies contradictions (Maria's memory of "no" vs. the actual approval email).

Phase 4: Analysis and Fact-Finding (Mostly human, AI helps thinking)

Once organized, you analyze. What can you establish as fact based on evidence?

Format your findings:

"We interviewed [names]. We reviewed [documents]. Based on this evidence:

• 
  [Established fact, with evidence] - "Maria requested training on March 1 (email evidence). David approved the request on March 1 (email evidence)."
• 
  [Disputed fact with credibility assessment] - "Maria recalls being told 'no' (interview). David's approval email contradicts this. The email is contemporaneous and more reliable than later recollection. We find it credible that approval was given."
• 
  [Remaining question] - "The basis for David's decision remains unclear. His email doesn't explain his reasoning."

This is fact-finding, not judgment. You're stating what the evidence supports.

AI can help by asking: "Does this finding follow from your evidence? Or are you inferring?"

Example: You write "David deliberately withheld training from Maria."

AI red flags: "Your evidence shows David denied approval then sent an approval email. That's contradictory and unclear on intent. Can you support 'deliberately' based on evidence, or is that interpretation?"

This is AI being useful. It's not drafting. It's helping you check your own logic.

Phase 5: Recommendations and Action (No AI)

Once you have findings, you recommend action. This is judgment + policy + risk assessment. AI shouldn't touch it.

Human decides: What does this finding mean for the employee? Discipline? Coaching? Remediation? This is the judgment call.

Phase 6: Communication (Careful AI use, mostly human)

You tell the people involved what you found. This communication must be:
- Factual (state the finding)
- Confidential (don't identify witnesses)
- Specific about action (here's what we're doing)
- Fair (explain the basis without revealing all details)

Example of good communication:

"We investigated your request. Based on our review of communications and interviews, we found that your training request was approved on March 1. You were entitled to attend the training. We've also reviewed [relevant policy], and [outcome of investigation]. Going forward, [action]. If you have questions, please contact HR."

This is factual, doesn't name witnesses, and explains action. It doesn't reveal every detail of the investigation.

AI can help draft this for clarity and tone, but a human must review it. The risk of miscommunication is high.

The Key Mistake to Avoid: Scope Creep

Investigations bloat. You start investigating "Manager didn't provide training" and you end up investigating "Is manager generally fair?" and "Has this happened to anyone else?" and "Are there systemic issues?"

That's not investigation. That's an audit.

Investigations need narrow scope:

Bad scope: "Look into all interactions between Maria and David."
Good scope: "Did David deny Maria access to training on March 1, and was the denial discriminatory?"

Bad scope: "Investigate whether the manager is a bad manager."
Good scope: "Did the manager exclude this employee from a key project because of the employee's protected status?"

Narrow scope protects everyone. It keeps the investigation bounded. It prevents "investigating everything" which takes forever and often uncovers unrelated issues that aren't your job to fix.

AI can help you stay focused. When you're tempted to expand scope, ask AI: "Is this new question part of my original scope, or scope creep?"

How to Use AI Without Creating Legal Risk

Safe pattern:

Input: Raw materials (interview notes, emails, documents, chat logs)
Process: AI organizes and identifies contradictions
Output: Organized summary, NOT findings or conclusions

Then you add human judgment: "Based on this organized evidence, here's what I find credible and why. Here's what I recommend."

Unsafe pattern:

Input: Raw materials
Process: "AI, draft the investigation findings"
Output: AI-generated findings document

You copy-paste it and send it to legal for review.

This creates liability because:
1. The findings came from AI pattern-matching, not your evidence judgment
2. If something's wrong, it's on you, not AI
3. The document looks generic and boilerplate (bad in court)
4. You haven't actually thought through the evidence

Always do the thinking yourself. Use AI to organize. Use your judgment to conclude.

Tip: Keep a separate "Investigation Workspace" file. Paste raw notes. Ask AI to organize. Copy the organized version into a working document. You then add your analysis and human judgment on top. This keeps AI's output separate from your conclusions.

The Confidentiality Principle: Protecting Identities

Investigation documents are confidential. You don't share detailed findings with everyone. You don't name witnesses publicly.

When you communicate to the parties involved (complaining party, respondent, witnesses), you share facts but not source identification.

Good: "We found that multiple people described communication issues."
Bad: "John and Priya said that the manager was dismissive in meetings."

Bad: "The manager's assistant told us the manager yelled in meetings."
Good: "We heard consistent reports of raised voice in meetings."

AI can help you write aggregated findings that are factual without identifying sources.

Prompt: "I have findings about [situation]. I need to communicate them to [party] without naming witnesses. Here's what I found: [details]. Rewrite this as aggregate findings: 'multiple people reported X' rather than 'John reported X.'"

AI helps you keep findings factual while protecting confidentiality.

When to Involve Legal: The Bright Lines

Must involve legal:
- Discrimination allegations (based on protected characteristics: race, gender, age, religion, disability, etc.)
- Harassment allegations (especially sexual harassment)
- Retaliation allegations (I complained, then bad things happened)
- Sexual misconduct allegations
- Physical violence allegations
- Violation of privacy rights
- Anything where employee might sue

These are legally protected areas. Mistakes create liability. Get a lawyer.

Probably involve legal:
- Anything leading to termination
- Anything involving policy violation with significant impact
- Anything where the outcome is unclear or the allegations are complex
- Situations where the parties might escalate to legal action

Can typically handle without external legal:
- Simple policy violation with clear evidence (person was late 12 times despite warnings; fired)
- Simple miscommunication with no policy violation and no harm
- Simple performance issue (not investigation-level)

When in doubt, involve legal. The cost of a lawyer is cheaper than the cost of a failed investigation.

Practical Workflow: Template You Can Use

Step 1: Define Scope (write it down)
"I'm investigating whether [manager] failed to [action], on [date/timeframe], with alleged impact [impact]. I'm not investigating [what's out of scope]."

Step 2: Gather Evidence
- Interview complaining party
- Interview respondent
- Interview witnesses
- Gather relevant documents
- Keep all in one file, dated, organized

Step 3: Ask AI to Organize
Paste notes + documents. Ask: "Create a timeline and fact summary showing contradictions."

Step 4: Review and Analyze
Read the organized output. Ask yourself:
- What can I establish as fact?
- What's contradicted by other accounts?
- What's my credibility assessment (who do I find credible, why)?
- What evidence do I need to gather?

Step 5: Draft Findings
Write your findings in plain language. Format:
- What happened (facts you can establish)
- What's disputed (contradictions and your credibility assessment)
- What you conclude (does the allegation hold up?)

Step 6: Legal Review (if needed)
Have counsel review before finalizing.

Step 7: Communicate
Tell parties what you found, in aggregate, and what action you're taking.

Try This Now: Two Hands-On Exercises

Exercise 1: Organize Mock Investigation Notes

Create mock interview notes for a hypothetical conflict scenario:
- Employee A claims they were excluded from a project for unfair reasons
- Manager claims it was a capability gap
- Witness claims they heard the manager criticize employee A's background

Prompt AI: "Organize these interview notes into a timeline. What's factual? What's disputed? Where are contradictions?"

Observe how AI organizes messy notes into clarity.

Exercise 2: Check Your Logic

Write a draft finding from your scenario (e.g., "The manager acted discriminatorily.").

Prompt AI: "I made this finding: [your finding]. Based on this evidence: [the evidence]. Does the finding follow from the evidence, or am I inferring/assuming?"

See if AI catches places where you're going beyond what evidence supports.

Practical Application - "What to Do Monday Morning"

• 
  Create an investigation protocol: Define scope upfront. Document your process. Train yourself on what to gather.
• 
  Build an investigation template: Timeline? Interview questions? Document checklist? Have a template so you're consistent.
• 
  Create an investigation workspace: Folder where you organize all investigation materials, interviews, documents, evidence summary.
• 
  Use AI for organization only: Timelines, fact summaries, contradiction identification. Not for drafting findings or recommendations.
• 
  Draft findings yourself: Based on organized evidence, you decide what you found and why.
• 
  Involve legal when needed: Discrimination, harassment, retaliation, anything high-stakes.
• 
  Communicate carefully: Factual findings, aggregate (not sourced), explanation of action.

Key Takeaways

• AI organizes and analyzes; humans judge and decide: This is the critical boundary.
  - Investigation documents are evidence: Every word matters. Clarity and accuracy are essential.
  - Never ask AI to draft investigation findings: This creates liability and removes your judgment from the process.
  - Use AI to organize raw materials: Timeline, fact extraction, contradiction identification are safe and valuable.
  - Scope tightly: What exactly are you investigating? Don't let it expand.
  - Protect confidentiality: Findings state facts. They don't identify witnesses.
  - Involve legal for protected-area allegations: Discrimination, harassment, retaliation.
  - Document everything: Your process, evidence, reasoning. It all matters in discovery.

FAQ

Q: How detailed should investigation findings be?
A: Detailed enough to support your conclusion. Show your evidence. Explain credibility assessments. But don't include unnecessary details that could identify witnesses.

Q: Should I investigate everything someone alleges, or just the core claim?
A: Just the scope you defined upfront. If allegations expand beyond scope, decide consciously whether to expand scope or note that in the findings.

Q: Can I investigate anonymously if the complaining party wants anonymity?
A: Generally no. The respondent has a right to know their accuser and respond. You can take steps to protect confidentiality, but you can't hide the complainant's identity from the respondent.

Q: How do I assess credibility when accounts conflict?
A: Consider: demeanor in interview, consistency of account, presence of bias, corroborating evidence (like emails), knowledge of details. Don't assume contemporaneous evidence (emails) is always more credible than later recollection, but it usually is.

Q: What if I gather evidence and it contradicts the initial allegation?
A: That's your finding. If evidence shows the allegation wasn't supported, that's the conclusion. Fair investigation means following evidence, not confirming allegations.

What's Next

Lesson 7.2 is about conflict resolution, how to mediate disputes between employees and help them work toward resolution when possible, and recognizing when conflicts can't be resolved.