TL;DR: The best AI prompts for HR investigations don't run the case; they turn messy notes into neutral documentation and catch the loaded language that’s easy to miss.
I've always treated HR like an engineering problem. Build the system, document the workflow, then go hunting for the edge cases. Employee relations is where the edge cases live.
In AI Prompts for HR: Onboarding, I shared a prompting framework and three prompts for employee onboarding. This is the second installment in the series, and it's the topic attendees asked for most: employee relations and investigations. I walked through it live on a recent webinar; watch the recording here.
Where AI belongs in an investigation
AI should not be the judge. Gathering information, knowing the right process, getting legal guidance, and deciding how a case plays out should all stay with you.
What AI is good at is two things:
- Turning messy notes into neutral documentation.
- Catching your own loaded language.
Every prompt below follows the same shape as the framework from my onboarding post: context, the ask, what good looks like, and guardrails. In employee relations, the guardrails do most of the work.
Keep the data safe first
Before you paste anything, do these three things.
Use placeholders. Every prompt here uses [Employee A] and [Employee B] instead of names. Swap in generic labels before the notes ever touch an AI tool. Tip: Anonymize in a doc first. When my notes include names, I drop them into a Google Doc, hit Ctrl+F for each name, and replace it with a generic placeholder. Takes a minute.
Check your company’s security posture. Ask your IT or security admin what's allowed. The answer depends on your company's broader policy, so don't guess.
Use an enterprise tier. We use the enterprise version of Claude at Warp because of the level of compliance and data processing addendums the enterprise plan supports, which is required by our IT & security team.
Prompts for Employee Relations & Investigations
Prompt 1: Turn raw interview notes into a neutral summary
After I've interviewed everyone involved, my notes are a mess. Shorthand. Secondhand claims. Sometimes my own reactions mixed in. They're what a well-meaning HR person writes at 4 PM after a very hard interview, and that's exactly when loaded language sneaks in.
Here's the prompt:
Turn these raw notes into a factual, neutral summary for an investigation file. Report only what was said and observed, with no conclusions about intent or credibility. Flag any word choices that sound judgmental so I can review them.
Three boundaries do the work:
- Report only what was said and observed. This sets the basic line with the AI.
- No conclusions about intent or credibility. This stops the model from playing judge.
- Flag judgmental word choices. This keeps me as the reviewer. It also surfaces secondhand statements, which show up constantly in interview notes (ie: "this other person said that...").
Prompt 2: Draft follow-up investigation questions
Once I have a clean summary, my next job is figuring out what I still don't know. This prompt drafts follow-up questions for my next interviews and requests.
Based on this initial complaint, draft a list of follow-up investigation questions. Cover both parties and any witnesses, and avoid leading language.
Prompt 3: Run a bias check on your own summary
When we're this close to a case, none of us can see our own bias. A second set of eyes helps, whether that's a manager, a colleague, or a well-prompted AI tool. This one costs about ten seconds.
Review this investigation summary I wrote. Point out any language that assumes intent, takes sides, or isn't supported by the notes.
Give it as much context as you can, because if the model can only see the summary, it can't tell you what the notes actually support. Again, anonymize PII where needed.
Additional AI prompts for HR Relations & Investigations
Those three cover the core of an investigation file. Here are three more I use. They all follow the same pattern: a clear task, boundaries on what the model should and shouldn’t do, and a request to show its work.
Build a timeline
Investigations are full of dates scattered across interview notes, emails, and chat messages. As we just saw, those dates don't always agree. This prompt pulls them into one place.
Build a chronological timeline from these notes and emails [paste]. One line per event, with date, who was involved, and the source. Flag any dates that conflict between sources.
If your company has an investigation template, follow up by showing the AI the format it needs to fit, then copy the result into your doc.
Match the complaint to your policies
As companies grow, policies end up scattered: an updated PTO policy in one doc, a handbook in another, a wiki that's out of date. This prompt connects a complaint to your own handbook.
Here's our handbook [paste] and a summary of the complaint [paste]. Which policies might apply? Cite the exact sections. Don't decide whether a violation happened.
That last sentence is the critical one. It keeps the AI from becoming the judge and keeps me as the reviewer. And if your handbook is out of date, fix that first.
Review a performance improvement plan
Review this performance improvement plan. For each goal, tell me if it's measurable and time-bound. Rewrite any that aren't.
FAQ
How do I handle PII in audio, video, or other evidence?
Start with your company's security posture and ask your IT or security admin for guidance. Then confirm your AI tool's enterprise tier handles sensitive data the way your policy requires. For text, anonymize with find and replace before you paste.
Should I set up a project for each investigation?
For repeatable tasks like these, I'd build a skill instead. A skill is a long doc that tells the AI what to check and do for a given task, and you can use it in any chat or project.
Next up: benefits, leave, and comp
This series runs every two weeks, and it’s built around what you want to hear. Coming up: benefits administration, leave of absence and leave management, compensation planning, and HR reporting best practices. See the upcoming sessions here.
