Name: AI Prompts for HR: Employee Relations
Uploaded: 2026-10-06T16:46:19.947Z
Duration: 34 min 48 s
Description: AI Prompts for HR: Employee Relations

Transcript for "AI Prompts for HR: Employee Relations": Great. Hello. Good morning. Good afternoon, to everybody. How are we doing on this beautiful, fall, fall afternoon? Awesome. Great to see, so many familiar faces, in, on this webinar. I think, yeah. Jeanette? Oh, in Upstate. Where are you in Upstate New York, actually? Is it fall yet? Watertown. Oh, beautiful. I've been meaning... North Of Syracuse. I've been meaning to do my, annual fall foliage trip. I'm calling here from our office, here in Flatiron. Already started noticing a bunch of the trees, changing, changing leaves here. Well, just wanted to say hello, everyone. It's great to see some familiar faces from our last session. If this is your first time with us, my name is Gabriela. You can call me Gabby. I am our chief people officer and chief operating officer here at WARP. A little bit on my background, I'm actually originally Colombian. I'm from BOTA, and then I later moved and, grew up in in The United States. Back in the day, I studied mechanical engineering and computer science, at MIT, so up in Boston. And then after that, spent my years as an engineer. So I've built hardware, at Apple. I built software products at Facebook. I was recently a product manager at Google, and then, most recently joined WARP about a year ago to run operations and people. I do come from a bit of a more nontraditional background, from, from the people side. But in general, I've always treated HR and people operations like an engineering problem. So building out systems, understanding workflows, driving great processes and documentations, and identifying the edge cases on what, what is working is what is not. And how do we design around that as our company grows and scales. And so that is very much a lens, that I bring, to AI as well. Maybe, in the meantime, I'll start, screen sharing, as I continue talking. Okay. Great. Thank you, Rachel. And so... Hey, Tammy as well. Fabulous. So this is our second, our second session in this series, which is on AI prompts for HR. So first, why are we doing this series? Recently, a couple of months ago, OpenAI came out with a fantastic economic, research study that found that when HR professionals use AI, only 10% of them are actually using it for HR work. We want to change that. We want to empower and enable folks like yourselves, to be really able, you know, to take advantage of what AI can use, for your day to days. And so today we're covering, a topic, that isn't very, very important to me, but is also the top, topic that you guys asked about, which is employee relations and investigations. Maybe some quick, house court... Housekeeping before we get started. This session is recorded, so we'll be sending it to you afterwards along with every prompt that we run today. So you don't need to copy anything down. Feel free to drop questions in the QA at any time. I do have a portion of the, later half, of today's session dedicated to answering all of your questions. So maybe, before, we kick things off, I'd love to start out with a poll. Feel free to, answer, in this section. If you've used AI... I'd love to hear, have you used AI for any employee relation related workflows? Okay. Great. We've got a few. Wow. Neck and neck. Yes and no. Okay. More so on the nose. Let's see. Any final holdouts? Any final holdouts? Okay. No worries. I know we've got some folks, who are a little, are a little bit nervous. Okay. Great. So why don't we, close this poll? So most of you guys said no. That's really what I expected. And it's a very good, place to start. I think a lot of us are very cautious, especially Leanne, as you put in the chat, around confidentiality, discoverability, what type of information are we inputting and how could that be tied up to ourselves? I personally think a lot and I care about and I'll talk about this later is how do we do this in a safe way? How do we not just continue digging deeper and deeper into our own inherent biases? And so today is really built, for you guys. Great. We'll move over, to slide two. Okay. Great. Perfect. So, today we'll be covering, three topics and prompts overall, you know, for for, investigations. One thing to touch on is specifically on safety. Some of you here might be thinking, is it safe to use AI in an investigation at all? Why should I be doing it? How can I be doing this safely? Here's how I think about it. I think AI is very good at drafting. It's very good at organizing in your approved corporate tools and what you can do, and you'll see this later in how I designed the prompts, is you can use placeholders for any PII. That's why you'll see in brackets, employee A and in bracket, employee B on stage on the screen today, instead of names. Great. So now looking at the prompts, I think I philosophically and I think a lot of you guys would agree as well is that AI should not be the judge in an investigation. At the end of the day, it is our responsibility for gathering information, understanding what is the right process. In many cases, you know, what is the legal guidance and the deciding how to run out the course of an investigation. However, I do think AI is good at very, at two things. Number one, turning messy notes into neutral documentation. And then two, catching your own loaded language. As I said, we'll be going over three prompts today. Starting out with a neutral summary. Then we'll be looking at some follow-up questions. And then finally, we will be covering, a bias check for ourselves. So now first, starting with the neutral summary. I'm sure many of you guys have been involved, you know, in, an employee investigation, whether it was a quick Slack DM that came to you about some questionable behavior or you heard something in the office. And so this is the prompt that I really use the most, which is after I do an interview with all involved parties, my notes are raw, right? They're messy. They have shorthand. They have what people assume to be right. I might even have my own reactions, mixed into what people say. And so this prompt is very helpful for turning it... Turning all of these messy notes into a fact... Into a factual a factual summary for our file. So as you can see, it's summarized and it's, it's actually structured in three different parts. So first, report only what is said and observed. This is very important for setting the boundary with the AI. Second, in the next line, you'll see no conclusions about intent or credibility. This... It places those guardrails so that the model stops playing the judge. And then third, flag any word choices that sound judgmental. This is really important because once again, the emphasis is keeping you as the reviewer. One thing as well, you know, in here as well, we can use the words judgmental and secondhand here. Secondhand also matters a lot because interview notes are frequently full of what people said, this other person said this or not. Great. Now we'll go into the next slide. Perfect. This is the thing that we'll do afterwards, which is once you have a clean summary, your next job is fragrant. What don't you know? Right? This is a great prompt for drafting follow-up questions for your next interviews or next requests. So based on this initial complaint, draft a list of follow ups, in follow-up investigation complaints, cover both parties perspectives. You can also include any witnesses as well, which is important, and avoid, using a leading language. Fabulous. So now we'll actually do this in practice. So I will switch over, to my Claude, Claude instance. We, overall as a company, we use a enterprise, version of Claude, which is a product from Anthropic. One note, I know we talked about this, on our last, on our last call as well. But we use the enterprise tier because there are different requirements in terms of keeping data, and what is what is required by our security and IT team as well. Great. So now starting off, with, the neutral summary. So I'm gonna take over and start inputting based on topic number one, which is turn these raw notes into a factual neutral summary for an investigation file. I also inputted, as you can see right here... Whoops. Let me move this to the side. The actual raw notes, which I pulled from my meeting tote... My, my meeting note taker tool. And you can see interview started with x, interview started y. This is what people said or not. Great. So maybe before I run it, let's look at the notes together. They're realistic. They're reasonable. They're what a well meaning HR person writes at 4PM after a very hard interview. But if you can start to notice as you read the notes, there is loaded language in here. So as this runs, we'll see, what the AI flags. And we'll see if it, if it if it highlights any other key phrases. Great. So here we have the interview summary. It has all of the locations, who is present, all of the remarks, any observations, any open, items for follow-up, which we'll talk on later. And then I think this is very important, you know, for us today, which is flagging that language, particularly from those notes. So let's see what what it says. Starts out with clearly upset. Clearly upset, is an observation. Crying is the actual item. The neutral version of this, as you put it in an investigation file, is to say a person was crying at certain points, which is what you saw rather than using a loaded language of clearly upset. Looking at the next part, right? Obviously didn't take it seriously. That's a very loaded statement to say. There's a lot of context with why, someone, you know, may have interpreted this as well. And overall, this is a conclusion about someone's intent, and is not what factually occurred. Great. And then maybe touch on, kind of the final point in here, you know, to be respectful of time. Using the term history of being aggressive, is a secondhand statement. It's not backed by facts. It's also highly nonspecific. And so this would get flagged to you as the reviewer to really chase down. It's important that some of these biases don't end up, don't end up snowballing, you know, in an investigation file. Great. Also, one thing to highlight, is, and this is why it's very important to give that prompt on being, you know, for, you to be factual, and to not sound judgmental is notice what the AI did not say. It didn't say whether employee a was credible. It didn't say whether employee b was non credible. And instead it really outlined what happened when. Great. Final thing, to note, and I think this is really why it's very helpful to throw these things into a well well set up prompt is if you had noticed in my initial raw investigation notes, there is a discrepancy between the dates. Employee a said the comment, happened at, on September 10 on her stand up. But the Slack message that she forwarded, the Slack message that she forwarded is actually time stamped to September 11. A good summary doesn't pick one. It records both and allows us as the witness and the reviewer to decide what to happen with that. So now let's see if, if the AI actually caught that discrepancy. Let's see. Happened September 10. Forwarded a message. Reported concerns. Fabulous. Here it is. Exactly to that point. Open items for follow-up. The September 10 stand up remark and the September 11 Slack message both... Oh, actually, that's the wrong thing. Great. So, actually, maybe this is a fabulous example of why you shouldn't be using, you should always be the reviewer, is I noticed this discrepancy, but, actually, the AI, didn't catch that. Great. So now we're gonna go, into prompt number two, which is, based on this initial complaint and the summary, draft a list of follow-up investigation questions. Cover both parties' perspective. Once again, avoid any leading language. Also, any, also include any witnesses. Great. One thing that's really powerful here, as I'm sure you guys understand, is there's always a mix of as the HR and people leader directly conversing with an employee versus... And then doing that investigation directly with them and then any other parties involved, and then also getting the group involved, and doing overall, you know, joint discussion on what occurred. What's really great about establishing a prompt like this is it does highlight directly per individual what do you need, what do you need to ask, to each one. So now here are the open questions for employee b. Can you describe the working relationship, etcetera, for the manager, for the witnesses, any other potential witnesses? How should be... We be recording these in, these calls in general? Once again, actually, the AI did not flag, the date discrepancy. So let's ask a follow-up. Is... Are there any date discrepancies in the notes? Great. Great. So it just highlighted it. I think this is why we should always be thinking about these AI systems as, in a best case, a friendly companion to your work, rather than delegating certain tasks to yourself. And in general, we should always be reviewing, their output. Great. Now we'll go back, and switch over, to the slides. So I'll stop screen sharing. Great. So here's an example, that, I pulled from my... Right here. Perfect. From one of my people ops, operations, colleagues, here at WARP who was testing out one of these prompts for a certain, for a certain event. And so this is the draft investigation summary, that they wrote and they sent to me. What jumps out, to you guys? Let me read out the summary in general. First, employee a credibly reported employee b repeatedly mocked her accent during team stand ups. B's comments on September 10 was clearly intended to humiliate a in front of colleagues. B has also deliberately excluding a from client calls, which appears retaliatory. Manager c failed to act on an... On a's initial report showing a dismissive attitude towards the complaint. Given B's known pattern of aggressive behavior, A's account is consistent with how B treats others. So why don't in the chat, before I go back, and run any prompts, are there any words that are loaded language, secondhand language, maybe showing biases that drop out at you? Janet. Yes. Great flag. Clearly intended, to humiliate. Rachel, also great note. Deliberately. Great job, Suzanne. Mocked, intended to humiliate, deliberately, dismissive, aggressive. Arian, thank you so much. Appears, I probably can't even pronounce that word. Retaliatory. Apologies. English is a second language. And then finally, Cherry, thank you so much for highlighting, once again, pattern of aggressive behavior. Great. So now we'll move over, you know, to slide, to the next slide, which is on the bias check. Great. So now I'll move over, to our live demo once again. Great. One quick note, for the folks on the call. I typically like, to continue working in a single chat, as it has all of the contacts and all of the questions of what I was working on before. You'll often see if you have, for example, a downgraded version of a chat or a free version of one of these AI tools that you can quickly run out of context. One quick tip that I typically do is if I can tell that the model is acting slower or it's flagged to me that it's running out of tokens, think of that as like space, in the chat. I will ask it to write a very detailed document on what happened in a chat, and then I'll throw that in to start up a new chat. This helps building up the context in a separate chat and also ensures that you don't run out of context. Given that I'm not running out of context, let's go over, and continue working in this single chat, and focus on checking out the bias check. So once again, this is my request, which is review the investigation summary I wrote. This is actually what my colleague wrote. Point out any language that assumes intent, take slides, take slides, and isn't supported by the notes. Great. So here's the summary. It's very important that you add as much context as possible. The AIs can only help given by what they know. And then notice once again, I'm including this here, you know, just as an example for you guys. I am uploading again the raw investigation notes. Great. If the model can only see the summary, it really can't tell you what it supports. Fabulous. That was pretty fast. Every sentence in the summary has at least one problem. I think a bunch of, you know, the folks in the channel, have highlighted certain things. Clearly intended to humiliate, assigns intent. This doesn't speak to that. Credibly, credibly is a credibility, determination. Deliberately excluding, once again, deliberately is employee a's belief, which it then restates as a fact, etcetera etcetera. One great thing, that I didn't ask it, to do, actually, which is quite helpful, is it went ahead, and actually revised my summary or, what my, colleague summary is to remove these loaded languages. So why is it important, you know, to remove via bias? I think in general, right, like, the role of the people leader and the role of HR, is difficult. Right? You're dealing with investigations. You're dealing with situations in which it's the people around you and it's okay, you know, for that to be sensitive and for that sometimes, you know, to, to be emotionally difficult for you. And so understanding where you can check yourself, check, what you're doing wrong, check your biases, and how can you improve is really critical. And then for myself as a manager, I am constantly thinking about how do I up skilled, upskill my team? How do I help them grow in the people function? And how can I give them very text, contextual, but also actionable ways, for them to improve their processes? Great. And so as we kind of highlighted, there's seven different phrases, or five phrases that highlights here, the known patterns of, aggressive. Great. Perfect. And so I think the overarching point here, and Susan, I'll go to your question right after this is, we have to assume that nobody wrote this in bad faith. It's really that none of us quite often can see our own bias when we're this close to a case. And so having a second set of eyes, whether that's the manager reviewing, one of your colleagues reviewing, or in this case, you know, with a very well prompted, AI tool can really enable you to have that second set of eyes. It costs you ten seconds. And in general, I would say, you know, those ten seconds are worth having to ensure that, this process works well. Great. So maybe going into the chat, Susan, you asked, how do you deal with PII in other evidence such as audio recordings, video, etcetera, that you may want to have the AI ingest and respond to as part of the investigation? Fantastic question. So the way that I do first, and this is what I would recommend you do at your company, is check the security, posture of your organization. So you can go to your IT admin or maybe your security admin and ask them for guidance. It... At the end of the day, it depends on the broader company posture on what should be included or not. Given that we are a business that deals with sensitive data and we have a variety of security, data security, and like, and like compliance requirements that we adhere to, we take security very seriously at Warp. And so in general, what I have to usually do is depending on the notes, I'll have it in a Google doc. I can press control F to find certain names, and then I can replace all of those names with broader, broader, and more generic answers. And so the real answer is check your company security posture. Number two, ensure that, like I said, for many of these AI tools, they do have, PII in mind if you're on the enterprise tier. So check that as well. And then there are little hacks here and there, you know, to be able to solve them. Great. So maybe I'll stop demoing here, because, to wrap up in the next couple of questions, in the next couple of minutes, as well as, to give, time for questions, I'm gonna go back to screen sharing to show some of my other additional prompts, that I would recommend. Great. So these three prompts cover the core of an investigation file. So before we open it up for questions, also, once again, we're recording the session. I'll be sending over, you know, what are the exact prompts that I use, you know, day to day. I won't be... I won't demo these, but overall, they're showing a clear, a pattern, a clear task, a boundary of what the model should do or should not do, and then a request for it to show its work. Great. So next one, building... First, example prompt that I'll give to you guys is building a timeline. As we all know, investigations are full of dates that are scattered across interview notes, emails, chat messages. And as we saw in the first investigation prompt, quite often these dates don't agree. So this is the prompt that I've designed. So build a chronological timeline from these notes and emails. I'll dump all of that data in here. Once again, given, my company security posture, include one line per event with the date, who was involved, and the source. Flag any dates that conflict between sources. Depending on what your template for, employee investigations is, we have a template that we use in which we output this into a, into a, into a table. You can then follow-up, with this prompt, give the example of what it needs to be fit in terms of a template, and then copy and paste that into your doc. Great. Maybe going to the next one. Matching policies. I hear this all the time. You know, as companies, grow, they have so... They've got a company handbook. They have updated PTO policies in one document, but that's not been translated over into the company Wiki. And so what this prompt helps you do is really connect complaints to your own handbook. So here's our handbook. In our case, it's a PDF, and a summary of the complaint, which I'll paste it in, which policies might apply, cite the exact sections, don't decide whether a violation happened. Once again, that sent... That final sentence is critical, because we want to ensure that the AI is not being the judge. We want to be the reviewer. Great. Maybe I'll skip the next ones and go straight, to the q and a section in the next couple of minutes. Fabulous. So maybe I'll go up to Tammy's question, which is, do you suggest having legal review AI suggested suggested follow-up questions or a set of questions used over multiple witnesses in general? I think number one, is, as always, these things are painful. You know, you don't want to have a long series of follow ups, double checking to a certain point. May... Double checking is good, but maybe not triple checking certain points. And so in general, I would recommend as you design what follow ups are needed and where you need more information, Be highly curated in, like, what is actually going to, like, move the needle with this investigation, who actually needs to know what. And so I start with a overall framework of what are the big questions that I have in terms of who did what, what is my policy. And then from those broader themes that I then need answers to, I typically go and design follow-up questions per an individual. In some cases, and this is why every investigation is different and I think we need to give ourselves grace as well, you know, to be flexible with these, is that we might be... Need to ask the same question to multiple parties. We might even need to design a meeting with multiple parties together at once, to really get to the root of the problem. Great. So thank you, Tammy. And now final follow-up question. I'll go over to Leanne. Do you typically create these as projects? I love this question. I imagine you use the same prompts for each investigation, but not sure if that, if it, that makes sense to structure it as an individual chat or project within Claude. Fabulous question. And so I overall use, once again, we use Claude, as a company. And so I have various projects for different teams, whether I actually have a Claude project for HR, I have a Claude project for people operations, and then I run a couple other teams. What I would actually recommend for these repeatable tasks, Leanne, is you can actually create something called skills, which is you write up a long doc on these are the things that I need to check, on this. And then what you can do is ensure that each of your chats within a project or within another project can then be using one of these skills to do these repeated tasks. Some examples of skills, that we've designed, as a team is, for example, on the copywriting front, here is a, a list of, of qualities on how I like to communicate. Now review this output in Gabby's words, you know, to edit. So I've designed copywriting skills. I've designed a couple document processing, types of skills. And so in general, I would recommend if you have these repeatable, repeated prompts, see if you can actually extend them to a skill. What's also powerful about that is depending on the skill, you can actually share that across a team. And so that's a really great way, you know, to empower not only yourself, but the rest of your people org to do these repeated tasks. Fabulous. And so I know we're out of time, but, before we go, here's what's coming up in some of our future sessions. Once again, we'll be hosting these AI prompts, for HR every two weeks, based. And a lot of this is also designed on what you guys wanna hear. So feel free, you know, in, in the email follow-up, you know, that we'll be sending out of the, after this as well as in certain polls to give your feedback on what is important to you. So in the next section we'll be covering benefits administration. This is actually very contextual, actually to hear at WARP because we're doing a mid year, plan change. And so how can you help empower, you know, like, benefits knowledge across a team? What are those types of tasks? Number two, leaves of leave of absence and leave management. I know it's a very timely, topic right now for a lot of people. After that compensation planning and then overall, HR reporting best practices. Once again, if your topic isn't on that list, drop it in the chat, follow-up with us in the survey, and then just drop it in the next webinars registration as well. Once again, thank you all for spending your time with us. It's an absolute pleasure. Great to see so many new faces as well as some repeated faces in today's chat. We'll send the recording and every prompt from from today, within the next twenty four hours. I'd love to hear your feedback on how they're working for you. Use them on your next case. And overall, you know, keep the one rule in mind. AI can draft and check, but you decide. Awesome. Take care everybody, and have a great rest of your Tuesday.