A working system for disciplinary write-ups, built around the one fact most managers never get told: the document you write today is the document a lawyer reads first if this employee is ever terminated.
The EEOC received 88,531 new workplace discrimination charges in fiscal year 2024, up 9.2% from the year before, and recovered nearly $700 million for victims (EEOC, FY2024 Performance Report). Wrongful termination settlements average around $40,000, and that’s before legal defense costs get added on top (Embroker). Gallup has also found that companies pick the wrong person for a manager role 82% of the time, so most disciplinary write-ups get drafted by someone who was never actually trained to write one (Gallup, Only One in 10 People Possess the Talent to Manage). I’ve seen AI close that skills gap surprisingly fast. I’ve also watched it make a bad write-up worse in about the same amount of time, if nobody checks the output before it lands in someone’s personnel file.
Picture a manager with five performance reviews due by Friday and a disciplinary write-up he’s been putting off for two weeks. He opens ChatGPT, pastes in a few rough notes about an employee who keeps missing deadlines, and asks for something that sounds professional. The AI hands it back polished in under a minute. He tweaks one sentence, signs it, and gets back to the rest of his week. I’ve sat across the table from managers doing exactly this, and honestly, I get it. Nobody trains you for this part of the job.
Three months later, that employee is terminated. Six months after that, a wrongful termination lawsuit lands on the company’s desk. During discovery, the manager admits under oath that AI wrote most of the warning that helped justify the firing. At that point the company is no longer defending one termination decision. Legal has to explain why nobody in HR knew a chatbot was involved, and why this warning reads nothing like the one issued to a different employee six months earlier for a near-identical problem.
That scenario isn’t a thought experiment. Employment law consultants tracking this are flagging it as a live, current risk, built directly on what’s happening right now inside companies with zero policy on managers using consumer AI tools for HR work [1]. I’m not telling you this to scare you off AI for disciplinary documentation, I use it myself when I’m training HR teams. I’m telling you because the fix is genuinely simple, and in my experience, almost nobody has actually done it yet.
I tell every manager I train the same uncomfortable thing: the write-up you draft this afternoon doesn’t just sit quietly in a folder somewhere. If this employee is ever terminated and pushes back, that document becomes the company’s star witness. Write every single one like a lawyer might read it in three years, because on a long enough timeline, some of them will.
Employment attorneys have a three-word mantra for defending a termination decision: “Documentation, documentation, documentation” [2]. I’ve heard managers repeat that phrase back to me in training sessions and still get the point backwards. A write-up is not a case file built to prove the employee is bad. Think of it instead as an honest, dated record: this conversation happened, this is the problem we discussed, and these are the expectations I laid out afterward [2].
A lot of managers assume at-will employment means none of this matters, that you can fire someone for pretty much any reason and skip the paper trail. That’s only half true. At-will employment means you don’t need good cause to end someone’s job. It never means you can fire someone for an illegal reason: discrimination, retaliation, or exercising a legal right [3]. Weak or missing documentation won’t make a termination illegal on its own, but I’ve watched it turn a defensible decision into an expensive one, simply because there was nothing specific to point to when a plaintiff’s attorney went looking for a pattern.
The EEOC’s own investigators are explicit about what they look for when a discrimination charge comes in: performance evaluations, disciplinary records, and personnel files, checked for dates, patterns, and whether other employees got treated the same way for similar conduct [4]. I’ve watched HR teams pull these files for an actual investigation, and vague or undated write-ups are the first thing an investigator flags. If yours look like that, you’ve handed them exactly the pattern they’re trained to spot.
And this isn’t some rare event. The EEOC took in 88,531 new discrimination charges in fiscal year 2024, a 9.2% jump over the year before, and recovered close to $700 million for the people who filed them [5]. Separately, the average wrongful termination settlement runs around $40,000, before legal defense costs, which climb fast once a case moves past the initial filing [6]. I’m not saying this so you panic every time you write someone up. I’m saying it because the document itself is worth doing well, whether AI helped write it or not.
I wish documentation alone could protect a company from a discrimination claim, but it can’t, not by itself. Its value is smaller and more specific than that: it’s the thing HR points to when they have to explain, in plain English, why a decision had nothing to do with a protected characteristic. Vague documentation leaves them with nothing to point to, and I’ve watched that gap sink termination decisions that were actually fine on the merits.
Most workplace guides skip this fact: Gallup has found that companies pick the wrong person for a manager role 82% of the time, and only about one in ten people naturally have the mix of talent the job actually needs [7]. In every workshop I run, I ask how many people were ever trained to write a disciplinary document, and it’s almost always nobody. AI is genuinely good at closing that particular gap, but only if you use it deliberately instead of typing “write me a professional-sounding warning” and hitting enter.
The single biggest improvement AI brings to a write-up is forcing specificity, and honestly, that’s the fix that matters most. Untrained managers default to character judgments like “has a bad attitude” or “isn’t a team player,” because that’s what’s actually bothering them day to day. Those phrases are close to useless in a write-up, and they can read as evidence of bias if the employee later disputes it. A well-prompted AI tool pushes the other direction: dates, observed actions, direct quotes instead of gut impressions.
This is the prompt template I hand out in workshops, and it works far better than typing “write a warning for an employee who’s been late”: “Based on these notes, draft a disciplinary write-up. Use only observable facts and direct quotes from what I’ve described, not assumptions about motive or character. Reference the specific policy number if I’ve given you one. End with a consequences section stating that failure to show immediate and sustained improvement may result in further disciplinary action, up to and including termination. Do not use vague terms like ‘unprofessional’ or ‘bad attitude’ unless I’ve described the specific behavior that earned that label.”
Notice what that prompt is actually doing: it front-loads the same rules employment law experts have been teaching managers for decades, long before AI ever existed, about sticking to facts, favoring behavior over character, and using clear consequence language instead of vague threats [8]. None of that is new. AI just applies it faster and more consistently than a busy manager relying on a training deck they sat through two years ago, if they sat through one at all.
If you’re building out a broader library of prompts for people-management work, pair this one with a structured approach to preparing for a difficult conversation with an employee. I always tell managers the write-up and the conversation that goes with it need to say the exact same thing, in the exact same words, or the employee will notice the gap immediately.
I run into this constantly: managers treat the AI draft like a finished document instead of a first pass. Using AI isn’t really the issue. Skipping the review step is, and it’s easy to do in a hurry. A workable process runs four steps, and step three, checking the new draft against precedent, is the one that gets skipped most.
The four-step process described in this guide: draft fast with AI, but never skip the human review step before it reaches an employee’s file.
Step one is boring but non-negotiable: gather the actual facts before you open any AI tool. Dates. Exact quotes if you have them. What policy got violated, and whether this is the first time you’ve raised it. AI can’t invent facts you never gave it, but it will happily fill gaps with something plausible-sounding, and plausible-sounding is not the same thing as true. I learned that the hard way once, reviewing a draft that sounded reasonable and was wrong on two dates.
Step two is the draft itself, using a specific prompt like the one above instead of a generic request. Step three is the part that actually protects you: pull up the last write-up issued for a comparable offense, on this team or a different one, and check whether the language and consequences match. It’s the single fastest way to catch the kind of inconsistency that turns into a discrimination claim later, and it’s also the step I watch rushed managers skip more than any other.
Step four is delivery. Discipline conversations should happen in person, not over email, so the employee can ask questions and you can confirm they understood what was said [8]. Get a signature acknowledging the conversation took place. If the employee refuses to sign, note that in writing, and where policy allows, have a second person in the room to confirm it [2]. AI involvement doesn’t change any of this. If anything it matters more, since a write-up that reads a little too polished can raise its own question about whether a real conversation happened at all.
My rule of thumb, the one I give every manager I train: if you can’t explain, in your own words, why this write-up says what it says, you’re not ready to hand it to the employee. Doesn’t matter how polished the AI-generated version sounds.
Employment law consultants tracking this issue describe a pattern that shows up constantly once managers start using consumer AI tools for HR work without any oversight: nobody in HR or legal even knows it’s happening, because it’s all occurring on individual laptops with tools nobody approved [1]. I see some version of this in nearly every company I walk into that doesn’t have an AI policy yet. Three specific failure modes come up again and again.
Two managers using two different AI tools, with two different prompts, will produce write-ups that emphasize completely different things for what’s supposed to be the same policy violation. One manager’s AI-assisted warning leans hard on “attitude.” Another leans on missed deadlines and never mentions attitude at all. Employees compare notes. They always do, and inconsistent treatment for comparable conduct is exactly the pattern a discrimination claim gets built on [1].
AI models are trained on enormous amounts of internet text, and that text carries the same biases people carry. In HR documentation, this shows up as language that quietly correlates with protected characteristics: different word choices to describe the same behavior in men versus women, or coded language around someone’s age or communication style. Honestly, most managers have no idea it’s happening. The AI suggests a phrase, nobody catches it, and it ends up sitting in an official personnel record [1].
If a write-up was drafted with AI help and this employee is later terminated, the prompts themselves can become part of what opposing counsel asks for in discovery. What did the manager tell the AI about this person. What did it hand back. A company with no idea AI was even involved, and no record of any of it, is stuck defending a decision it genuinely can’t explain. “We didn’t know managers were doing this” carries zero weight as a legal defense, and it’s a miserable place to be sitting during litigation. I’ve watched companies get there [1].
I don’t read any of these three problems as a reason to ban AI from disciplinary documentation. Banning it just pushes the same behavior further underground, out of sight of HR entirely. What actually fixes this is a real policy: name the approved tools, spell out what has to be disclosed, and say who signs off before a draft becomes permanent. Most companies I talk to still don’t have one, and it’s the missing policy that’s the real gap, not the AI itself.
This is the part that separates a write-up that protects the company from one that quietly turns into a liability, and none of it is new, it predates AI by decades. A well-known guide for HR practitioners lays out documentation habits that get managers, and their companies, in trouble [8]. Left unsupervised, AI will happily reproduce every one of them. I check for these first whenever someone asks me to review a draft.
Large language models also carry a documented tendency researchers call sycophancy: a pull toward smoother, more agreeable-sounding output, even when the source material was sharp and specific [10]. In a disciplinary write-up, that tendency can sand a clear, factual account of repeated lateness into something too soft to support the consequence you’re about to hand down. I’ve read drafts like that and nearly missed it myself. Read the AI’s draft against your original notes before you send it anywhere. If it sounds nicer than what actually happened, something got lost.
My fast gut check before signing anything: hand it to a stranger with zero context. Can they tell exactly what happened, on what date, and what happens next if it continues? If the answer is no, the language is still too vague, whether AI drafted it or you did.
Let’s be honest about what a free, consumer AI tool actually is: a system outside your company’s access controls, run by a third party, with no real guarantee about how your input gets stored or used. Pasting an employee’s real name, performance history, and personal details into ChatGPT to draft a disciplinary record treats sensitive personnel data like it’s disposable. It isn’t, and I wince every time I see it happen.
One data point worth sitting with: in a mid-2025 survey of over 600 HR professionals, the share with zero plans to use AI at work dropped from 35% to just 7% in about a year [11]. That shift already happened, policy or no policy, and I’d rather see a company spend a quarter writing an actual policy than another one debating whether AI belongs here at all.
Quick test: can you say, in one sentence, which AI tool your managers are actually allowed to use for disciplinary write-ups? If not, the honest answer right now is “whatever’s open in their browser,” and I’ve heard HR leaders admit that out loud. Fix that sentence before you touch anything else on this list.
Yes, using AI as a drafting tool is legal. The real risk shows up when nobody reviews the output for accuracy, consistency with past write-ups, and language that could read as biased or vague. I treat every AI draft as a first pass, something a manager or HR reviews and edits before it ever reaches an employee, never as the finished document.
Potentially, yes. If a terminated employee sues, the write-up becomes evidence, and if AI was involved, the prompts used to generate it can become part of what’s requested during discovery. I don’t think that means you should avoid AI assistance. It means you need a clear policy on which tools are approved, plus a real review step on every draft before it goes in the file.
Avoid pasting an employee’s full name, other identifying personal details, or anything already flagged for legal review, things like accommodation requests, harassment complaints, or anything an attorney is already handling. Use a placeholder like “the employee” while you draft, and finalize the real document inside your company’s actual HR system, not a public chatbot. I tell every group I train to treat this like a habit, not a one-time reminder.
Give every manager the same specific prompt template instead of letting each person ask AI for “a professional warning” in their own words. Build in a review step too, where someone checks the new write-up against the most recent one issued for a comparable offense. Inconsistent language and inconsistent consequences for similar conduct are exactly the pattern that turns into a discrimination claim, and it’s the first thing I check when I audit a company’s files.
No. At-will employment means you don’t need good cause to end someone’s job. It never permits firing someone for an illegal reason, including discrimination or retaliation. Weak or missing documentation won’t make a termination illegal on its own, but it makes the real reason far harder to defend, which is exactly what a plaintiff’s attorney goes looking for. I’ll say this plainly: this is general HR practice guidance, not legal advice. For your specific situation and jurisdiction, talk to employment counsel.
I researched this by checking SHRM’s and the EEOC’s own published guidance on disciplinary documentation and discrimination-charge investigations, cross-referencing wrongful termination and EEOC enforcement statistics against Embroker’s and the EEOC’s own fiscal year 2024 reporting, and grounding the AI-specific risks in reporting from an employment-law-focused AI consultancy and a published survey of HR professionals’ AI adoption. The point on AI summarization softening specific language is grounded in published research on sycophancy in language models. This article explains general HR documentation practice; it is not legal advice, and specific situations should go through your employment counsel.
Weekly tips. Real examples. Practical help for busy professionals.
We care about your data, check out our privacy policy.