How to Draft a Performance Improvement Plan Using AI

What a performance improvement plan using AI changes

AI does not write the plan. It converts what you already know into language that survives review, and it does that in minutes instead of the afternoon this normally costs you.

The gap between a weak PIP and a defensible one is almost never effort. It is specificity. “Needs to show more ownership” is a sentence nobody can pass or fail. A sentence with a finish line is a different object entirely.

The second version is not softer. It is harder to argue with, which is the entire point. An employee can dispute an opinion. They can only respond to a fact — and responding is what a plan is supposed to make possible.

The method we teach at PromptPeakAI for this is the Defensible Draft Method: five gates, run as five separate AI passes, each with exactly one job. Strip, State, Standard, Support, Stress-test.

Running five passes instead of one matters more than it sounds. Ask a single prompt for a whole PIP and you get a plausible-looking template with your situation sprinkled over the top. Chaining the steps keeps each output small enough that you can actually check it.

Set your privacy boundary first

Before a single note goes near an AI tool, settle one question: what type of account are you signed into? Not what you pay for it — what type it is. That distinction is where most managers get this badly wrong.

Consumer accounts and business accounts sit under completely different terms. On the consumer side, including the paid tiers, your conversations may be used to train the model unless you switch that off yourself. On the business side, OpenAI’s enterprise privacy documentation states that business and enterprise data is excluded from training by default unless a customer explicitly opts in. Anthropic draws the same line for Claude.

The trap is the paid consumer tier. People upgrade, quietly assume the money bought privacy, and paste an employee’s full name into a chat window. It did not buy privacy. It bought higher limits. Our guide to whether ChatGPT is safe to use at work walks through the specific settings.

None of this means you cannot draft a performance improvement plan using AI on a personal account. It means you sanitise first and let nothing identifying cross the line — which, as it happens, produces a better draft anyway. Stand-in tags stop the model anchoring on a name and inventing a personality to go with it.

Whatever account you are on, strip the notes first. Redaction is not only about the vendor — it is about your own chat history, which on a shared workspace a colleague may be able to search.

Strip and state: the first two gates

Gate one removes anything that identifies a person. Gate two turns your raw observations into sentences you could read aloud without wincing.

Pull together what you already have: calendar records, delivery data, notes from your last one-to-one, and the themes from any 360 feedback you have summarised. Then run one pass whose only job is sanitisation. Do not ask it to analyse anything yet — a model given two jobs does the interesting one properly and the boring one carelessly.

You are a data sanitisation assistant. Do not analyse, summarise, or comment on the content.

Task: rewrite the notes below with every identifying detail replaced by a neutral stand-in.

Replace real names with [Employee A] and [Manager B]. Replace company or client names with [Client One]. Replace project or product names with [Project One]. Replace team names with [Team One]. Replace any date that would identify a single person with [Month Y].

Keep job responsibilities, quantities, and the sequence of events exactly as written.

Output only the rewritten notes. No preamble, no commentary.

Notes:
[paste your raw notes here]

Read what comes back before you go further. Models miss things — a product codename, a date only one person could own, an incident so specific it names itself.

Now convert. The instruction doing the real work here is a negative one: ban the vocabulary of judgement outright. Words like “lazy”, “unprofessional” and “poor attitude” are what turn a legal document back into an opinion.

Give the request the Role, Goal, Context, Format shape and it holds that line far better than a one-line ask does.

Role: You are a senior HR business partner reviewing a manager's rough notes.

Goal: Convert the notes below into three to six factual issue statements for a performance improvement plan.

Rules:
- Describe observable behaviour and its effect on the work. Nothing else.
- Do not use judgement words: lazy, careless, unprofessional, poor attitude, difficult, negative, unmotivated.
- Do not invent any number, date, or incident that is not in my notes. If a detail is missing, write [detail needed] rather than guessing.
- One statement per issue, two sentences maximum.

Format: a numbered list. Nothing before it, nothing after it.

Notes:
[paste your sanitised notes here]

Watch for invented specifics. If your notes said “missed some deadlines” and the output says “missed four deadlines in Q2”, the model filled a gap for you — that is how hallucination works. Delete any number you did not personally supply.

Standard and support: make it a plan

Issue statements describe the past. Gates three and four describe the future — what success looks like, by when, and what the company will do to help. Skip the fourth and you have not written a plan. You have written a warning with a countdown on it.

Feed the model the actual role. The job description and your real team metrics are what stop it producing generic goals about improved communication. Four elements make a standard testable:

Role: You are an HR business partner writing measurable success criteria.

Goal: For each issue statement below, write one standard the employee can objectively pass or fail.

Context, the role: [paste the sanitised job description]
Context, how the team is already measured: [paste your real metrics, or write "none defined"]

Every standard must contain all four of these or it is not finished:
1. The specific behaviour or deliverable
2. How it will be measured
3. The frequency or the deadline
4. The review point: day 30, 60, or 90

Do not use the words improve, better, more consistent, or as needed. None of them can be measured.

Format: a table with the columns Issue, Standard, Measure, Review point.

Issue statements:
[paste your output from gate 2]

If your team already runs a defined KPI framework, pull the standards from it rather than inventing parallel ones. A PIP target that contradicts the metric everyone else is judged on is the first thing an employee will point at.

This gate is the one managers skip, and it is the one that decides whether the document reads as a genuine attempt or a paper trail. A plan with standards but no support is a test with a predetermined result.

Role: You are an HR business partner making sure this plan is fair.

Goal: For each standard below, propose the support the company should provide so the employee has a genuine chance of meeting it.

For each one give: the support item, who owns it, and when it starts. Prefer concrete things — a named weekly check-in, a specific course, a reassigned piece of workload, a defined point of contact. Avoid vague offers such as "coaching as required".

Flag any standard where you think no realistic support exists, and explain why.

Format: a table with the columns Standard, Support, Owner, Starts.

Standards:
[paste your output from gate 3]

Stress-test the draft before HR does

Gate five turns the model against your own document. Ask it to find every sentence that could be argued with and it will hand you several you stopped noticing three drafts ago.

Role: You are an employment lawyer reviewing this draft on behalf of the employee. Your job is to find every weakness in it.

Goal: Identify every sentence in the draft below that could be disputed, misread, or argued as already met.

Report in this order:
1. Subjective or vague wording, quoted exactly, each with a specific replacement
2. Standards with no stated baseline, so nobody can tell what changed
3. Standards that could be argued as already satisfied today
4. Deadlines that look unrealistic given the support offered
5. Anything that reads as a decision already taken rather than a plan

Do not rewrite the document. List the problems only.

Draft:
[paste your full draft]

Fix what it finds, then read the result against your own written rules — the wording needs to match the HR policies your organisation has published, not a generic template’s idea of them.

Write the whole thing as though a stranger will read it in a year, because they might. The EEOC’s recordkeeping requirements oblige US employers to keep personnel records for at least a year, and to preserve anything relevant until a charge reaches final disposition. Other countries differ on the detail. None of them differ on the principle.

What AI cannot do here

Every gate above ends with you. That is not a disclaimer bolted on for safety — it is how the method works. Drafting a performance improvement plan using AI shifts the labour, not the accountability.

Two things worth being blunt about. First, if you cannot write three specific examples yourself, the problem is your evidence, not your wording — and no prompt fixes thin evidence. Second, a PIP is not a performance review. Reviews evaluate a period; this is a corrective document with legal weight and a deadline. If what you actually want is the annual cycle, our guide to writing performance reviews for your team is the right page.

And yes — people ask whether the employee will be able to tell. Occasionally, if you paste the output unedited and it arrives sounding like a template. They will notice the tone before the technology. What they will definitely notice is whether the standards are real, which is a content problem, not a phrasing one. The conversation still needs preparing properly.

When the output goes wrong

Four failures account for nearly everything that goes sideways with this workflow.

The tone fix is worth spelling out, because it is the one people get backwards. Do not soften the standards. Keep the standards exactly as they are and rewrite only the framing sentences around them — ask for supportive, plain language while stating that the criteria themselves must not change. Softening the criteria to sound kinder is how a plan becomes unenforceable.

If a pass comes back worse rather than better, diagnose which part of the input caused it instead of rerunning the same prompt hoping for a different afternoon.

Your final check before delivery

Before this leaves your desk, walk it once more. Not to polish the language — to confirm it would still hold up if you had to explain each line to someone who was not there.

Then check the last few details by hand, because these are the ones AI will never flag:

• Every stand-in tag from gate one has been swapped back to the real name.
• The dates are the actual dates, not the ones the model assumed.
• Nothing in the wording implies the outcome is already decided.
• Your HR contact has seen the final version, not an earlier draft.

Save the five prompts somewhere you will find them again. The gates do not change between plans — only the notes you pour into them do, and the second time through this takes about twelve minutes.

The wider habit is worth keeping too. Building a performance improvement plan using AI is really one instance of a pattern that works across every document with consequences attached: separate the sanitising from the thinking, the thinking from the structuring, and the structuring from the critique. One prompt asked to do all four does none of them properly.