Daniel Acheampong leads a six-person data analytics team at a healthcare-services company. For three weeks he had been carrying a knot in his stomach about one of his analysts, Tomas, whose work had slipped. In Daniel's head the problem had a name, and it was not a useful one: "Tomas has an attitude problem." When he tried to imagine actually saying that out loud in a one-on-one, he froze, because he knew it was both unfair and unactionable. A label is not feedback. You cannot coach someone out of a word. So before the conversation, Daniel sat down with an AI assistant, not to write his script, but to help him turn that vague, loaded label into specific, evidence-backed observations he could stand behind. By the end he had three concrete feedback points and a far steadier sense of what he actually needed to say.

What This Lesson Covers

Performance conversations are where management gets real. They affect a person's livelihood, confidence, and career, which is exactly why they are hard to do well and tempting to avoid or rush. This lesson is about using AI to prepare for these conversations, and it draws a hard line around that word. AI helps you organize evidence, structure your feedback, anticipate how the person might react, and sketch a development path. It does not have the conversation. The judgment, the empathy, the reading of the room, and every word you actually speak remain yours.

The core tool you will learn here is the SBI model, a simple structure for feedback that stands for Situation, Behavior, and Impact. SBI is how Daniel converts "attitude problem" into something fair and specific. You will also learn the guardrails that matter most in this domain, because performance data is among the most sensitive information you handle, and AI introduces real risks of bias and dehumanization if used carelessly.

One rule sits above all the others: never feed identifiable employee data into an AI tool your organization has not approved for it. We will return to why that line is non-negotiable.

The Conversations This Applies To

"Performance conversation" covers a wide range, and the preparation shifts with the type. Broadly they fall into three registers. Developmental conversations say "here is how you grow." Constructive ones say "this needs to change." Difficult ones say "this role may not be working." The register determines how much of the conversation is exploration and how much is clarity about consequences.

In practice, seven situations account for most of them. There is the periodic review, whether annual, half-year, or quarterly. There is the mid-year or mid-cycle check-in, which is a conversation rather than a formal assessment and is often the most useful of the lot. There is constructive feedback, where something specific needs to change. There is positive feedback, which managers systematically underprepare for even though recognition that names the specific behavior is what makes it repeatable. There is the promotion conversation, where you are describing the next opportunity and what it demands. There is difficult feedback, where performance is genuinely at risk and the job may be too. And there is the continuation decision, where you and the person work out whether this is still the right role for them. Naming which of the seven you are walking into is the first act of preparation, because it sets the tone you owe the person before you have written a single point.

Prepare, Do Not Script

The difference between preparing and scripting is the difference between this lesson working and backfiring. A script is a set of lines you read. If Daniel walks in reciting AI-generated sentences, Tomas will feel it within seconds. People can tell when they are being managed by a template instead of talked to by a person, and nothing destroys trust in a hard conversation faster.

Preparation is different. It means walking in with your evidence organized, your three or four key points clear, your sense of the likely reactions sharpened, and a genuine plan for how to support the person. Then you have a real conversation, in your own voice, that can go wherever it needs to go. Think of AI as a thinking partner who helps you get your reasoning straight before you speak, not a ghostwriter who speaks for you.

AI can help you figure out what is true and fair to say. It cannot say it for you, and it should not try. The hard part of the conversation is supposed to be hard.

Organizing the Evidence

Strong conversations rest on evidence, not impressions. Before Daniel opened any AI tool, he wrote down what he had actually observed, the raw facts, stripped of Tomas's name and any detail that could identify him. That sanitizing step is the guardrail in action: he described "an analyst on my team" and the behaviors, never the person's identity, before pasting anything into the tool.

Good evidence covers four angles. What the person delivered: the work shipped, the metrics, the deadlines met or missed. How they delivered it: collaboration, communication, reliability. Where they fell short: specific gaps, and crucially whether each gap is a one-off or a pattern. And what opportunity exists: the strengths you can build on. AI is useful here as an organizer. You give it your raw, anonymized observations and ask it to sort them into delivered, process, gaps, and opportunities, and to flag which gaps look like patterns versus isolated incidents. That sorting is structural work the AI does well, and it leaves you with a clear picture instead of a swirl of frustration.

As you organize, run a fairness test on each item, because this is where bias creeps in. Ask yourself: would I say this to their face? If not, it is gossip, not feedback. Is this supported by something observable, or is it just my impression? Have they actually had a fair chance to succeed, with clear expectations? Am I judging this person more harshly than I would judge someone else for the same behavior? And one more that is easy to skip: is this feedback, or am I venting? Both feel urgent in the moment and only one belongs in the room. AI can amplify whatever bias is already in how you framed things, so these questions are yours to ask, not the tool's.

The SBI Model: Situation, Behavior, Impact

SBI is the structure that makes feedback specific and hard to argue with, because it sticks to observable facts. It has three parts:

• Situation. When and where did this happen? You anchor the feedback to a specific moment, not a general "you always." For example, "In last Thursday's client review meeting."
• Behavior. What did the person actually do or say? This is the observable action, described without judgment or interpretation. Not "you were dismissive" but "you interrupted the client twice and said the request was not worth the team's time."
• Impact. What was the effect of that behavior? On the work, the team, the client, the goal. "The client emailed me afterward worried that we were not taking their priorities seriously, and I spent an hour rebuilding that trust."

The power of SBI is that it makes feedback impossible to dismiss as a personality attack. "You have an attitude problem" invites denial and defensiveness, because it is a verdict on who someone is. "On Thursday you interrupted the client twice, and afterward they emailed me worried" is a description of what happened and what it cost. There is nothing to argue with, only something to address. AI is genuinely helpful at restructuring your raw observations into clean SBI form, because separating the behavior from your interpretation of it is exactly the kind of restructuring it is good at.

Two elements finish the structure once the observation and impact are on the table. Name the opportunity: here is how we could address this. And name your part in it: here is what I will do to help. Observation and impact tell someone where they stand. Opportunity and support tell them there is somewhere to go and that they are not going alone. Feedback that stops after impact is a verdict, not a conversation.

Worked Example: From "Attitude Problem" to Three Specific Points

Here is how Daniel used SBI with AI to fix his vague, unfair starting point. He gave the AI his anonymized raw observations and the prompt: "I have a team member whose performance has dropped. I keep thinking of it as an attitude problem, which I know is not fair or specific. Help me restructure these observations into specific feedback points using the SBI model, Situation, Behavior, Impact. Flag anything that is interpretation rather than observable behavior."

His raw, loaded notes going in were: "Bad attitude in meetings. Doesn't care anymore. Sloppy work lately. Checked out." Every one of those is a judgment, not an observation. Working with the AI to map each frustration back to a specific incident, then running each through SBI, Daniel landed on three points he could actually deliver:

• Point 1. Situation: In the last three sprint reviews. Behavior: Tomas's dashboards shipped with calculation errors that QA caught, three in the most recent sprint alone. Impact: The analytics team's numbers were pulled from a leadership deck at the last minute, and two stakeholders now double-check our outputs, which slows everyone down.
• Point 2. Situation: In team standups over the past month. Behavior: Tomas has stopped volunteering updates and answers questions in one or two words. Impact: I cannot tell when he is blocked, so problems surface late, and twice we discovered a stalled task only at the deadline.
• Point 3. Situation: On the patient-cohort project two weeks ago. Behavior: Tomas delivered the analysis four days late without flagging the slip in advance. Impact: The clinical team had to compress their review, and we nearly missed a regulatory checkpoint.

Notice what happened. One vague, accusatory label, "attitude problem," became three specific, behavior-based, evidence-backed points. None of them mentions attitude. Each names a moment, an observable action, and a concrete consequence. And here is the part the AI could not do: as Daniel reviewed the three points, the late delivery and the sudden quietness made him wonder whether something was going on outside work. The AI saw only the behaviors he typed; Daniel, knowing Tomas, recognized a pattern that looked less like not caring and more like someone struggling. That insight reshaped how he planned to open the conversation, leading with curiosity rather than accusation.

Running the Difficult Version Well

Tomas's case had crossed into serious territory, and Daniel had to prepare for that without either flinching or overreacting. The tone he aimed for was serious, direct, and fair: not angry, not soft. Anger makes the conversation about him. Softness leaves the person genuinely unsure whether anything is wrong, which is its own unkindness.

The shape he planned had six moves. State the pattern plainly, and say it is a pattern rather than an incident: six to eight weeks of missed deadlines, dipped quality, and changed engagement is a trend, and naming the timespan is what separates a trend from a bad week. Ask what is going on, and mean it, because the context may change everything. Be direct about the impact, on the team and on the person's own standing. Be clear about what has to change, in observable terms rather than adjectives. Offer real support alongside real clarity about the stakes. And end with a commitment and a next step, including a specific date to review whether things have moved: if we are heading the right way, good, and if not, we will need to talk about what that means for the role.

Two branches are worth planning in advance. If there is context, whether a personal situation, unclear expectations, or an unreasonable workload, the conversation turns toward what changes on both sides, and the standard stays the same while the path to it adjusts. If there is no clear context, the conversation stays where it started: this is the pattern, this is what has to change, this is when we will look at it again.

Afterward, send a short written recap. Not a legal document and not a threat, just a note that says thank you for the conversation, here is what we discussed, here is what we each committed to, and here is when we will check in. It protects both of you from remembering the conversation differently, and it is the single easiest thing to skip and later regret.

The Quieter Conversation: Solid, But Not Growing

Not every performance conversation involves a problem, and the hardest ones sometimes involve none at all. Daniel had a second analyst, Sarah, two years in the role, who executed reliably, met her deadlines, worked well with everyone, and had learned almost nothing new since she arrived. There was nothing to criticize. There was also a real conversation to have, and the trap here is doing neither: not criticizing, but also not saying "you are doing fine" in a way that quietly closes the door on growth.

Daniel prepared it as six moves. Open by appreciating what she genuinely does well, because it is true and it sets the frame. Share the observation about growth, in specifics: she is excellent at executing when the requirements are clear, and he would like to see her spot problems before he does, propose solutions, and drive work more independently. Then open the dialogue and actually stop talking, because the whole conversation rests on one question he cannot answer for her: does she want to grow, or is she content at this level? Then, depending on that answer, describe what growth would concretely look like, name the support he would provide and what he would need from her, and close with clear next steps.

The three branches matter. If she wants to grow, the conversation becomes specific about the path: leading a project, owning a domain, mentoring someone. If she is content where she is, that is a legitimate answer and the conversation turns to what makes the work fulfilling for her, not to persuasion. And if she is unsure, the useful move is to explore what growth would even mean to her before assuming it means promotion. Preparing all three branches is what stopped Daniel from walking in with a conclusion already reached.

Assume Growth, and Own Your Half

Underneath every one of these conversations sits an assumption about whether people can change, and the person across the table will detect yours whether you state it or not. Assume they can improve. Be explicit about what improvement looks like, provide the support and resources to make it possible, allow genuine time for change to show up, and then follow up and name the progress when it appears. People tend to rise or sink to the expectations set for them, and a manager who has privately decided someone is a lost cause will run the conversation in a way that proves it.

That said, assuming growth is not the same as being vague. Performance management holds together on four things at once. Clarity: is the person genuinely clear on what is expected, or have they been guessing? Support: are we actually providing the resources and help to meet it? Honesty: are we being straight about the gaps, or softening them into meaninglessness? And consistency: are we holding people to comparable standards relative to one another, or does the bar move depending on who is standing at it? Any of the four missing turns the conversation into theater.

Anticipating Reactions and Planning Support

Once your points are clear, AI can help you pressure-test them by anticipating how the conversation might go. Ask it: "Given these three points, what are the likely ways the person might react, defensive, surprised, upset, in agreement, and how might I respond to each while staying fair and direct?" This is rehearsal for your own thinking, not a script. It helps you walk in less likely to be thrown off balance.

Equally important is the development path. Feedback without support is just criticism, and people rarely improve from criticism alone. For each point, plan what help you will offer. For Daniel's first point, that meant proposing a peer-review step on Tomas's dashboards before they ship and offering to pair him with a senior analyst for two weeks. The GROW coaching structure, Goal, Reality, Options, Will, is a useful frame here: clarify the goal you both want, agree on the current reality, explore options together, and end with what the person will commit to do. AI can help you draft questions for each stage, but the commitment has to come from the person, in the room, not from your plan.

How This Goes Wrong

Using AI to avoid the difficult part. The most seductive misuse is drafting something so polished that it lets you deliver the words without ever having the conversation. AI organizes your thinking. You still have to sit across from a person and say a hard thing and stay in the room while they react. That discomfort is the job, not a defect in your preparation.

Landing too harsh or too soft. AI will happily produce something that reads reasonably on screen and is, in reality, either brutal or so hedged it says nothing. The test is mechanical: read it aloud. Would you actually say it that way, to this person, in this room? Does it sound like you? If either answer is no, rewrite it.

Reaching into someone's personal life. Gathering or citing information about behavior outside work, unless it bears directly and legitimately on the job, is a line not to cross. Stick to work performance. Their personal life is theirs, and inviting AI to reason about it compounds the problem.

Collecting one-sided evidence. It is easy to assemble only the examples that confirm the conclusion you already reached, especially when you are frustrated. Deliberately look in the other direction: what evidence would argue against your view? Who saw this differently? A case that survives that test is one you can defend to the person's face.

Feedback with no support attached. Telling someone they are falling short and offering nothing to help them stop is not accountability, it is abandonment with paperwork. Every point you raise should arrive with a sentence that starts "here is what I will do to help."

Your Judgment Checkpoints

Before you walk in, run six quick checks on your own preparation. The fairness check: is this fair, would I say it to their face, would I want to hear it said this way about me? The evidence check: is each point supported by something observable, or am I venting in a professional accent? The context check: do I actually understand this person's full situation, and should I be asking rather than concluding? The support check: am I offering genuine help, or only criticism with a friendly opening line? The tone check: am I direct and human, or have I drifted into corporate and cold? And the sensitivity check: what is this person likely to be feeling when they hear this, and how will I acknowledge it? Each of these is a place where you override or adapt whatever the preparation produced, including anything AI helped you draft.

The Guardrails That Matter Most

Never feed identifiable employee data into an unapproved tool. Performance information is sensitive and often legally protected. Names, employee IDs, health details, anything that identifies a real person, stays out of any public or unsanctioned AI tool. Anonymize ruthlessly: describe behaviors and situations, not identities. If your organization provides a private, approved tool with the right data protections, use that and only that for anything identifiable.

The human owns the judgment. AI can draft a structure, but it cannot decide whether a gap is fair to raise, how sensitive a situation really is, or whether someone has had a genuine chance to succeed. Those are judgment calls, and they are yours.

Protect authenticity. AI-drafted talking points drift toward corporate language, and corporate language in a conversation about someone's career reads as distance. Rewrite everything in your own words. If it does not sound like you, they will feel it, and they will conclude that the words were produced rather than meant.

Hold on to human dignity. A tool can optimize phrasing and remain entirely blind to what is actually happening in the room. You are talking to a person whose livelihood, confidence, and sense of themselves are in play. Nothing about the efficiency of your preparation should make you forget that.

Watch for bias. AI learns from patterns and can quietly reinforce the bias already in how you framed something. Test your own assumptions. Are you describing observable behavior or projecting a story? Would you frame the same action the same way for a different person on your team? Look for evidence that contradicts your view, not just evidence that confirms it.

Read it aloud in your own voice. Before any conversation, say your key points out loud. If they sound corporate, robotic, or unlike you, rewrite them. The person across the table will feel the difference between your words and a machine's, and authenticity is what makes hard feedback land as respect rather than process.

Practice and Reflection

Take two minutes on this before you move on. Think about the past week and find one conversation, one piece of feedback, or one decision about a person where these ideas would have changed your approach. What would you have done differently, and what would the outcome have been? Then pick one real situation on your team right now, this week rather than next quarter, and prepare it properly: write the raw observations, strip the identifying detail, sort them into what was delivered, how, where the gaps are, and what the opportunity is, run each point through SBI, and decide what support you will offer alongside it. The connection between the concept and an actual conversation on your calendar is where this becomes a skill rather than a framework.

Related Lessons

Difficult Conversations Preparation applies a similar preparation framework to the wider set of hard conversations you have as a manager, not only performance ones. If the difficult version of this lesson is the part you dread most, that lesson goes deeper on holding your nerve in the room.

Coaching and Development Planning picks up where a developmental performance conversation ends. Once someone knows honestly where they stand, the next question is what growth looks like and how it gets built, which is the whole subject of that lesson.

Feedback Crafting goes further into the specific techniques of shaping feedback so it lands. SBI is the backbone; that lesson adds the range of moves for different people, different severities, and different moments.

Key Takeaways

• Prepare, never script. AI organizes your evidence and structures your thinking; you have the actual conversation in your own voice. The moment it becomes a recited script, the person feels it and trust collapses.
• Use SBI to convert labels into feedback. Situation, Behavior, Impact turns a vague "attitude problem" into specific, observable, evidence-backed points that cannot be dismissed as a personality attack.
• Finish with opportunity and support. Observation and impact tell someone where they stand; naming how it could be addressed and what you will do to help turns a verdict into a conversation.
• Ground every point in observable evidence. Separate what you saw from what you concluded. AI is good at flagging where you have slipped from behavior into interpretation.
• Never feed identifiable employee data into an unapproved tool. Anonymize the situation and behaviors, keep names and sensitive details out, and use only an organization-approved tool for anything identifiable.
• Run the fairness test yourself. Would I say this to their face? Is it evidence or impression? Have they had a real chance to succeed? Am I judging them as I would judge anyone else? Is this feedback or venting? AI cannot ask these for you.
• Be clear, and follow up in writing. People should leave knowing exactly where they stand and when you will look at it again. A short recap note after a serious conversation prevents two different memories of it.
• Pair every piece of feedback with support. Plan the concrete help you will offer, peer review, pairing, coaching, before you walk in. A frame like GROW keeps the conversation developmental, but the commitment must come from the person.
• Assume people can grow, and hold the four pillars. Clarity, support, honesty, and consistency together are what make performance management fair; people rise or sink to the expectations you set.
• Stay alert to what AI cannot see. It only knows what you typed. The pattern that suggests a struggling person rather than a careless one is something only you, who know them, can recognize.
• You are talking to a human whose livelihood may be at stake. Read your words aloud, keep them in your voice, and let empathy lead. That is the part no tool can do for you.