Prepare Feedback Conversations With AI: A 1:1 Script and Limits

Prepare Feedback Conversations With AI: A 1:1 Script and Limits

To prepare feedback conversations with AI, use the tool to organize what you have already observed: facts, concrete examples, impact and clear requests, in a short script for the 1:1. AI helps you structure the material and review your wording. Judging the person, setting the final tone and handling any sensitive data stay with the manager.

Key takeaways

• AI is for structuring your preparation (facts, examples, impact, request). It is not for evaluating the person or deciding consequences.
• The more specific the material you provide, the more useful the script. Vague impressions produce vague feedback.
• Sensitive data, such as health, family or disciplinary proceedings, stays out of the tool unless there is a clear policy and legal basis.
• The manager reads, cuts and rewrites the script before the conversation. The AI text is a draft, never a script read aloud.
• Test the process on a few conversations, with a simple usefulness criterion, before turning it into a team habit.

What AI does well in preparation, and what it should not do

A feedback conversation usually fails from lack of preparation, not lack of goodwill. The manager arrives with a general impression ("she seems a bit checked out"), no dated examples, no clarity about what should change and no idea what to propose. AI is good at turning loose notes into a readable structure, and that solves a large part of the problem.

What it does well in this context:

• Organizing notes into blocks: what happened, what the impact was, what is expected going forward.
• Separating fact from interpretation. When you write "he is careless", it can point out that this is a judgment and ask for the observable behavior behind it.
• Suggesting open questions so you can hear the other side.
• Proposing clearer, less accusatory versions of a sentence.
• Anticipating likely objections so you can think through responses in advance.

What it should not do:

• Decide whether the person deserves a promotion, a warning or a dismissal.
• Infer intent, character or personal circumstances from messages and metrics.
• Replace your knowledge of the context: what the team went through, what was agreed before, what the person has already tried to fix.
• Generate examples you did not witness. A language model can invent plausible details, and a false example in a feedback conversation destroys trust.

For a broader view of this boundary, read what not to delegate to AI in business. Feedback about people sits very close to that line.

What information can go into the tool?

Before pasting anything, ask: if this text leaked or were read by someone else, would it cause harm? Feedback is HR material by nature, and Brazil's LGPD applies to employees' personal data, not only customers'.

A practical criterion, as a decision guide:

• Can go in, with care: descriptions of work behavior, deliverables, deadlines, excerpts from work documents with no third-party information, and the goal of the conversation. If possible, replace the name with a label such as "employee A".
• Goes in only with a company policy and a decision from the responsible person: consolidated performance history, individual targets, the content of previous formal reviews.
• Stays out: health, medical leave, family life, religion, sexual orientation, union membership, ongoing disciplinary proceedings, complaints and anything the person confided in you in confidence.

Where the tool processes the data also matters. Cloud services may send information outside the country, which has implications described in international data transfer under LGPD when using AI providers. If your company has a usage policy, it should say which tools are approved for this kind of text. If it does not have one yet, a company policy for employee ChatGPT use is a good starting point. For the ANPD, the authority that guides LGPD enforcement in Brazil, the starting point is always purpose, necessity and minimization: use only what the task requires.

One principle also helps quality: giving the AI only the relevant context usually produces better answers than dumping everything in. Anthropic's discussion of context engineering addresses exactly this, choosing with care what goes into the model's window.

Step by step: building the 1:1 script

This is a simple flow that fits in twenty or thirty minutes of preparation.

• Step one, take unfiltered notes. Before opening the AI, write down for yourself what you observed: dates, deliverables, situations, what was said. Use only what you witnessed or can prove.
• Step two, remove identification and sensitive data. Replace names with labels and remove anything from the "stays out" list.
• Step three, ask for structure, not a verdict. Ask the AI to organize your notes into four blocks: observed situation, impact, what you expect and how you can support.
• Step four, ask for a language check. Ask it to flag sentences that sound like a judgment of the person ("you are disorganized") and suggest versions about behavior ("in Thursday's delivery, three items went unreviewed").
• Step five, ask for questions that invite listening. Two or three open questions, such as "what made that deadline hard to meet?".
• Step six, validate every fact. Reread the script and confirm that every example is real and verifiable. Cut whatever the AI added on its own.
• Step seven, define the request. Write in one sentence what you want to change, by when and how you will follow up. This step is yours.
• Step eight, rehearse the opening. Write in your own words the first thirty seconds of the conversation. That is the part that sets the tone most.

A prompt that tends to work well:

Organize these notes into observed situation, impact, expectation and support; flag any sentence that is a judgment rather than a fact; do not add examples I have not written.

The last clause matters. Telling the AI explicitly not to invent examples reduces the risk but does not eliminate it. That is why step six exists.

Illustrative scenario: feedback about deadlines on a support team

Illustrative scenario: an operations manager needs to talk with an analyst whose weekly deliverables keep slipping. She writes her notes: three weeks in a row, the report agreed for Friday arrived on Monday, and the most recent delay made the sales team redo a spreadsheet. She also remembers that the analyst took on an extra project in the same period, but she knows he mentioned a family situation in confidence, and she decides not to include it.

She replaces the name with "analyst A", pastes the notes and asks for the four-block structure. The AI returns a draft with the situation (three deliveries on Monday instead of Friday), the impact (rework for the sales team), the expectation (delivery on Friday or an early warning) and the support (reviewing the workload). It also flags that her note "he is unmotivated" is an interpretation, and suggests replacing it with a question: "how have you been seeing your workload over the last few weeks?".

The manager cuts a sentence that sounded harsh, confirms the dates against the delivery history and defines the request: a heads-up by Wednesday if a deadline is at risk. The decision about workload, tone and any consequence stays with her, and the family situation never entered the tool.

Checklist before you walk into the conversation

Use this as a final review, with the script on screen and the AI closed:

• Is every example something I saw, read or can show?
• Did I separate fact from interpretation, and did the interpretation become a question?
• Is the impact described in terms of the work (deadline, rework, customer) and not personality?
• Is the request specific, with a deadline and a way to follow up?
• Is there room in the script for the person to respond and for me to change my mind?
• Did no sensitive or confidential data enter the tool or the text?
• If the person asks to see my notes, would I be comfortable showing them?
• Does everyone who should be informed (HR, senior leadership) already know, when the situation requires it?

If any answer is no, adjust before the meeting. This checklist also protects against a common mistake: using AI to gain speed and skipping the reflection that feedback requires.

When feedback preparation starts to become surveillance

There is a trap right next to good use. When preparation starts to depend on continuously collected metrics, monitored messages or automatic summaries of each person's behavior, the AI stops helping the conversation and starts producing a dossier. That affects trust, and it can create transparency and purpose problems under LGPD. The boundary between following the work and watching the person is discussed in how to measure AI productivity without surveillance on your team.

It is also worth remembering that decisions that significantly affect a person, such as promotion or dismissal, should not come out of an automated analysis. NIST, in its AI risk management framework, addresses keeping people accountable and able to intervene in systems that affect other people. In the manager's day to day, this translates simply: AI prepares, the human decides and answers for the decision.

Applying the map, limit, pilot and transfer method

For a team that wants to adopt this use in an organized way, the method has a light version.

• Map: describe how managers prepare feedback today, where they lose time and where conversations tend to fail (missing examples, tone, vague requests).
• Limit: define which data can go in, in which approved tool, and what is prohibited. If the company needs to keep everything in its own environment, consider a local model, as in running Ollama local models in a company, knowing that this requires operations work and is not automatic.
• Pilot: choose a few managers and some real conversations, with a shared prompt template and a simple success criterion, for example whether the manager felt clearer going into the 1:1.
• Transfer: document the prompt template, the checklist and the policy, and leave the decision to expand, adjust or stop with the team.

There is no promise of results in this. The pilot exists precisely to find out whether the use helps in that context.

Frequently asked questions

Can I paste the person's full performance review into the AI?

In general it is not recommended. Formal reviews concentrate personal data and sometimes sensitive data. If the company decides to use it, it should be with an approved tool, a clear purpose and data reduced to what is necessary, and HR or the person responsible for data protection should take part in the decision.

Can the AI write the feedback so I just read it in the meeting?

It is not a good idea. A text read aloud sounds distant, and the person notices. Use the AI to organize facts and requests, and lead the conversation in your own words, with room to listen.

How do I stop the AI from inventing examples?

Instruct it explicitly to use only what you wrote, then check each fact against real records, such as deliverables, emails and dates. If an example cannot be verified, remove it from the script. This care matters more than any prompt adjustment.

Do I need to tell the person I used AI to prepare?

There is no single rule, and it depends on company policy and the type of data used. If their personal data was processed in the tool, transparency is an LGPD topic and should be covered in internal guidelines. In any case, being honest when asked helps maintain trust.

Sources

• ANPD: Autoridade Nacional de Proteção de Dados
• NIST AI Risk Management Framework
• Anthropic: Effective context engineering for AI agents