AI in employee conversations: support, limits and human responsibility

Before an employee conversation, you have a handful of observations, outstanding commitments and development goals. AI can turn these into a draft agenda, summarise the discussion and suggest next steps. That can help with organising information and putting it into words. Whether it leads to a better conversation also depends on what information is missing, how suggestions are checked and whether employees feel able to speak openly.

A well-written AI output is a draft to work with. It does not establish that an observation is complete, an assessment is accurate or a suggested action will work. For HR and managers, the specific task therefore matters: will the tool organise information, support a shared discussion or influence an employment decision?

This article follows four stages: preparation, the conversation, follow-up and ongoing development. Our guide to a privacy- and works-council-friendly rollout covers the organisational framework.

What technical capabilities tell us about the benefits

Research on automated meeting summarisation shows that systems can condense conversations. Shang et al. (2018) studied this using the AMI and ICSI meeting datasets; Li et al. (2019) combined text and video information from AMI project meetings. Improvements in automated text comparison measures such as ROUGE or BLEU indicate progress on those tasks. They measure neither the quality of an employee conversation nor a person's development.

The word error rates of 36 and 37 per cent in Shang et al. refer to the transcripts used in those two datasets at the time. They are not general error rates for today's transcription systems, nor percentages of incorrect employment decisions (Shang et al., 2018).

Experiments on performance feedback are closer to the HR context. Marót et al. (2026) studied reactions to identical feedback text with different descriptions of how it had been produced. This is research on a hypothetical appraisal scenario, not a test of actual development conversations. The studies reviewed here establish neither general time savings after full checking nor long-term development benefits from using AI throughout the conversation process. The workflows below are practical recommendations, not an empirically validated method as a whole.

1. Preparation: organise topics and check the starting point

Agreed goals, outstanding actions, selected observations and both participants' concerns can provide the material for a draft agenda. A useful instruction might be: “Organise this information into discussion topics. Separate observations, open questions and previously agreed actions. Do not add claims about the person.” Use only the intended data and a tool approved for that purpose.

Then check whether the agenda reflects the current situation. Are older entries still relevant? Do several notes describe the same event? Are less visible contributions, counterexamples or the employee's own topics missing? AI can only work with the context provided. How often a topic appears in the notes is not a reliable measure of its importance.

Fictional example: an observation becomes an assumed weakness

A manager notes: “Nora answered the customer's question after the team meeting.” The AI suggests “Improve communication skills” as an agenda item. During the conversation, it turns out that Nora had been waiting for technical approval.

Bringing the unresolved situation back into view was useful. Assigning it to a personal weakness was not justified. A suitable discussion topic would be: “What delayed the answer, and what coordination do we need?” Our article on recording workplace observations explains how to separate observation from interpretation.

Form a provisional view of your own before reading AI suggestions. In a randomised scenario experiment with preservice teachers, this sequence helped participants avoid adopting misleading recommendations. The study involved educational tasks at a Chinese university, not management conversations; it does not establish guaranteed protection in HR (Yan, 2026). In practice, write down your own questions, compare them with the AI draft and deliberately look for another explanation. Your own assessment must remain open to correction too.

2. During the conversation: support documentation and make room for openness

Transcription converts speech into text. A summary selects information from that text and rewrites it. These are separate processing steps. Even when a transcript reads smoothly, check names, numbers, negatives and who said what. Test the specific tool with the relevant languages, specialist vocabulary, speaker changes and actual recording conditions.

One potential benefit during the conversation is having less to write down while listening. Whether this actually frees up attention needs to be assessed in practice. Live suggestions can themselves distract participants or shape which topics seem important. Keep the prepared agenda open to concerns that only emerge through discussion.

Recording also changes the setting. Interviews with 50 employees from the US, China and Germany identified both opportunities and concerns relating to control, learning and psychological safety (Cardon et al., 2023). This qualitative study of meetings does not measure a causal effect of AI recording on openness in employee conversations.

Explain in advance what will be recorded or analysed, how the output will be used and who can see it. Establish separately whether recording is lawful. Agree on a process for handling sensitive passages and for holding conversations without recording when appropriate. Afterwards, also ask whether documentation caused anything to go unsaid. Agreement to use the technology alone does not answer that question.

3. Follow-up: check meaning, attribution and commitments

An AI summary can present topics, open questions and next steps clearly. Information can also be lost or changed. A study of 200 generated meeting summaries distinguishes, among other errors, complete and partial omissions, incorrect references, and content that contradicts the transcript or does not appear in it (Kirstein et al., 2025). These findings concern a research dataset, not the error rate of every product.

Fictional example: a condition becomes a commitment

During the conversation, Elias says: “I can run the workshop in November if we arrange someone to cover support first.” His manager replies: “I'll sort out the cover.” The AI draft reads: “Elias will organise the November workshop and support cover.”

The draft highlights a next step but drops the condition and assigns the cover arrangement to the wrong person. An accurate version would be: “Manager to clarify support cover by the next one-on-one. Elias will run the workshop provided cover is secured.” Checking the draft preserves the actual agreement.

Check the draft in both directions: is every statement supported, and is anything essential missing? Look beyond wording and grammar. Pay particular attention to:

• Meaning: Are reservations, uncertainty and differing views preserved? “Could” must not quietly become “will”.
• Attribution: Who observed, proposed or agreed to something? A wish is not yet a commitment.
• Completeness: Does the summary include conditions, support needs and the manager's responsibilities?
• Additions: Has the AI inserted reasons, assessments or deadlines that were never discussed?

Check consequential statements against the source information lawfully available and clarify uncertainties with the participants. A transcript can contain errors too. If the source is no longer available, mark the uncertainty instead of presenting a reconstruction as established fact. Traceability does not justify keeping recordings indefinitely.

Assign someone to review the draft. Discuss the relevant commitments with the employee, allow corrections and record any remaining disagreement. Make it clear which version was checked and who is responsible for its further use. Errors already carried into later documents need correcting there as well. An approval click without time, context and the ability to make changes does not provide this review.

4. Ongoing development: offer options and assess progress together

AI can group similar observations and suggest connections to goals or competencies. Similar wording does not establish a stable behavioural pattern or explain its cause. Check the situation, role, expectations, counterexamples and the person's perspective before turning it into a development topic.

Once the topic is clear, AI can help assemble possible actions: a practice assignment, peer feedback, shadowing, coaching or training. For each suggestion, check the specific learning goal it should support, whether the offer actually exists and whether time, access and support are available. A relevant course title is not evidence that the course will help in everyday work.

Fictional example: training or a missing decision?

Several notes about Amir's project mention late handovers. The AI suggests time management training. A joint review shows that Amir had finished the documents, but approval kept moving between two people.

Grouping the notes helps identify a recurring obstacle. The training suggestion, however, assumes a cause that has not been checked. The manager first clarifies who owns approval. Amir also wants to practise raising outstanding decisions earlier. They agree to try this over the next few handovers and then discuss specific examples. Neither a course booking nor an AI score counts as evidence of development.

For an agreed action, establish who will try what, what support is needed and how you will recognise progress together. Revisit it later using new observations. Repeated automated summaries must not turn an outdated assumption into an apparently established personal trait.

Applications and the checks they need

This table is a practical review aid. Each potential benefit needs to be assessed in your own setting.

Trust also depends on the role AI plays

There is a difference between asking AI to improve the wording of a manager's own assessment and asking it to produce the assessment itself. In the experiment with 192 employees mentioned above, feedback described as entirely AI-generated prompted less favourable reactions than feedback described as human-written. Text described as human-written and then refined by AI did not differ statistically significantly from purely human feedback (Marót et al., 2026). That proves neither equivalence nor a lasting effect on workplace trust.

The practical implication is to explain the actual use clearly and make personal responsibility visible. Which observations came from the manager? What was merely reworded? Which interpretations were checked together? Transparency is necessary for informed use, but it does not guarantee trust. Employees also need to be able to disagree and see their concerns reflected in the documentation.

Legal context: purpose and use remain decisive

As of 7 October 2026. This overview does not replace an assessment of the specific use case.

In the EU, the GDPR requires a lawful basis, purpose limitation, data minimisation and accuracy, among other safeguards. Health data is subject to additional conditions; a data protection impact assessment is required where processing is likely to result in a high risk. Article 22 addresses solely automated decisions with legal or similarly significant effects. Other data protection duties also apply outside that provision. Human approval does not make unlawful data collection or reuse lawful (European Union, 2016, Arts. 5, 6, 9, 22 and 35).

The EU AI Act is already in force. The general prohibition on biometric emotion recognition in the workplace has applied since 2 February 2025, with exceptions for medical or safety reasons. AI used to evaluate performance or make certain employment decisions may fall within the high-risk rules. Purpose and actual influence require assessment; describing a system as “assistance” or leaving the final decision to a person does not create a blanket exemption (European Union, 2024, Arts. 3, 5, 6 and Annex III). Following the AI Omnibus, which entered into force on 27 July 2026, the high-risk rules for Annex III systems apply from 2 December 2027. This transition does not suspend existing prohibitions or data protection duties (European Commission, 2026; European Commission, n.d.).

In Germany, section 26 of the BDSG and, where applicable, works council co-determination under section 87(1)(6) of the BetrVG must also be considered. Employment-related dependency matters when assessing consent (Bundesrepublik Deutschland, 2017; Bundesrepublik Deutschland, 1972). In Austria, the GDPR likewise applies directly; sections 96 and 96a of the ArbVG may additionally require works council consent. The conditions differ according to the type of system (Republik Österreich, 1974).

Switzerland has its own framework under the DSG and employment law. For private-law employment relationships, Article 328b of the Code of Obligations limits processing to information concerning suitability for employment or necessary to perform the employment contract. EU rules do not apply simply because there is a Swiss employment relationship; any additional EU scope must be assessed separately (EDÖB, n.d.; European Union, 2016, Art. 3; European Union, 2024, Art. 2).

Questions for selecting and using a solution

Ask for answers demonstrated in the actual workflow and supported by current documentation:

1. Task and boundaries: Which steps does AI perform, which outputs remain drafts and which feed into assessments or decisions?
2. Source information: Which sources and time periods are used? Can irrelevant, duplicate or outdated information be excluded?
3. Quality: Has the system been tested with comparable languages and conversations? What was measured: text similarity, factual accuracy, review effort or the participants' experience?
4. Traceability: Can a reviewer see what supports a statement and where information is missing? Do suggestions remain distinguishable from actual commitments?
5. Correction: Who can change or reject drafts? How are disputed statements marked and corrections carried into later documents?
6. Responsibility: Who checks which content, with what expertise and how much time? Who decides what to do when the evidence is insufficient?
7. Conversation setting: How are employees informed? How can they raise concerns, and what process is available for sensitive conversations?
8. Data processing: Who can access recordings, transcripts, notes and derived information? Which service providers, retention periods, processing locations, third-country access arrangements and uses for model training are documented?
9. Legal assessment: Have recording, processing purposes, employee representation rights and possible high-risk classification been assessed for the specific use case?
10. Benefits in practice: Is total time measured, including checking and correction? How are missed errors, openness and follow-through on agreed actions assessed?

Start with a bounded use case. Compare the workflow with your existing practice and consider the severity of errors as well as the effort involved. A short draft containing an incorrect commitment can create more work to resolve than it saves in writing. Reassess the workflow when the model, data sources or intended use change.

How riteful supports this work

riteful supports AI-assisted agenda creation, meeting summaries, observation analysis and one-on-one plan creation. Journal entries, goals, outstanding actions and previous conversations can provide context when preparing an agenda. Observation analysis offers AI-suggested links to competencies and areas of responsibility. These connections need checking in the relevant work context.

Managers and employees jointly define a development and conversation plan. Notes are private; HR can see whether one-on-ones take place but cannot access conversation content or private notes. AI outputs assist people; riteful does not make autonomous employment decisions. The employee review software page explains the workflow.

References

Bundesrepublik Deutschland (1972). Betriebsverfassungsgesetz (BetrVG), § 87. Current version, accessed 7 October 2026.

Bundesrepublik Deutschland (2017). Bundesdatenschutzgesetz (BDSG), § 26. Current version, accessed 7 October 2026.

Cardon, P. W., Ma, H., & Fleischmann, C. (2023). Recorded Business Meetings and AI Algorithmic Tools: Negotiating Privacy Concerns, Psychological Safety, and Control. International Journal of Business Communication, 60(4), 1095–1122.

Eidgenössischer Datenschutz- und Öffentlichkeitsbeauftragter (EDÖB) (n.d.). Datenbearbeitung durch den Arbeitgeber. Accessed 7 October 2026.

European Commission (2026, 27 July). AI Omnibus enters into force. Accessed 7 October 2026.

European Commission (n.d.). Timeline for the Implementation of the EU AI Act. AI Act Service Desk. Accessed 7 October 2026.

European Union (2016). Verordnung (EU) 2016/679 (Datenschutz-Grundverordnung). Arts. 3, 5, 6, 9, 22 and 35. Accessed 7 October 2026.

European Union (2024). Verordnung (EU) 2024/1689 (Verordnung über künstliche Intelligenz). Arts. 2, 3, 5, 6 and Annex III; text checked as amended on 27 July 2026. Accessed 7 October 2026.

Kirstein, F., Ruas, T., & Gipp, B. (2025). What’s Wrong? Refining Meeting Summaries with LLM Feedback. Proceedings of the 31st International Conference on Computational Linguistics, 2100–2120.

Li, M., Zhang, L., Ji, H., & Radke, R. J. (2019). Keep Meeting Summaries on Topic: Abstractive Multi-Modal Meeting Summarization. Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, 2190–2196.

Marót, J. L., Palcsó, T., & Szabó, Z. P. (2026). Source matters, AI anxiety less so: comparing employee reactions to human, AI, and hybrid performance feedback and the limited role of AI anxiety. Frontiers in Psychology, 17, 1924487.

Republik Österreich (1974). Arbeitsverfassungsgesetz (ArbVG), §§ 96 and 96a. Current version, accessed 7 October 2026.

Shang, G., Ding, W., Zhang, Z., Tixier, A., Meladianos, P., Vazirgiannis, M., & Lorré, J.-P. (2018). Unsupervised Abstractive Meeting Summarization with Multi-Sentence Compression and Budgeted Submodular Maximization. Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 664–674.

Yan, D. (2026). Independent judgment before viewing generative AI recommendations reduces susceptibility to misleading advice: a randomized scenario-based experiment with preservice teachers. Frontiers in Psychology, 17, 1936522.