AI coaching uses software to help you think through goals and choices. Tools differ. Some follow set questions. Others write new replies. Some help a human coach with their work. You might use a tool to think through a choice, rehearse a talk, or draft a project update. It can also add noise or expose private details. It may take the place of help you need. A weak first attempt can waste time, leak private data, and add errors. Choose the task first. Check the evidence and what data leaves your hands. Then decide how you will judge the result.
What can AI coaching help you decide?
It can help you prepare, not decide. Try a chatbot on a low-risk task. Draft a message, name what you want from a meeting, or practice how you will open an update. Before a second pass, weigh the time, privacy risks, and errors from the first.
Think about the stakes. If a wrong answer would affect someone's job, health, money, or rights, keep the chatbot out of the decision. That includes hiring, discipline, health, law, and money decisions. Use people with the right skills and authority.
Set limits on how you use the tool. It does not replace your judgment. You stay the decision-maker and you can stop at any point. A clear task, stated in advance, matters more than the label on the tool.
What do comparison studies actually show?
The evidence is mixed, and it does not crown a winner. Three studies test different things, in different settings, with different limits. Read them side by side and no single support type wins everywhere.
| Study design | Population and size | What it found | Main limit |
|---|---|---|---|
| Two separate UK undergraduate randomized trials (Terblanche et al., 2022) | First assigned 210 people (105 per arm); second assigned 268 (134 per arm); 327 complete cases across 8 timepoints | Both the scripted goal chatbot and the human interventions raised self-rated goal attainment over their own comparison controls. Final scores were similar | Two trials, not a direct randomization of people between human and AI. Narrow chatbot, self-rated |
| Non-preregistered three-arm randomized comparison of human coaching, a generative chatbot, and a waiting list (de Haan et al., 2026) | 114 coachees, mainly leaders in a global education company. 99 completed the intended intervention | Reported substantial benefits across outcomes only for human coaching. The 99 reflect adherence, not questionnaire attrition | Short, restricted setting and chatbot. Acknowledgements disclose commercial interests in the chatbot and a measure |
| Preregistered randomized design, preprint, not peer reviewed (Schimpf et al., 2026) | 517 employed adults assigned; 323 passed attention checks and returned at two weeks | AI improved reported goal progress versus no support, not significantly versus structured written reflection | Preprint, attrition, two-week follow-up, one model setup. Nonsignificance is not equivalence; mediation is not causation |
Read these results with care. The studies test different tools and compare different kinds of support. They do not show that structure alone caused the results. They do not prove all tools are equal.
Which tasks fit a limited trial?
Pick a task that is small, frequent, and easy to observe. Choose work where you can check the output soon. You could draft a one-page update or an agenda, or practice how to open a meeting. Keep private work out of a tool that is not cleared to receive it.
Run one limited trial, and keep a record. Use it to guide your choice. It is not a controlled experiment. Keep the record short enough that you will actually complete it.
Here is one chosen fictional example. It shows a plan, not a proven method or a promised result.
- Task: rehearse a three-minute spoken update on a delayed project.
- Fully fictional inputs: a project called "Orion" that is two days late, and a fictional team. No real client, budget, or headcount.
- Activity: ask the tool to draft a message telling the team that Orion is two days late and asking them to confirm a revised review date.
- Prompt: "Using only these fictional details, draft a message under 30 seconds that tells the team Orion is two days late and asks them to confirm a revised review date." No outcomes are invented, and the draft must add no fact or promise.
- Exact criterion: the draft runs under 30 seconds, the ask is clear, and it matches the intended request.
- Record the date and tool. Note its version if shown, or mark it unknown. Before use, read the draft aloud and check it against the supplied notes.
- Note cost and correction time, and compare a written alternative.
- Check privacy before typing. Approval must cover both the tool and the input. Stop on unsafe output. Do not use wrong facts. Fix only a clear, low-risk gap within the limit. Keep the tool only if the checks pass and its use is approved.
Keep it to a fictional 15 minutes, review included, and buy nothing for this trial. Stop if paid access is needed. This time limit is part of the example, not a research finding. Check that the draft adds no facts to the made-up input.
Redaction reduces risk but does not guarantee anonymity. Other details can still identify a project. Treat removal as risk reduction, not protection.
What should you check before sharing information?
Treat data handling as a set of questions to answer. Ask who sees the inputs and whether your employer receives reporting. Ask how long content is kept and whether it trains the system. Ask how you reach a human, what the exact version and setup are, and what the total price and accessibility are. If no acceptable clear answers exist, use an alternative. If identifying data was already sent, redacting it later does not undo that disclosure. Stop sharing, follow your employer's incident route, and use provider controls. Deletion is not a guarantee. A provider that mentions a standard has not proven compliance with it.
Professional guidance in this field does address disclosure, system limits, data handling, testing, and the separation of coaching from therapy. That guidance is a standard, not an efficacy trial, and it does not guarantee how a specific product behaves (International Coaching Federation, 2024).
Separate voluntary risk guidance notes that generative systems can confidently produce false content, flawed reasoning, or invented citations, and that privacy is a distinct risk. Risk varies with the use and the system, so there is no fixed error rate and not every output is false (Autio et al., 2024). Keep sensitive material out, and verify anything factual before you rely on it. Answer one question first: is this task reflection, rehearsal, a factual answer, an authorized decision, or specialist support?
When do you need a person or another kind of help?
Some situations need a person, and that is not a failure of the tool. Coaching, with or without AI, is not a substitute for qualified professional help or authority. Serious distress calls for qualified support, not coaching. An urgent safety concern calls for appropriate local urgent support. A chatbot should not sit between you and a doctor, lawyer, or clinician. Coaching is not authority for medical, legal, financial, or employment decisions.
Consider what you are buying. Ask a human coach how they keep track of context over time. Do not assume that fit. For health, legal, or money questions, seek the right qualified help. If the task is reflection or rehearsal, a bounded trial may be enough. If a decision affects other people, check who has the right to make it.
How can you review the trial without mistaking use for progress?
Review the deliverable, not the usage. Frequency of use and a pleasant conversation are not evidence that the task improved. Check the actual output against the requirement you wrote down at the start. Did the opening line get tighter? Did the ask get clearer? If the output did not change, review whether the extra work was worth it.
Set a review date and decide: keep, change, or stop. You can also review whether a goal still fits before spending more time on a task that no longer matters, or turn learning into a test at work so the review lives in the workflow rather than in memory. The goal is not to prove the tool works. It is to make one small decision honestly.
If you would like to think this through with a person, you can discuss advisory fit, and for a wider view, explore the full insights library.
Key takeaways
- Match the support to the task. Reflection and rehearsal suit a limited trial; legal, medical, hiring, and financial calls do not.
- Check what each study compared, who took part, and how well it fits your task.
- Check data handling, retention, and escalation before sharing anything. Treat a standards claim as unverified until you confirm it.
- Review the deliverable, not the usage. Record keep, change, or stop on a set date.
Frequently asked questions
References
- Autio, C., Schwartz, R., Dunietz, J., Jain, S., Stanley, M., Tabassi, E., Hall, P., & Roberts, K. (2024). Artificial intelligence risk management framework: Generative artificial intelligence profile (NIST AI 600-1). National Institute of Standards and Technology.
- International Coaching Federation. (2024). Artificial intelligence coaching framework and standards (Version 1.01).
- Schimpf, M., Voigt, J., & Bohné, T. (2026). AI-assisted goal setting improves goal progress through social accountability [Preprint]. arXiv, 2603.17887.
- Terblanche, N., Molyn, J., de Haan, E., & Nilsson, V. O. (2022). Comparing artificial intelligence and human coaching goal attainment efficacy. PLOS ONE, 17(6), e0270255.
- de Haan, E., Terblanche, N., & Nowack, K. (2026). A randomised controlled comparison of the effectiveness of human and AI chatbot coaching with goal attainment, wellbeing and self-efficacy. Human Resource Development International, 29(4), 754-783.
This article is for informational and educational purposes only and does not constitute financial, legal, tax, medical, or professional advice. Individual results vary.