CommunityPricing

How it works

OverviewFind a practice partnerSchedule and callTranscripts and feedbackLogbookExam practiceSession recordingRobin’s feedbackAI Practice

Who it’s for

For coachesFor coaching schoolsFor coaching communities

Resources

GuidesBlogHelp centreFAQ
← All posts

11 September 20267 min readChecked 11 September 2026

Will AI replace human coaches?

Anyone offering you certainty in either direction is going beyond what has been measured.

Some coaching-shaped tasks are already being done by software, at least one narrow tool has produced results worth taking seriously, and the studies discussed here establish neither imminent profession-wide automation nor protection from it. Anyone offering you certainty in either direction is going beyond what has been measured.

Three questions that get mixed together

Coaching tasks. Some software already asks questions, records goals and follows up on progress.

Client purchases. Substitution depends on price, access, preference and what someone wants when they book a session.

The profession. Profession-wide replacement is a labour-market question. These coaching trials measure outcomes rather than market adoption.

What the evidence shows

A narrow AI coach produced similar self-rated goal-attainment results

One comparison put a chatbot called Vici against human coaches on self-rated goal attainment (Terblanche, Molyn, de Haan and Nilsson, PLOS ONE, 2022). Both interventions beat their controls. The two intervention groups "never significantly differed between each other throughout the experiment".

These were two separate ten-month trials run in different years, because the AI coach did not exist when the first ran: human coaching from October 2017 to July 2018, AI coaching from November 2019 to August 2020. Participants were undergraduates at a UK business school, and 327 of the 478 allocated submitted data at all eight time points. The outcome was self-rated. The authors also note that between the seventh and eighth measurement, human-coached participants received no more coaching while AI-coached participants could keep using the app, which is a plausible reason the lines converged at the end.

That is not a head-to-head test of whether AI can do a coach's job. It reports similar self-rated goal-attainment results for a scripted goal-support tool and human coaching, not an equivalence or non-inferiority result. Most goals were concrete (60%); study-related goals were the largest category, at 38%.

The paper's discussion goes further, speculating about which levels of coach maturity might be replaceable. That is the authors interpreting, not reporting. It should not be quoted as a finding.

The relationship study did not involve an AI

In one study, fifty-two graduate students had one 60-minute session structured by the CLEAR model, half of them believing they were being coached by a conversational avatar (Frontiers in Psychology, 2025). The avatar was operated by five professional human coaches acting as blinded confederates, a Wizard of Oz design chosen deliberately "to sidestep the rapid obsolescence of technology". Working-alliance ratings did not differ significantly. Although recruitment met the separate target of 25 per group, the paper's power analysis required N=102 for 80% power to detect a medium effect; N=52 fell short. Non-significance does not establish equivalence.

This offers limited evidence about willingness in that setting. Clients who believe they are talking to an AI can still form a working alliance, at least in one session. It says nothing about what an autonomous system could do, because there wasn't one.

A 2026 direct trial favoured human coaching

A 2026 randomised controlled experiment compared accredited human coaches and automated AI coaches with 114 coachees (de Haan, Terblanche and Nowack, Human Resource Development International 29(4), 754–783, DOI 10.1080/13678868.2026.2633990). Its abstract reports substantial effectiveness across several coaching outcomes only for human coaching, with effect sizes described as mid-to-high.

This account is limited to the public abstract. The allocation details, intervention design, attrition and outcome tables are not available in that summary. It does not establish that every human coach outperforms every AI system. One trial of 114 people cannot forecast employment across a profession.

The publisher page also acknowledges commercial interests: one author developed the chatbot used in the study and another has an interest in a measure used. The full methods would be needed to assess how the trial handled those interests.

A systematic review reports task-specific findings

The abstract of a 2026 systematic review reports searches of eight databases up to March 2024 and 16 included studies covering 2,312 participants (Passmore, Olafsson and Tee). It says AI coaches can be effective, accepted and useful, and match human coaches in competence for specific tasks.

The abstract cannot establish how well each included comparison was designed. Its search date also means it cannot assess products or trials introduced after March 2024. Neither task-specific findings nor a review's abstract establishes profession-wide equivalence.

What the profession has said

The ICF's AI Coaching Framework and Standards sets out guidance "across six domains: foundational ethics, co-creating the relationship, effective communication, learning and growth facilitation, assurance and testing, and technical factors like privacy and accessibility". It names accessibility and efficiency as opportunities alongside risks including bias and confidentiality, and positions AI as "a valuable complement to human coaching" (ICF).

That is a professional body stating a position and setting expectations for providers. It is not evidence about what will happen, and a framework describing AI as a complement does not prevent buyers from treating it as a substitute.

Where this becomes judgement

From here on, this is Robin's editorial reading of the evidence above, not a research finding.

The evidence pattern points towards substitution by task rather than by profession, at least for now. What has been shown to work is narrow, structured and repeatable: holding a goal, asking about progress, prompting an action plan on a schedule. The PLOS authors themselves attribute Vici's performance to the "rigor and mechanistic execution of goal theory" and its inability to deviate from a set process. Some of what people currently pay coaches for looks like that.

Other parts of the work look nothing like it: contracting with an organisation about what coaching is for, holding a conversation in which the stated goal turns out to be the wrong one, noticing what the client is avoiding, deciding this person needs a therapist rather than a coach and saying so. The evidence reviewed here does not establish autonomous performance of those tasks, or show that it is impossible. Their difficulty is an argument for caution, not a guarantee of employment.

Price affects the choice: a subscription tool competes for a client's budget even if coaches dislike it. Accountability also needs checking. A human coach can offer a named practitioner, but check their professional membership, insurance and complaints arrangements rather than assuming those protections exist. A software provider has its own terms and accountability arrangements. What clients want varies; a study of undergraduates pursuing concrete goals cannot tell you what your clients will buy.

"AI-proof" is not a defensible claim and shouldn't be sold as reassurance. Neither is the mirror-image line that only mediocre coaches will be replaced, which flatters the reader and rests on nothing.

What you can actually do

Ask what job your clients are hiring you for. Someone who wants a structured nudge towards a goal they have already named is buying something a tool can imitate. Someone who wants to work out what is going on before they decide what the goal is, or who needs a person to be accountable to, is buying something else. If your practice contains both, separating them can help you consider where you're exposed.

Review with clients what they wanted from coaching and what changed. Keep their reports distinct from proof that coaching caused the change; these trials do not establish which way of describing your work will protect your practice.

Be straightforward about tools you use. If AI touches a client's material at any point, in note-taking, transcription or preparation, say so, agree it, and check it against your body's ethics and your obligations to the client. The ICF framework addresses providers, developers and organisations. Your own professional duties and agreements still apply when you use their tools.

Keep developing your contracting, reflective work in supervision, ethical judgement and attention to what a person isn't saying, without assuming these skills are immune to technological change.

What would change the answer

Look for concurrent trials comparing autonomous systems with experienced human coaches in the populations coaches serve, with longer follow-up and measures of harm as well as benefit. Adoption and pricing data from real buying decisions would also help answer the separate substitution question.

Sources