Yes, with two conditions, and both are quieter than the thing most coaches are worried about. Your profession’s own standards body published an AI framework two years ago. It assumes you’ll use these tools, and it says which kinds it’s watching. What it asks of you is this. Know which side of one line your software sits on, and be able to account for where a client’s words went.
- The permission question is settled, and the ICF settled it. The published framework treats offloading scheduling and reminders as the normal case.
- The line that matters runs between software that helps you work and software that reaches your client on its own. One ordinary piece of coaching practice crosses it quietly.
- Disclosure is a narrower duty than it feels like. The standard’s “tell them it is not a person” requirement is scoped to systems that talk to clients. What reaches your scheduling tool is a duty about data.
- The condition that costs you something is being able to answer for your vendor. The standard tells you exactly what a vendor is supposed to have published.
What the profession has already said
Most coaches researching this expect silence, or a warning. The profession has published something warmer than either.
The ICF Artificial Intelligence Coaching Framework and Standards is marked V1.01, with the revision mark rev. 2024.10.08 (checked September 6, 2026). It came out of a working group the ICF convened in April 2021 and finalized in 2024. Its foreword is direct:
This blended approach can provide the best of both worlds where coaches can offload tedious coaching activities such as delivering assessments, scheduling appointments or sending reminders to journal. As a result, AI used well allows coaches to invest more in the high value work of complex, transformational coaching in the human-to-human domain.
That’s the professional body, in its own standard, describing the thing you’re wondering whether you’re allowed to do. The Code of Ethics is marked “Board approved October 2024” and “Revised May 2026” (checked September 6, 2026). It points the same way from the other direction. Standard 2.5 has you fulfill your obligations “directly and through any technology systems I may utilize (i.e. technology-assisted coaching tools, databases, platforms, software, and artificial intelligence).” It names AI to say your duties travel through it. That’s a sentence you write about something you expect people to use. Handing coaching admin to AI and staying yourself works that standard through properly.
It helps to be clear about what’s at stake here, because coaches often picture a board that could stop them practicing. In the United States, coaching is an unlicensed occupation. The Code describes an ICF credential as “a professional certification indicating a person has met specific standards and requirements designed to develop and refine their coaching skills.” Your certification and your reputation are what’s exposed. Both matter, and both respond to different precautions from the ones people picture.
So the permission question is closed. Three questions were hiding behind it, and they’re the rest of this guide. Which of your tools the standards are watching. What you owe a client, and when. Whether you can answer for the company holding the mail.
Which side of the line your inbox falls on
The framework’s most useful move is a taxonomy. Learn it, and an anxious question becomes a sorting job you can do in five minutes.
It divides AI coaching applications into four types, and puts a wall down the middle:
in general, the first two: scheduling and data processing, are coach assisting tools. They primarily support a coach in their provision of a human or blended Coaching Service. The latter two, interactive and conversational are different in that they are providing Coaching Services to the client directly.
Scheduling is what it sounds like: meeting planning, following up, reminders. Data processing is defined as “Analysis tools which process data that is accessible in a given environment (e.g. emails, calendars, feedback)” that “generate relevant outputs or measurements that can provide information for the coaching process.” The framework adds that “These outputs then require a coach or other entity to use them in order to be impactful.”
Emails and calendars, named. A tool that ranks your inbox, summarizes a thread, extracts a task or prepares a draft you read and send is a data processing application in the profession’s own vocabulary. It sits on the coach assisting side of the wall.
Then the framework says which side it’s watching. It and the associated standard “focus primarily on the latter two categories of application,” meaning the interactive and conversational ones. The framework gives two reasons. Client-facing systems process more sensitive material. More importantly, they “may be providing a Coaching Service independently of a human coach.” What the standard is watching for is software that reaches your client on its own.
That produces the one finding in this guide worth writing down somewhere. The interactive type is a Coaching Service Application. Under it, the framework gives these examples: “habit-building quizzes made through a mobile application and email prompts relating to a specific goal or topic.”
Email prompts relating to a specific goal. That’s the automated between-session nudge sequence. It’s the accountability check-in that goes out on day four. It’s the “how did the conversation with your director go” that fires on a timer. Coaches build these constantly and file them mentally under admin, because they’re automated and short. In the profession’s own taxonomy they’re coaching delivered by software, and they land in the category the standards are actually written about, alongside chatbots.
That’s the line. It isn’t who typed the words. It’s whether the system reaches your client on its own initiative and does something coaching-shaped when it gets there. Sorting your practice against it takes one pass:
Coach assisting, and uncontroversial. Triage and ranking. Summaries you read. Task extraction. Scheduling, invites, reminders about times. Drafts you edit and send. Filing. Chasing an unsigned agreement.
Coaching service, and a different order of decision. Goal nudges sent on a schedule. Reflective prompts. Anything conversational the client talks to. Progress feedback generated and delivered before you’ve read it.
The second list is allowed. It’s a product, with a standard attached that runs to thirteen sets of requirements, and building one is a different undertaking from buying an email tool. If you’re running one already, start with A.1.7. Its basic form is that the system “shall require the Client to agree to the AI coaching, stating that a human will not be involved.”
The boundary drawn on principle rather than category, including the material that stays away from all of this, belongs to what AI should never touch in a coaching practice.
What you actually owe a client
Now the question people came for. Do you have to tell them?
The standard is more precise than the debate. Element A.1.1 is called AI Disclosure, and it says what you’d expect: “The user should know when they’re interacting with an algorithm and not a person. People behave differently when they think they’re speaking to a human, and being deceived can damage trust.” Its basic requirement is that the system “shall indicate to the Client that it is not human prior to the provision of Coaching Services.”
And then the line that decides your case. A.1.1 is marked Applicable to: Interactive, Conversational. Those are the systems a client talks to. Your scheduling and your data processing tools sit outside it.
The duty to announce the machine attaches to a machine the client talks to. The software that decided which of your forty threads to show you first sits outside that duty. So does the software that wrote the first pass of a reply you read and sent. The Code does the same thing elsewhere, and it’s worth noticing, because it shows the drafting is deliberate. Standard 3.7 requires you to disclose when you’re “acting in a capacity other than the role of an ICF professional.” The Code knows how to require a disclosure. It requires them about who you’re being.
Three requirements do reach your admin tooling, and all three are about data. Element A.1.6 is Data Transparency. F.13.1 is Data Management Disclosure to Clients. F.13.2 is Consent for Data Processing. Each is marked applicable to all four application types. F.13.1 asks that clients be informed “about data management practices, specifically regarding security and confidentiality, before providing Coaching Services.” F.13.2 asks for “explicit consent from the Client for any data processing activities.”
Those land on the software vendor rather than on you. Read them next to Standard 2.1, though, which asks you to “Maintain the strictest level of confidentiality with all parties involved, regardless of the role I am fulfilling.” The shape of your own obligation is then plain. Your client is owed the truth about who now holds what they wrote you. A note explaining that a program formatted your session confirmation is beside the point.
So the disclosure worth making is a custody disclosure, and it goes in the agreement. Standard 1.1 puts the timing beyond argument. It asks you to communicate “(before coaching begins)” to “explain the nature of coaching and to co-create a coaching agreement regarding roles, responsibilities, confidentiality, financial arrangements, and other aspects of the coaching engagement.” The Code’s own definition of a coaching agreement says it typically covers “confidentiality policies.” You already have the document. It already has the section.
Three or four sentences do it. Something in this shape, adapted to what you actually use:
Our sessions are private. Between sessions we correspond by email, and my mailbox runs on software that sorts it, summarizes it and prepares draft replies for me. I read and send everything that comes to you from me. My provider’s list of who can access that data is at [link], and if you’d rather a particular conversation stayed out of it, tell me and we’ll use another channel.
That’s short, true and checkable. It reads as terms rather than as a confession. That matters more than it should, for a reason worth being honest about.
Volunteering it costs something. Jessica A. Reif, Richard P. Larrick and Jack B. Soll ran “four preregistered experiments (N = 4,439)” and found “that people who use AI at work anticipate and receive negative evaluations regarding their competence and motivation.” Their paper is Evidence of a social evaluation penalty for using AI, published in the Proceedings of the National Academy of Sciences in 2025. Its conclusion is stated as a dilemma rather than a recommendation: “Although AI can enhance productivity, its use carries social costs.” The related finding, that people are judged on suspicion whether or not they used anything, is in keeping your voice when AI drafts your email replies.
The lesson there is about placement. A disclosure delivered as an announcement invites the judgment. A disclosure delivered as a term of business gets read the way your cancellation policy gets read. Put it where terms live. Send it as its own email in month four and you’ve staged an event.
Concealment is the more expensive path anyway, and the Code says why in one sentence. Standard 5.2 asks you to “Make verbal and written statements that are true and accurate about what I offer as an ICF professional.” Say on your site that every word comes from your own hands, while software drafts your replies, and you’ve made a written statement you’d have to defend. The exposure is the statement, rather than the smooth paragraph somebody noticed. The Code’s professionalism value asks for “Behaving with respect and transparency in all business dealings related to coaching.” The cheapest way to hold it is to claim only what’s true.
The legal floor and what it does not reach
Federal law is quiet here. No statute requires a coach to tell clients that AI helped write an email, and the professional standards leave it alone too. Several states have begun legislating in the neighborhood. Utah’s statute speaks most directly to a question like yours, and it’s worth reading even though it binds only there.
Utah Code Title 13, Chapter 77 covers generative artificial intelligence, consumer disclosures and enforcement. It’s been effective since May 7, 2025. Section 13-77-103(1) is the rule most coaches would guess at:
A supplier that uses generative artificial intelligence to interact with an individual in connection with a consumer transaction shall disclose to the individual that the individual is interacting with generative artificial intelligence and not a human, if the individual asks or otherwise prompts the supplier about whether artificial intelligence is being used.
Answer when asked. The statute adds that the question “must be a clear and unambiguous request to determine whether the interaction is with a human or with artificial intelligence,” so it takes a direct question to trigger the duty. A heavier duty, a prominent up-front disclosure, applies under 13-77-103(2) only to a “regulated occupation.” That’s defined as one “regulated by the Department of Commerce” that “requires an individual to obtain a license or state certification.” Coaching sits outside that definition. The Division of Consumer Protection can impose an administrative fine “of up to $2,500 for each violation.”
Two limits on how far any of this reaches you. The chapter binds suppliers in Utah consumer transactions, so a coach in Ohio with clients in four states falls outside it. And its definition of generative artificial intelligence is a system “designed to simulate human conversation with a consumer” that “generates non-scripted outputs” with “limited or no human oversight.” A drafting tool whose output you read and send is a different thing.
The reason to know the rule anyway is that it’s a good rule, and it’s the one to run your practice on wherever you are. Let it sit in your terms rather than in an announcement. Answer plainly when somebody genuinely asks. And when the answer is yes, say yes.
When a client asks you outright
Clients rarely ask the question they mean, so answer the one underneath. It’s almost always one of three, and each wants a different length of answer.
“Do you use AI?” Curiosity, usually prompted by something in the news rather than by anything you did. Match the register and stop.
For scheduling and for sorting my inbox, yes. My software drafts replies for me sometimes, and I read and send all of them. The coaching is me.
“Is what I tell you going into ChatGPT?” This one is about custody, and it deserves specifics. Name the company. Say where the terms are. Say what you know, and say where your knowledge stops.
My mail runs on [product]. Their terms and their list of who can access data are published at [link]. What I can tell you for certain is what leaves this practice, and nothing about you goes anywhere I haven’t named.
If those blanks are hard to fill in, that’s the finding, and the next section is about fixing it.
“Did you write that, or did a machine?” This is the wounded version, and it’s a different conversation. Something in your message landed wrong, and the tool is what they have to point at. Answer the factual question in one sentence. Then go to the real one.
I wrote it. Something in it landed wrong, though. Tell me what.
Two rules hold across all three. Answer in the register you were asked in, because escalating a light question into a policy statement is alarming in itself. And when the answer is yes, say yes. A client who finds out later has caught you saying something untrue, which is the version of this that does permanent damage.
If the answer is that they’d rather you didn’t, take it. A client’s refusal settles the matter for that client, and honoring it is cheaper than the alternative: keep their correspondence off the tooling, or move it to a channel the software can’t read.
The four things a vendor is supposed to have published
This is the condition with teeth, and the standard makes it easy to check. Four of its requirements apply to scheduling and data processing tools as well as to chatbots. Each asks the provider to publish something. So the test is whether the documents exist.
Who can reach the data. Element A.1.6 asks the provider to “supply documentation identifying who within the Provider organization will have access to coach and client data.” At the advanced level it asks for “publicly available documentation identifying all parties who may have access to coach and client data and where the data will be stored.” All parties means the subcontractors as well as the vendor. A published subprocessor list is what meeting this looks like.
What the software does. Element A.1.3 asks for “publicly available documentation on the types of algorithms used and their basic functionality,” and it explicitly stops short of asking for trade secrets. A plain description of what the software does clears it.
Why it did what it did. Element A.1.4 asks for “publicly available documentation on the general ways in which decisions are made by the AI Coaching System.” For an email tool, that’s the difference between software that ranked your inbox and software that can show you why.
Where it might be wrong. Element A.1.5 asks the system to notify you of “any potential biases in the Application.” In practice this one yields less than the others. A vendor who says something specific and unflattering is the better sign.
The companion practical guide turns these into a buyer’s checklist, and its security section is worth borrowing:
- encryption in storage and transmission
- secure sign-on such as multi-factor authentication
- collecting only necessary data
- giving users “control over their data access and sharing preferences”
One item on that list, “Adheres to privacy laws like GDPR,” points at European law, which reaches European practices rather than American ones, so read it as information rather than as a badge. The rest of the list is regime-neutral, and that’s the part to use.
Then the part the standards leave alone, because it’s a promise problem rather than a documentation problem. The commitments hold for as long as the vendor keeps them. On February 13, 2024, the Federal Trade Commission said that “It may be unfair or deceptive for a company to adopt more permissive data practices,” giving as its examples starting to share consumer data with third parties or using it for AI training. The rest of that sentence covers how the change gets made: “and to only inform consumers of this change through a surreptitious, retroactive amendment to its terms of service or privacy policy.” That’s a real constraint on the company, and a poor substitute for a copy. Save the privacy page you relied on, dated, on the day you sign up. It costs a minute, and it’s the only version of the promise that will still exist in two years.
The general skill of questioning a vendor is in questions to ask any AI tool about your data, including what to accept as an answer and which claims you can test rather than ask about. Whether your mail becomes training material is in keeping client data out of AI model training. Both are written for a general reader, and both hold for a coaching practice.
What Point publishes
Point runs on the mailbox you already have, Gmail or Microsoft 365. Everything stays where it is, and your clients keep writing to the same address. In the framework’s vocabulary Point is a data processing application. Point reads the mail and calendars already in your account, and produces rankings, summaries, tasks and drafts for you to use. That’s the coach assisting side of the wall, and Point stays on it by design. Point sends your clients no goal or accountability sequences, and there’s no chatbot here for a client to talk to.
Against the four documentation asks. The privacy policy and the subprocessors list are the published answer to who can reach the data and where it goes. That’s the A.1.6 question, and the one your client is really asking. On explainability, every action taken for you is written down with a timestamp and the reading it was made on. That’s the activity log, and most entries can be reversed from there. What Point has learned about how you work is kept as ordinary sentences you can read down, edit and delete. That’s seeing what an AI learned about you.
How far Point goes is set per kind of action rather than once for everything. The range runs from suggest-only, through preparing work and waiting for you, to handling that kind of work outright. A new account starts every action type in the middle. Where each position stops is laid out on the dial’s page. A message that has already gone out sits on somebody else’s server and can’t be recalled, here or anywhere, which is the limit on undo.
For the client who asks you to keep something out of it, a conversation can move onto a channel locked end to end. It stays between the two of you, and Point has no way to read it. The cost is real and worth saying in the same breath. A locked thread gets no summary, no ranking and no draft.
Three honest limits. Pricing is private beta and per seat, so anyone comparing costs should ask rather than infer. Point is an email tool, and calls and session transcripts sit outside what Point does, so the recording question is one you settle with your client and your video platform. And your agreement clause is yours to write, as is the answer you give a client who asks. The full inventory lists everything Point handles, and the coaching page walks the same ground a week at a time.
What this does not settle
Four things this guide leaves to you, and it’s better to know which they are.
Your sponsor’s contract. In a corporate engagement, the organization paying may have vendor rules about what systems can hold data on its employees. Those rules are contractual rather than ethical, and they’re the ones that get enforced. Read what you signed before you decide anything on the basis of a professional standard.
A client who has said no. A refusal settles it, whatever the framework says. If somebody asks you to keep their correspondence off the software, that’s your answer.
Whether you’re still doing the job. The standards cover permission and custody. They’re silent on whether the person on the other end is getting a coach, and that’s the question that matters. Only you can say whether the reply you sent last Tuesday was one you’d have written.
A caseload problem. If you want the tooling because you have more clients than you can hold in mind, it will take the symptom away and leave the cause. Software that remembers for you is a genuine relief. Having room for somebody is a different thing.
Common questions
Is it ethical for a coach to use AI?
Yes, for the work that supports coaching. The ICF has published an AI framework whose foreword describes offloading scheduling and reminders as what AI used well makes possible. The Code of Ethics runs your existing duties through whatever technology you use, which is what Standard 2.5 says, rather than setting up a separate AI regime with a permission to seek. The ethical content sits in three places. What the software is allowed to do on its own. Whether your clients know who holds their correspondence. Whether you’ve kept the parts of the work that are the work.
Do I have to tell my clients I am using AI?
Tell them about custody rather than about authorship. The ICF standard’s “tell them it is not a person” requirement is marked applicable only to interactive and conversational systems, meaning things a client talks to. What reaches a scheduling or inbox tool are the requirements about data. Clients are informed of data management practices before services begin, and consent is obtained for data processing. So the disclosure worth making is about custody rather than about who typed the sentence. Its natural home is the coaching agreement, which the Code already expects to cover confidentiality policies.
What should I actually put in my coaching agreement?
Three or four sentences, covering four facts. Your mail is handled by software that sorts, summarizes and drafts. You read and send anything that comes to them from you. The provider’s data policies can be read at a place you name. And a client can ask for a conversation to be kept off it. Standard 1.1 asks you to co-create the agreement before coaching begins, so this belongs in the version a new client signs rather than in a message sent later to people already working with you.
Is there a law requiring me to disclose AI use to clients?
No federal one. Utah’s statute has been effective since May 7, 2025. It requires a supplier using generative AI to interact with someone in a consumer transaction to say so, if that person makes a clear and unambiguous request to know. A heavier up-front duty is reserved for licensed occupations, and coaching is an unlicensed one. The statute binds only in Utah, and its definition of generative AI covers systems that converse with the consumer rather than drafting tools whose output you read and send. Answer honestly when asked is a good rule everywhere, whatever a statute compels.
Will using AI put my ICF credential at risk?
The use itself is fine. The risk sits in the specific standards that reach the practice however the work gets done. Confidentiality under 2.1. How you handle records. The agreement about how information is exchanged. And the accuracy of what you say about your own service under 5.2. That last one is underrated. Marketing a practice as entirely handwritten while software drafts your replies is a truthfulness problem, and it’s a worse position than leaving the claim off your site.
What if a client asks me not to use AI at all?
Honor it, and say what it means concretely rather than agreeing in general. In practice their correspondence stays off the tooling, or it moves to a channel the software can’t read. It may also mean you answer them more slowly. Say that at the time. A boundary you negotiated honestly holds up better than an agreement you can’t keep.
Are automated check-in emails to clients admin or coaching?
Coaching, by the profession’s own classification, which surprises most people. The ICF framework lists “email prompts relating to a specific goal or topic” as an example of an interactive application. Interactive applications are Coaching Service Applications, in the category the standards are principally written about. A confirmation of Thursday’s time is admin. A nudge about the goal is the work, arriving by machine.
Is an AI coaching chatbot allowed?
The standards allow it, and it’s a different product from an assisted practice. The ICF standard is aimed at exactly this, and it runs to thirteen sets of requirements covering ethics, agreements, listening, growth, testing and security. The one to know before anything else is A.1.7. Its basic requirement is that the system have the client agree to the AI coaching “stating that a human will not be involved.” Read that sentence and ask whether you’d put it in front of your clients. Your answer there is the answer to the whole question, and the rest of the standard gets harder from there.
Can I let AI write the note I send between sessions?
That one is a no, and permission has little to do with it. The between-session note is delivered coaching rather than correspondence about coaching. A drafting model’s instinct is to answer, reassure and close, which is the opposite of the move. Keeping your voice when AI drafts your email replies sets out which messages can start from a draft and which are yours from the first word.
The short version
You’re allowed. The ICF published an AI framework in 2024 that treats a coach offloading scheduling and reminders as the normal case. Its Code of Ethics names artificial intelligence to say your existing obligations travel through it. In the United States coaching is an unlicensed occupation, so what’s exposed is your certification and your reputation.
The line the standards draw runs between tools that assist you and tools that reach your client on their own. The second category includes something most coaches file as admin: automated email prompts about a client’s goal. Sort your practice against that line first, because it decides everything after it.
What you owe clients is narrower than the worry, and harder to fake. A true account of who holds their correspondence, put in the agreement before coaching begins. A plain answer when somebody asks, and a yes when the answer is yes. The one condition that takes real work is being able to name the company, point at what it has published about who can reach the data, and keep a dated copy of the promise you relied on.
The test a year from now is simple. A client asks you the direct question. You give a specific answer in one sentence, and it’s the same answer you gave your last client.