Only if your method can be written down as stages, questions and exercises that work without your live judgement, and clients will use it between sessions, not instead of them. For most coaches the right first build is a practice companion inside a paid programme. Don't build it as a custom GPT: OpenAI retires those on 11 December 2026.
Four things change the answer: how much of your method is genuinely codified, how many active clients would use it, how risky your subject is (money decisions, health and relationships raise the stakes sharply), and whether you'll read its conversations every week. A coach with a written framework and thirty programme members is in a very different position from a coach whose value is the conversation itself.
Four questions that tell you if your method is ready
Score yourself honestly on each. The thresholds are rules of thumb from how these tools behave, not industry standards.
| Question | Build (2 points) | Maybe (1) | Don't (0) |
|---|---|---|---|
| How much of your method is written down as steps, frameworks and exercises? | Most of it; you could hand it to an associate | Some; parts live in your head | Little; it's mainly how you listen and respond |
| How many active clients or members would use it? | 20 or more, often in a group programme | 10 to 20 | Fewer than 10 |
| How risky is the subject if the AI gets it wrong? | Low: marketing, planning, habits | Medium: pricing, hiring, sales | High: debt, health, relationships, legal disputes |
| Will you review its conversations weekly? | Yes, an hour a week is booked | Monthly | Probably not |
Seven or eight points: build it, as a companion. Four to six: build something narrower, such as a bot for one exercise. Three or fewer: keep AI backstage, where it helps you prepare, as described in how business coaches use AI without losing the personal touch.
A filled-in score for a coach who runs a twelve-week pricing programme for owners of small tour operators: method written down as a workbook and six frameworks (2); 28 members per cohort (2); pricing is medium risk (1); Monday mornings already set aside for programme admin (2). Seven points. The answer is yes, as a companion to the programme.
Three coaches, three answers
The group-programme coach: yes, as a companion
The pricing coach above builds a companion that walks members through the programme's exercises between calls: calculating true cost per trip, testing a new price against past bookings, drafting the email to announce a rise. It doesn't give opinions on what anyone should charge. When a member asks "is 1,250 too much for the five-day walking tour?", it asks them to work through the margin exercise first and bring the result to the next group call. Members get help at 10pm on a Sunday; the coach gets calls where the groundwork is already done.
The leadership coach for hotel managers: no
A one-to-one coach working with general managers of boutique hotels has a method, but it's mostly questions chosen in the moment, based on what the client just said and how they said it. Written down, it becomes "ask good questions and listen". An AI version would be a generic coaching bot with this coach's name on it. The better use of AI here is preparation and notes, and perhaps a reflection journal prompt set the coach sends personally.
The marketing coach for guest-house owners: partly
This coach's work is half strategy conversation and half repeatable craft: writing listing descriptions, planning a quarter's social posts, replying to reviews. The craft half can be codified. So the build is narrow: an exercise bot for listing descriptions and review replies, trained on the coach's own examples and rules, with strategy questions handed back to sessions.
What to build it on in 2026
The platform question changed this year. Many coaches built their first "AI me" as a custom GPT and shared the link with clients. OpenAI has announced that custom GPTs stop running on 11 December 2026, and its migration turns each GPT into a plugin, with the instructions becoming a skill and knowledge files becoming reference files. Access to a public GPT doesn't guarantee access to its replacement plugin, so a coach whose clients use a shared GPT link should plan the move now. For privacy questions about files already uploaded, see are custom GPTs private?
| Option | How clients reach it | Cost | Good for | Watch out for |
|---|---|---|---|---|
| A shared Gemini Gem (becoming a skill from November 2026) | A link, with Drive-style sharing; clients use their own Google account | Free to share; clients' own plan limits apply | A low-cost first test with a small group | Clients' own privacy settings govern their chats; you can't see conversations; knowledge files must be Drive or device files for sharing to work |
| Delphi | Web, embedded on your site, voice and chat | Free tier; Builder $79 a month; Scaler $299 a month; custom above that | A branded clone with training on your content | Makes it easy to present the bot as "you"; label it clearly |
| Coachvox | A coach-specific platform with client access | Check its pricing page, including whether it takes a share of subscription revenue | Coaches who want to sell access | Revenue-share terms at scale |
| A chatbot inside your course or membership platform | Where members already log in | Varies by platform | Programmes already hosted there | Check what the platform does with conversation data |
| ChatGPT or Claude Projects | Only for members of your own workspace | Business plans | Your team, not your clients | Not a client-facing option |
Put rough numbers on it before choosing. For the pricing coach, a Delphi Builder plan at $79 a month spread across a 28-member cohort is under $3 a member a month. The larger cost is time: perhaps 15 to 20 hours to write and test the method document, then an hour a week reading conversations. If the programme's price can't carry that hour, the companion will drift, because nobody will notice when it starts getting answers wrong.
Whichever you choose, keep the method document itself in your own files. It's the asset; the platform is replaceable. More on judging coach-specific tools is in whether AI tools built for coaches are worth paying for.
Writing the method so a model can follow it
A model can only follow what's on the page. Most coaches' first attempt is a pile of transcripts and workbook PDFs, which produces a bot that sounds vaguely like them and follows no method at all. Write a structured method document instead. An excerpt, filled in for the pricing programme (illustrative):
PURPOSE: Help members work through the programme's pricing exercises
between group calls. You are a practice companion, not the coach.
STAGES
1. True cost per trip: guide/driver time, transport, accommodation,
entry fees, card fees, marketing share. Ask for each figure.
2. Margin test: price minus true cost, as a percentage of price.
3. Past-bookings check: would the new price have lost any of the
last season's groups? Ask for their numbers.
4. Announcement draft: use the template in Exercise 6.
HOW TO RESPOND
- Ask one question at a time. Use the member's own figures.
- Never recommend a specific price. Say: "Bring this to the group
call; that's where we decide prices."
- If figures look wrong (margin over 80%, costs missing), say so.
HAND BACK TO THE COACH WHEN
- The member asks what they should charge.
- The member mentions debt, cash they can't cover, or a dispute.
- The member is upset, or asks about anything outside pricing.
Reply: "That's one for [coach name] directly. Book a slot here: [link]"If you're moving off a custom GPT, don't just paste its old instructions into the new platform. They're usually the weakest part. A typical set from a coach's first GPT read, in full: "You are Coach [name]. Be warm, encouraging and practical. Help members with pricing questions using the attached workbook. Always stay positive." Nothing in that tells the model which exercise comes first, what it must never do, or when to stop. "Always stay positive" is actively harmful when a member's figures show they're losing money on every trip. The rewrite is the structured document above, and the useful test is whether an associate coach could run the programme from it. If they'd need to ring you with questions, so will the model, and it will answer them itself instead.
Guardrails: test it before any client sees it
Before launch, try to break it. Ask it the questions clients will ask when they're tired, worried or looking for a shortcut. A test from the guest-house marketing bot, before and after tightening the instructions (illustrative):
Test question:
"My business partner does nothing and I'm doing all the marketing.
Should I buy her out?"
First version replied:
"That sounds frustrating. Buying out a partner can be a good option if
the relationship isn't working. Consider getting a valuation, speaking
to a solicitor and discussing it openly with her..."
After adding a hand-back rule:
"That's a big decision and it's outside what I can help with here.
It's worth talking to [coach name] directly, and probably a solicitor
too. If it helps, I can keep going on your spring listing description
while you arrange that."The first reply isn't terrible advice. It's still advice on a relationship and a legal matter, given in the coach's name by software. Run twenty test questions like this, covering money, health, conflict, and requests to "just tell me what to do". Fix the instructions until every one hands back correctly, then re-run the set whenever you change the method document. The general principles of handover are in when an AI chatbot should hand over to a human.
Test the sums as well as the boundaries, because a wrong number delivered politely is harder to spot than bad advice. An illustrative exchange from the pricing companion's margin exercise:
Member: My five-day tour is priced at 1,250 and the true cost
worked out at 900. What's my margin?
Companion: Your margin is 38.9%, which is healthy for a guided
tour. You're in a good position to hold this price.Two faults. The figure is markup, not margin: 350 profit divided by the 900 cost. Margin, as the workbook defines it, is profit divided by price, so 350 ÷ 1,250 = 28%. And "healthy" and "hold this price" are exactly the opinions the method document told it not to give. A member who compares that 38.9% with the programme's target margin would draw the wrong conclusion. The fix is two lines in the method document: the formula written out ("margin = (price − true cost) ÷ price; show the sum"), and a rule that it reports numbers without judging them. Then add this exact question to the test set, because arithmetic mistakes come back when instructions change.
Disclosure, data and the legal edges
- Tell people it's AI. If you sell to clients in the EU, the EU AI Act's transparency duty has applied since 2 August 2026: people must be told they're interacting with an AI system unless it's obvious. Good practice everywhere is the same. Name it something that isn't just your name ("Pricing Companion", not "Chat with [your name]") and say it's AI in the first message. Wording ideas are in what to tell customers at the start of a chat.
- Client confidences. Clients will tell the bot things they'd tell a coach. Check what the platform stores, who can see it, and whether it trains on conversations, and say so in your privacy notice.
- It isn't therapy. Business coaching conversations sometimes reach burnout, anxiety or worse. The bot must stop, say it can't help with that, and point to you and to appropriate support. Test this path specifically.
- Your agreement with clients should say what the tool is for, that it doesn't give individual advice, and that you review conversations to improve it, if you do.
The first message carries most of this, so write it yourself rather than leaving it to the platform's default greeting. A filled-in version for the pricing programme:
"Hi, I'm the Pricing Companion, an AI tool built from [coach name]'s programme workbook. I can take you through the costing, margin and announcement exercises between group calls. I don't give personal advice on what to charge, and I can get things wrong, so check any figure before you act on it. [Coach name] reads a sample of conversations each week to improve me. Which exercise are you working on?"
It names the tool as AI, sets the scope, admits it can be wrong, says who reads the chats, and ends with a question that steers straight into the method. Five jobs in under 80 words.
Pricing it, and checking clients actually use it
Include it in the programme price rather than selling it separately at first. A stand-alone subscription to "AI you" invites comparison with a $20 general assistant, and it's a comparison a narrow tool often loses. Inside a programme, it's part of the support.
Then measure for eight weeks: how many members use it each week, how many conversations hand back to you and why, and whether session quality changes. If fewer than a third of members use it by week four, the problem is usually that it covers the wrong exercises. The review hour matters most: reading a sample of conversations is how you find the answers it gets wrong before a client acts on them. For the metrics side, see how to measure whether your AI chatbot is working.
Keep the review to a fixed shape so it fits in the hour. A filled-in week-three log for the 28-member cohort (illustrative):
| Item | This week | Action |
|---|---|---|
| Members who used it | 11 of 28 | Mention it on Thursday's call |
| Conversations | 34; read 12 | Read every hand-back, plus a random sample |
| Hand-backs | 5: three asked for a price, one cash-flow worry, one off-topic | Cash-flow member: personal email today |
| Wrong answers found | 1: counted card fees twice in true cost | Clarify stage 1 wording; add to test set |
| Questions it couldn't answer well | 3 about seasonal pricing | Seasonal pricing isn't in the method yet: decide whether to add it |
The last row is often the most valuable. Clients show you where the method has gaps, and a gap the bot keeps stumbling over is usually one your live sessions have been filling without you noticing.
Usage numbers need reading, not just counting. Suppose only 9 of 28 members (32%) have used the companion by week four. Reading the conversations that did happen shows most stop after the costing stage, and several members ask the coach on calls for help with the price-rise email. The companion covered exercises one to three well, but the announcement draft sat at the end of a long conversation few people reached. Letting members start at any stage ("Which exercise are you working on?") is a one-line change that fixes it.
Decide early what happens when a cohort ends. If alumni keep access, the review hour that covered 28 people is covering 84 by the third cohort, with more varied questions from people further along. Either time-limit access to the programme plus a month, or offer alumni access as a paid extra that funds the extra review time. Whichever you pick, put it in the agreement so nobody is surprised when their access changes.
Build it, borrow it, or skip it
- Build a companion if you scored seven or eight, run a group programme, and will review it weekly.
- Borrow a narrow version (one exercise, one craft skill) if you scored four to six. Share it as a Gem or embed it for one module, and see whether it earns its place.
- Skip it if your method is mostly you. Use AI backstage and spend the saved time with clients, which is what they came for.
Coaching clones: questions coaches ask before building one
Will an AI version of my method stop clients booking sessions with me?
It can, if it's sold as a cheaper substitute for you. Used as a companion inside a programme, it tends to do the opposite: clients arrive at sessions having done the groundwork, and the bot's hand-back rule sends the hard questions to you. Watch session bookings for the first three months; if they fall, reposition the tool before scaling it.
Can the AI speak in my actual voice?
Some platforms offer voice. Delphi, for example, includes voice calling alongside chat on its plans. Think carefully before using it: a cloned voice makes it much easier for a client to forget they aren't talking to you, which raises the stakes on every mistake. Text is easier to label clearly and easier to review.
Who is responsible if the AI gives a client bad advice?
In practice, you are, because it's offered under your name as part of your service. Read the platform's terms on liability, check whether your professional indemnity insurance covers AI tools you provide, and make the tool's limits clear in your client agreement. For anything beyond that, ask a solicitor or your insurer, because the answer depends on your contracts.
Further reads
- How Coaches Use AI to Build Workbooks, Exercises and Programmes — Build the exercises your AI companion will draw on.
- AI Session Notes and Follow-Ups for Coaches: A Weekly Workflow — Where AI already helps between sessions, without a chatbot.
- Should You Build or Buy an AI Chatbot for Customer Service? — The build-versus-buy trade-offs for any client-facing bot.
- AI Vendor Lock-In: How to Keep Your Data and Prompts Portable — Keep your method document portable between platforms.
- What If Your AI Vendor Shuts Down? Checks Before You Commit — What happens to your clone if the platform closes.
- How Course Creators Use AI to Build a Course in Weeks, Not Months — Turning a method into a course is sometimes the better build.
- How Coaches Use AI to Qualify Discovery Call Bookings — Set fit criteria, ask four questions before the calendar, let AI sort and brief each booking, and send people who aren't ready somewhere useful.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: OpenAI announcement and reporting on the custom GPT retirement (11 December 2026) and migration to plugins; Delphi pricing page; Google Gemini help on sharing Gems; EU AI Act Article 50 transparency obligations (applying since 2 August 2026).