It's a Sunday in March and I'm at my desk with a cup of green tea that stopped being warm about an hour ago. I have five sessions' worth of notes to write. My shorthand from Tuesday says "board dynamics, loss of trust, consider 360?" and I'm sitting here trying to reconstruct what that actually meant. Something about how his CTO stopped showing up to one-on-ones. Or was that the Thursday client? Outside, it's one of those afternoons where the light is doing something nice and I should be anywhere but here.
This was my life for the better part of seven years. Not every Sunday, but enough of them. I'd tell myself I'd write notes after each session, and sometimes I did, but most weeks the afternoon sessions would stack up and I'd push them to evening, then to the weekend, then to that guilty Sunday window where the details had already gone soft at the edges.
Here's what I've figured out about using AI to write session notes. Not the theory, not the hype. What actually works, what I tried that was terrible, and where I've drawn lines I won't cross.
The first attempt was bad
I should tell you about the first time I tried this, because if your experience was anything like mine, you probably gave up too early.
About a year ago, I had a particularly brutal Thursday. Four sessions back to back, the last one running over by fifteen minutes. I sat down at my desk afterward, looked at my scribbled notes, and on impulse opened up an AI chat and typed something like: "Here are my notes from a coaching session. Please write a professional session summary."
What came back was technically accurate and completely useless. It read like an HR report. Clinical language, passive voice, phrases like "the client expressed concern regarding interpersonal dynamics within the leadership team." Nobody talks like that. I certainly don't write like that. If I'd put those notes in a client's file, they would have been worse than having no notes at all, because they would have flattened everything that actually mattered about the session into corporate wallpaper.
So I closed the tab and went back to writing notes by hand. For about three weeks.
What changed
The thing that changed wasn't the tool. It was how much context I was giving it.
That first attempt failed because I handed over the raw material and expected the AI to know what to do with it. But it doesn't know my clients. It doesn't know my voice. It doesn't know that when I write "trust rupture, board level" I mean something very specific about how this particular leader's relationship with his board has been eroding since Q3, and that we've been working on how he shows up in those meetings for the past four sessions.
The AI had none of that. So it did what it does when it has no context: it produced something generic.
The approach that actually works is what I've started calling the context sandwich. Three layers, every time.
The context sandwich
Here's the actual prompt structure I use. I'm giving you the whole thing, not a cleaned-up version, because the messy specificity is what makes it work.
PROMPT — Session Notes CLIENT CONTEXT: [Name/initial] is a [role] at a [company type]. They've been in this role for [duration]. We've been working together for [duration]. The current coaching arc is focused on [2-3 sentence description of the main themes]. Key ongoing threads: - [Thread 1: brief description] - [Thread 2: brief description] - [Thread 3: brief description] SESSION NOTES (raw): [Paste your shorthand, bullet points, voice memo transcript, whatever you captured during or immediately after the session. Don't clean it up. The messier and more specific, the better.] WHAT I NEED BACK: Write a session summary in my voice. I write in first person, present tense for observations and past tense for what happened. I'm direct and specific. I note what the client said that mattered, what shifted during the session, what I want to return to next time, and any patterns I'm starting to see. Keep it to one page. Don't use clinical language. Don't use the phrase "the client." Use their initial.
That last section, the "what I need back" part, is where most people don't go far enough. You need to tell it how you write. Not how a coach writes. How you write. The first few times, I went back to old notes I'd written by hand and pulled out patterns in my own language. I noticed I almost always end a note with a question I want to open the next session with. I noticed I use the word "noticed" a lot, and I tend to write observations as "there's something happening with..." rather than declarative statements. Those quirks are what make your notes yours.
Building the client file
The real shift happened when I stopped starting from scratch every session.
I keep a running context file for each client. After every session, the updated summary goes into the file. The file grows. So the next time I sit down to write notes, the AI already knows that R has been working on his relationship with his board since October, that we tried a role-play approach in session six that he found surprisingly useful, that his default under stress is to withdraw into data and stop making eye contact.
I ran my notes from last Thursday through Margaret and the draft that came back referenced a pattern I hadn't consciously named yet. R's language about his CTO had shifted from "I need to hold him accountable" to "I need to understand what he's reacting to." That's a significant shift. I caught it in the notes because the AI had the full arc and could surface the change. If I'd been writing by hand at 9 PM on a Thursday, exhausted, I would have missed it entirely.
That's not the AI being insightful. It's the AI being a good pattern-matching machine, fed enough context to actually match something. The insight was already in the sessions. I just needed something that could hold more history than my tired Thursday brain.
The ethics question
I want to talk about this directly because I know it's the reason some of you have tried this and stopped, or haven't tried it at all.
Coaches aren't bound by HIPAA the way therapists are. But we hold sensitive information about real people. Things they've told us in confidence. Things about their organizations, their teams, their personal lives. The first time I considered putting that into an AI tool, it felt wrong. It felt like a violation of something, even if I couldn't point to a specific rule.
Here's where I've landed, and I want to be clear that this is my line, not the line. Other coaches will draw it differently and that's their right.
I never use a client's full name. Initials only, and sometimes I change even those. I never include the company name. I describe the organization in general terms: "mid-size tech company," "family-owned manufacturing business." I never paste raw transcripts from sessions. What goes in is my notes about the session, not the client's words verbatim.
I use a paid API that doesn't train on my inputs. That matters. The free tier of most AI tools explicitly tells you they may use your conversations to improve the model. For coaching notes, that's not acceptable. A paid, no-training-data plan is the minimum.
And there are two clients I still write notes for entirely by hand. One is going through something so personally sensitive that I don't want any version of it in a digital system beyond my encrypted files. The other is someone in a public-facing role where even anonymized details could be identifiable if the context were specific enough. For those two, the old way is the right way.
The ten-minute workflow
Here's what my actual end-of-session process looks like now.
During the session, I jot shorthand in a notebook. Three or four words per thought. Enough to anchor the memory, not enough to pull my attention out of the room. On my walk between sessions (I take twenty minutes between every two, no exceptions, it's where my actual thinking happens), I pull out my phone and record a quick voice memo. Sixty to ninety seconds. Just the things I want to remember: what felt important, what shifted, what I want to come back to.
After my last session of the day, I sit down for ten minutes. I open the client's context file. I type or paste my shorthand and the voice memo notes. I send it through the prompt. What comes back is a first draft that's about 80% right. I read it, adjust the parts where the AI missed something or smoothed over something I wanted left rough, and save it.
Ten minutes. Sometimes twelve if it was a complex session. The notes are better than what I used to write in forty-five, because they're written while the session is still close, with a context file holding what my memory can't.
The first time I read back a set of these notes a week later, preparing for the next session, I actually stopped and reread a paragraph. It was more useful than anything I'd written by hand in years. Not because the AI wrote better. Because I'd captured more, sooner, with the full arc available, and the result had a texture that my exhausted Sunday notes never did.
What I haven't figured out yet
I'm still not great at the "what I need back" section for new clients. The first two or three sessions, before I really know someone, the notes come back flatter. I don't have enough context to give, and I don't yet know my own patterns with this person. For those early sessions, I tend to over-write the prompt, trying to compensate with instructions for what I can't yet provide in context. It's clunky. I'm still iterating.
I also haven't solved the voice memo problem. I want a faster way to get my between-session observations into text without manually transcribing them. There are tools for this. I haven't found one I trust with the content yet.
And some sessions resist being captured in notes at all. The ones where something cracked open that can't be reduced to a summary. I've learned to just write one line for those: "Session went somewhere important. Don't try to summarize. Be present to where it picks up." Sometimes the best note is the one that reminds you to pay attention, not the one that tells you what happened.
Those sessions are the reason I do this work. The AI handles everything around them so I show up ready for the next one.