I spent years on client calls before any of this existed, and the thing that cost me deals was never the pitch. It was blanking on a name, or forgetting a promise I made three weeks earlier.
Short answer: the agent listens, you don't watch it. Set it up before the call so it actually knows the client, give it one narrow job during the call, and let it do the heavy work in the ten minutes after you hang up. If you're reading a screen while someone is talking to you, it isn't a co-pilot yet.
What should an AI co-pilot actually do during a client call?
Not talk. That part matters.
During the call the job is narrow: catch what you'd otherwise lose. The name of someone's business partner. A number you said you'd confirm. The one detail from the last call that's relevant right now. Good output is one line, glanceable in half a second.
Tools that handle the listening side well: Fathom, Granola, Otter, Fireflies. If you want to build against the meeting stream directly, Recall.ai hands you the transcript as it happens. Claude or ChatGPT handles the thinking.
What none of them do on their own is know your business. That's on you, and it's the whole difference between an agent and a note-taker.
I've been teaching this stuff since 1998 and running businesses since 1995, and I've watched a lot of people buy the tool and skip the setup. The setup is the product.
How do I prepare the agent before the call so it has real context?
This is context engineering for founders, and it takes me about four minutes.
Step one: build a client file. One row per client in Notion or Airtable. Names, roles, the side comments, open promises, deal stage. I have clients I've worked with for a decade, and I can't hold all of it in my head anymore. No reason to.
Step two: feed it the last two or three transcripts, not everything. Too much context makes the output worse, not better. The agent starts summarizing the summary.
Step three: write a five line brief. What I want out of this call. What I'm worried about. Two things I must not forget. Mine usually says something like "she mentioned her mother last time, ask" or "do not quote a number on this call."
The file structure and the prompts that connect all of it live in one place at https://a.mastermindshq.business/ai-os-book.
How do I keep it from making me look distracted?
One screen, one panel, glance only.
The rule I use: the agent surfaces at most two or three items for the whole call. If it's pinging you every thirty seconds, your threshold is wrong and you're now managing a tool while someone is trying to buy from you.
A few things that help:
- Turn off the scrolling live transcript. Reading is not listening, and your face shows it.
- Put it on a second device. A tablet flat on the desk, phone face down. Nothing should move in your eyeline.
- Rehearse once with a friend. Twenty minutes. You'll find out fast whether you can take a glance without losing the thread.
- Tell the client. "I've got an assistant taking notes so I can stay with you." I've never had anyone push back on that. Some states require consent to record, so ask. It's one sentence.
The founders who look distracted are the ones reading. The ones who look present are the ones who glance and come back.
What should the agent do after the call is over?
This is where the time comes back. The live piece is maybe ten percent of the value.
My flow looks like this:
- The transcript lands in a folder tied to that client.
- Claude drafts the follow-up email in my voice. I give it three emails I actually wrote as samples, not a description of my tone.
- n8n or Make pushes that draft into my Gmail drafts, creates the next three tasks in Notion, and moves the deal stage forward.
The line I don't cross: it never sends. I read the draft for ninety seconds, fix the two things it got wrong, and hit send. A transcript is not an outcome. A sent follow-up and three scheduled tasks is an outcome.
That's what an agentic workflow actually is. Not a chatbot answering questions. A chain of steps that finishes a job you used to finish at 11pm.
Where do founders get this wrong?
Three ways, and I've done all of them.
Treating it as a note-taker. A transcript you never read is just a recording. If nothing changes in your CRM or your task list, you bought a very expensive memory.
Skipping the context. Without the client file and the brief, you get a generic summary that could belong to any call with any company. That's the "my AI is busy, not useful" feeling people describe to me.
Letting it speak. Auto-replying to a client is how you have a bad week. The agent prepares, you talk. Keep that line.
One more: too many tools. If you're running four transcription apps, you don't have an AI agent operating system, you have a junk drawer.
FAQ
Do I have to tell clients I'm using an AI agent? Yes, at least one sentence. "I have an assistant taking notes" covers it. Recording consent laws vary by state, and honestly, most clients care less than you think. The ones who care usually prefer that you were present.
Do I need to code any of this? No. Fathom or Granola plus Claude plus Zapier gets you most of the way, and AI for founders and operators is mostly assembly at this point. Every piece is no code AI agents territory now. The technical part was never the hard part. The context is.
Can the agent answer client questions live? It can, and you shouldn't let it. The moment a client is talking to a bot instead of you, the call is over. Let it feed you the answer privately and say it in your own words.
The part that actually matters
I've trained more than 90,000 people at this point, and the pattern is always the same. The tool is easy. The system around the tool is what separates the people who get their time back from the people who just added a subscription.
I wrote the full playbook for this. How to Build Your Own AI Agent Operating System walks you through the exact architecture I use, step by step. You can get it at a.mastermindshq.business/ai-os-book.
And if you want to tell me about your worst client call story, reply and tell me. I read everything.