I run about five client engagements at once, and roughly eleven AI agents run alongside me while I do it. Most of them live in a system I built on OpenClaw - my own agent setup, not a product I bought. The short version: AI handles the research, drafting, and operational grunt work across all five accounts, and I keep the judgment. That split is the whole game.
I have started at nine or ten companies over the years, and for most of that career the constraint was never ideas. It was hours. A fractional practice makes that brutal. Five clients means five sets of context, five sets of goals, five sets of people, and one me. Before AI, the only way to serve that many was to serve each one worse. Now I don't. The agents carry the load that used to eat my evenings, and I show up to every call actually prepared.
This is not a story about a magic tool that thinks for you. Nothing thinks for me. But a lot of things now fetch, draft, summarize, and remember for me, and that turns out to be most of the work.
How many AI agents does it take to run a fractional practice?
For me, eleven. Spread across five engagements, with a few shared ones that serve the whole practice. They live in the agent system I built on OpenClaw, which is the closest thing I have to a real answer to the hours problem.
The agents are not clever. That is the point. Each one has a narrow job and a clear spec. One pulls research on a market before a strategy session. One drafts the first version of a client update from my notes. One watches a shared inbox and flags the threads I actually need to answer. A couple are client-specific and carry the context of that account so I don't start cold every Monday. I wrote about the whole build in my OpenClaw setup guide, but the design principle is simple: an agent should do one thing I would otherwise do myself, and it should do it the same way every time.
Eleven sounds like a lot until you count what a fractional exec actually holds in their head across five companies. It is not a lot. It is barely enough.
An agent should do one thing I would otherwise do myself, and it should do it the same way every time.
What do I actually use AI for every day?
Three things, every single day: drafting, research, and operations. Those are the categories where the work is real but the judgment is low, and that is exactly where an agent earns its keep.
Drafting is the biggest one. I do not start documents from a blank page anymore. I dump a messy outline or a stream of voice notes into a drafting agent and tell it what shape I want - a strategy doc, a client update, a growth plan I can react to. It builds the scaffold. I bring the substance. The document that comes back is never the one I send, but it gets me from nothing to something I can argue with, and arguing with a draft is ten times faster than writing one.
Research is the second. Before a session with a client I have an agent pull the market context, the competitive picture, whatever public signal exists on the question we are chewing on. I still read it with a skeptical eye, because half the value is knowing which parts to throw away. But it means I walk in with a point of view instead of building one live in the room.
Operations is the quiet one. Logging what happened, tagging it, keeping track of which client I have not touched in two weeks, turning a pile of notes into next steps. None of that requires me. All of it used to cost me a Sunday. This is the same instinct I bring to a short engagement, where every hour you spend on admin is an hour you are not spending on the actual mandate.
How do I turn meeting transcripts into action?
I use Granola to capture the transcript, then run an action-extraction workflow over it. Granola listens during client and coaching calls and gives me a clean record without me typing a word. The workflow pulls the decisions and the open items out of that record and turns them into a short list I can actually act on.
The capture and the extraction are two different problems, and I kept them separate on purpose. Granola solves capture - I stay present in the conversation instead of scribbling, and I trust that the record is there. The extraction step solves the part that always used to slip: turning forty minutes of talk into three things that have to happen and who owns them. An agent reads the transcript and hands me that list. I edit it, because it usually gets the nuance of a commitment slightly wrong, and then it flows into wherever that client's work lives.
The compounding effect is memory. Across five clients, the thing that quietly kills a fractional relationship is forgetting what you agreed to three weeks ago. When every conversation leaves a searchable record and a clean set of actions, I stop dropping threads. That is worth more to my clients than any single insight I bring.
Where does AI fail in a real advisory practice?
It fails when you let it be confident. These models are agreeable to a fault - they will hand you a clean, well-structured answer that is quietly wrong, and the polish is exactly what makes it dangerous. So I treat every output as a draft and I verify anything that carries a number or a claim.
My habit is to ask the agent where it is most likely to be wrong before I trust anything it gave me. It is surprisingly good at naming its own weak spots when you make it. That one question has saved me from putting a bad figure in front of a client more than once. The pattern I have learned across every one of these tools is that speed and accuracy pull against each other, and the last twenty percent - the part where you decide what is actually true and what actually matters - is the part you cannot hand off.
The other failure mode is subtler. AI will happily do the wrong work beautifully. If I point an agent at the wrong question, it produces a gorgeous answer to a question no one asked, and the gloss makes me less likely to catch it. Direction is still mine. The agent supplies velocity, and velocity in the wrong direction is just a faster way to be wrong. I dug into this tension between speed and false confidence in a piece on what your AI-obsessed CPO is not telling you.
Here is how I split the work in practice, and why the middle column never bleeds into the right one:
| Work type | What the agents handle | What stays with me |
|---|---|---|
| Drafting | Scaffold a doc from my notes and voice memos | The argument, the point of view, what to cut |
| Research | Pull market and competitive context before a session | Which parts are real and which to throw out |
| Meeting notes | Capture the transcript and extract open items | Whether a commitment actually got made |
| Operations | Log interactions, tag threads, flag stale accounts | Which client actually needs me this week |
| Verification | Surface its own likely weak spots on request | The final call on what is true |
Where should you start if you run a portfolio of clients?
Start with one narrow agent that does a task you already hate, and get it working before you build anything fancy. The mistake is trying to design the whole system on day one. You cannot, and you will quit.
Pick the most repetitive, lowest-judgment thing on your plate - for me it was drafting client updates - and hand exactly that to an agent with a clear spec. Live with it for a week. Once the time it buys back is obvious, you will build the second one without me telling you to, and the third, until you have a practice that runs on rails you laid yourself. That is roughly the arc I followed on OpenClaw, and it is the same logic I bring to any fractional CPO engagement: prove the smallest useful thing, then compound from there.
Do not start with the tutorials and the prompt guides. You do not need them. You need one working automation that saves you a real hour this week, and the momentum takes care of the rest.
AI has not made me smarter. It has made my habits visible. If I were sloppy, it would help me be sloppy faster across five clients at once. Because I already know what good looks like in a product and growth practice, it gives me leverage that compounds every week - eleven agents clearing the noise so I can spend my hours on the one thing they cannot do, which is decide what actually matters.
