When a health clinic here in Austin hired me to help implement AI, I didn’t build anything on the first day.
I showed up with a microphone and a notepad. Asked for permission to record. And then I spent the entire day watching.
Their weekly team meeting. Patient intake. How staff communicated with each other between appointments. The scheduling process. What happened when something went wrong.
Just observing. No solutions. No recommendations. Just watching how the business actually worked.
Why Most AI Consulting Starts in the Wrong Place
The typical consulting pattern goes something like this: you get on a call with a client, they explain their problems, you propose solutions, you build things.
The problem is that the problem description is almost always incomplete — and not because clients are withholding information. It’s because people describe their work the way they think it should work, not the way it actually does.
When you ask someone “what are your biggest time sinks?” they’ll tell you about the things they’ve consciously identified as problems. They won’t tell you about the 15-minute daily workaround they’ve been doing for three years without thinking about it. They won’t mention the verbal system their team uses to track tasks because the real system broke and nobody replaced it. They won’t describe the friction that’s become invisible because it’s been there so long.
Observation surfaces all of that. You see things the client can’t see precisely because they’re too close to their own work.
At the health clinic, one of the things I noticed early in the day was how their weekly team meeting ran. The doctor would delegate tasks to different staff members verbally. There was no central system for tracking those tasks. No automated follow-up. No accountability loop. Action items were assigned in the meeting and then tracked entirely in people’s memories.
This would never have come up in a discovery call. Nobody would have listed “our verbal task assignment system” as a problem. It just was how things worked.
The Microphone as Research Tool
The thing that makes the observe-first approach practical — and not just an expensive way to spend a day — is the recording device.
I use a small microphone I can wear or clip to a surface. I record everything throughout the day with the client’s permission.
At the end of the observation day, I don’t have notes. I have a full transcript of hours of actual work happening in real time: meetings, workflow discussions, problem-solving conversations, interruptions, questions from staff, explanations of processes.
Then I feed all of that through AI.
The AI processes the full transcript, identifies patterns, surfaces recurring themes, and helps me generate a structured implementation proposal: here are the workflows worth automating, ranked by frequency and time savings, with specific recommendations for each.
The clinic got a proposal the following week that was grounded in their actual workflows — not in how they thought their workflows worked, but in what I’d seen and recorded.
What the Data Shows That People Don’t
The key principle in agent building: automate the tasks that happen daily or weekly, not the rare impressive ones. High-frequency pain creates compounding return.
This is exactly why observation matters. The tasks people consciously identify as problems tend to be the big visible ones. But where time actually goes is in the small, daily, repeated frictions. The verbal task assignment with no tracking. The 15-minute email thread that happens every Monday. The three-step manual process everyone does every single day without realizing there’s a pattern.
You don’t find those by asking. You find them by watching.
How to Do This Yourself
If you’re consulting on AI for clients — or if you want to apply this to your own business — the observation framework is simple:
Before your first working session: Ask for permission to record. Explain that you’re there to understand how work actually happens, not to pitch solutions.
During the observation: Sit in on any meetings that happen that day. Ask people to walk you through their actual workflow in real time, not just describe it. Note every time you see something manual and repeated.
After the day: Transcribe your recordings. Feed the transcripts through AI with this prompt: “Identify the 5 most time-consuming repetitive tasks in this transcript, ranked by frequency.” Use that analysis as the foundation for your proposal.
The proposal you generate from actual observation data will be more specific, more useful, and more likely to get approved than any proposal generated from a 45-minute discovery call.
The observe-first methodology is one of several frameworks I cover in the 4-Day AI Sprint — a structured program for building practical AI workflows for your business.
Recommended for you
Want the full system? 25X is the flagship productivity system we teach.
Explore 25X →