People keep asking me for the playbook. “Just tell me what to automate and I’ll do it.” As if there’s some clean checklist you can follow to hand your business over to AI agents and go sit on a beach.
I’ve been running my entire operation from my iPhone for six weeks now. AI agents handle my content pipeline, my job search, my SaaS product debugging, my outreach. I’ve written about all of it in this series. And the thing nobody talks about — the part that keeps tripping me up — isn’t the technology.
It’s figuring out what you absolutely should not delegate.
The stuff that works embarrassingly well
Let me give credit where it’s due. My Slack agent team handles morning digests, content drafts, job listing research, and basic triage across every project I’m running. Transkrybe signups, blog performance, LinkedIn engagement numbers — I wake up to a summary instead of forty tabs. That part is genuinely life-changing.
Scheduling and cross-posting content? Solved. My agents draft blog posts, generate LinkedIn variants, create carousel outlines, and schedule everything via APIs. I review and edit. The process that used to eat three hours of my week now takes about forty minutes. And the quality of the first drafts has gotten noticeably better as the agents learn my voice from previous edits.
Bug triage on Transkrybe is another win. When a user reports an issue, my agent pulls the logs, checks if anyone else hit the same problem, and gives me a diagnosis before I’ve even opened a Cowork session. Last week it correctly identified a rendering edge case in our PDF export that only showed up on scores longer than four pages. I didn’t write the fix — but I knew exactly what the fix needed to be before I asked for it.
Where I keep getting burned
Here’s the part that doesn’t make it into most AI productivity threads.
I let my outreach agent send a follow-up to a recruiter two weeks ago. The original message was good — personalized, referenced the company’s recent product launch, asked a smart question. The follow-up? Generic. “Just circling back on my previous message.” I caught it in my sent folder the next morning. That recruiter never responded. And I get it. I wouldn’t have responded to that either.
The pattern I keep seeing: AI agents are exceptional at first contact and terrible at nuance over time. The initial draft is sharp because it has clear context. But the follow-up requires reading the room. Did the person seem interested? Were they being polite but not really engaged? Did the tone of the conversation shift? Agents don’t track that. Not yet.
I also burned myself on a consulting proposal. My agent drafted a scope document based on a discovery call summary. It was technically correct — hit all the right deliverables, reasonable timeline, fair pricing. But it missed the client’s actual anxiety. They didn’t want a big scope. They wanted to start small and prove it worked. The proposal felt tone-deaf because the agent optimized for completeness instead of reading the emotional subtext of the conversation.
I rewrote it from scratch. Took me an hour. But we landed the engagement.
The delegation line I’m drawing now
After six weeks, here’s roughly where I’ve landed.
Fully delegated: morning digests, content scheduling, job listing research, log analysis, data aggregation, first-draft generation, cross-platform posting, basic email responses to people I don’t know yet.
Partially delegated (I review everything): blog posts, LinkedIn content, outreach messages, bug fix suggestions, carousel outlines, meeting prep summaries.
Not delegated at all: follow-ups with people I’m actively in conversation with, proposals for potential clients, any message where the relationship matters more than the information, responses to my son’s care team, creative direction for Transkrybe’s product roadmap.
See the pattern? Anything involving an ongoing relationship or emotional context stays with me. Anything that’s high-volume, pattern-based, or first-touch gets delegated. The gray zone — and this is where most of my mistakes happen — is when a task starts as delegatable and crosses into relationship territory without me noticing.
What the “delegation economy” actually feels like
I saw a bunch of articles this past week about the “delegation economy” — the idea that trusting AI agents is becoming a competitive advantage. And yeah, I’m living that. But the framing is wrong.
It’s not about trusting AI more. It’s about getting better at knowing where trust breaks down. Every week I find a new edge case where my agents did something technically correct but contextually wrong. And every time that happens, I get a little sharper at drawing the line before the mistake lands.
That’s the real skill. Not prompt engineering. Not building agent pipelines. Knowing when to take the wheel back.
Six weeks ago I was excited about everything I could hand off. Now I’m protective of the things I keep. The late-night recruiter email where the tone matters. The Transkrybe user who’s frustrated and needs to feel heard, not processed. The proposal that needs to reflect not just what the client asked for but what they’re actually worried about.
Those are mine. And I don’t think I want AI agents to learn how to do them. Not because they can’t eventually. But because those are the moments that make this a business instead of a machine.
The meta-lesson
If you’re building an AI-augmented workflow — and you should be — don’t start by asking “what can I automate?” Start by asking “what do I want to keep?” The answer tells you who you are as a professional. Everything else is leverage.
I’ll keep documenting how this evolves. Because six weeks from now, the line will probably be in a different place. And that’s kind of the point.
Want to run your own company this way?
I run my entire business with a team of AI agents. If you want to build the same setup, book an hour with me and I will walk you through exactly how it works on your own stack.
