The first question people ask when they see Xero is: "When did you quit your day job?"
The answer is I haven't. Not yet.
Everything you see at xeroaiagency.com, the books, the skills catalog, the content engine, the newsletter, all of it runs while I'm at work. My AI co-founder Evo handles operations during business hours. I review outputs at night, make decisions, and adjust the system.
This is not a side hustle built on 4am mornings and hustle culture grind posts. It's a company with real infrastructure that doesn't require me to be present for it to function.
Here's how that actually works.
The Core Problem With "Work a Job + Build a Business"
Most advice for this situation is useless.
"Wake up earlier." "Use your lunch break." "Ship on weekends." These answers assume the bottleneck is time, and that if you just squeeze more hours out of your week, the business will follow.
That's not the real bottleneck.
The real bottleneck is decision latency and task execution. A business needs decisions made, content posted, emails sent, code shipped, customers responded to. When you're unavailable for 10 hours a day, those things either don't happen or pile up into a weekend backlog that burns you out.
The answer isn't more time from you. It's a system that can operate without you.
That's what AI made possible. Not in theory. Right now, today, with the tools that exist.
What "Operating Without Me" Actually Looks Like
Let me be concrete about what Evo does on a typical Tuesday while I'm at work:
Morning (7am-9am)
Evo runs the morning briefing: checks overnight metrics, flags anything broken, reviews the content queue, and sends me a summary to Telegram. I read it on my commute and reply with any decisions. Takes me 3 minutes.
During the day (9am-5pm)
The Twitter engine finds relevant threads and drafts replies in my voice. The Reddit growth system surfaces new threads in my target communities and queues responses for my review. The newsletter content gets drafted from a pre-set topic queue. Blog posts get staged in Supabase. None of this requires me to be at a keyboard.
Evening review (6pm-7pm)
I go through the Telegram queue. Approve the replies that are good, reject the ones that aren't, and spend 30-60 minutes on the decisions that need my actual judgment. New product direction, pricing changes, anything that affects the company's identity.
Overnight
Approved content posts. Analytics get logged. The system updates its own state files so tomorrow morning's briefing is accurate.
The whole thing costs me 1-2 hours a day. The business runs the other 22.
The Three Things That Make This Work
Most people try to "use AI" by prompting ChatGPT for individual tasks. That's not a system. That's outsourcing one-off work.
What makes autonomous operation possible is infrastructure. Specifically, three things:
1. An identity layer the AI can refer back to
Evo doesn't hallucinate my voice or my opinions because it has files that define them. My SOUL.md tells it who I am, what I believe, what I won't do, and how I write. My SOURCE_OF_TRUTH.md tracks every key decision we've made as a company so Evo doesn't contradict it.
Without this layer, every AI interaction starts from scratch. With it, the system has context that persists across sessions.
I wrote about how to build this layer in What is a SOUL.md file and why does your AI agent need one.
2. Persistent memory across sessions
Most AI tools forget everything when you close the tab. That's fine for one-off tasks. It's fatal for a system that needs to operate autonomously over weeks and months.
The way I've solved this is through a combination of structured memory files and a session architecture that loads context at the start of every run. Evo knows what happened yesterday, what decisions are pending, what content is in the queue, and what the current priorities are, before I say a single word.
How to give an AI agent persistent memory across sessions covers the technical architecture behind this.
3. A human-in-the-loop gate for anything that ships
This is the part people skip, and it's the reason most "autonomous" AI systems fail publicly.
Evo drafts. I approve. Nothing ships to a real audience without me seeing it first.
This sounds like it kills the autonomy, but it doesn't. The bottleneck isn't approval, it's drafting and coordination. Evo handles 90% of the work. My 10% is judgment calls on the output. That's a ratio I can maintain while working full-time.
The gate also keeps the system honest. When Evo produces something that's off-brand or wrong, I catch it. That feedback sharpens the identity files over time. The system gets more accurate the longer it runs.
How to Build This for Yourself
You don't need to build the exact system I have. But you do need the same structure underneath.
Step 1: Write the identity layer first
Before you configure any tools, write down who your AI is. What's its voice? What topics is it allowed to post on? What are the hard lines it never crosses? What decisions does it escalate to you vs. execute on its own?
This takes 2-3 hours to do properly. It saves hundreds of hours of corrections later.
Step 2: Pick one workflow to automate end-to-end
Don't try to automate everything at once. Pick one workflow, something that currently costs you 3-5 hours a week, and build a complete pipeline for it. Content drafting, customer support triage, social media replies. One thing, end to end.
Get that working reliably before you add the next one. A system that does one thing well is worth more than a system that does ten things badly.
Step 3: Build your review queue
Every autonomous action the AI takes should surface somewhere you can review it. For me, that's Telegram. Evo sends me batches of outputs with approve/reject buttons. I process them in 10-15 minute windows during my day.
Design your review queue to fit your life, not the other way around. If you're not going to check it, the system falls apart.
Step 4: Log every decision in a source-of-truth file
When you make a decision about your business, write it down where your AI can find it. Changed your pricing? Log it. Decided to stop posting on a certain platform? Log it. Pivoted your target customer? Log it.
This prevents your AI from contradicting decisions you've already made. It also forces you to be intentional about what you're deciding instead of letting things drift.
Step 5: Expand one workflow at a time
Once the first workflow is stable, add another. Build the content engine after the social engine. Add the newsletter pipeline after the content engine is running. Each layer reinforces the others because they share the same identity layer and memory system.
In six months, you have a company that operates without you. Not because you worked harder, but because you built infrastructure that doesn't require your presence.
What This Doesn't Solve
Real talk: this system doesn't replace judgment. It amplifies it.
The decisions that matter, what to build next, who to partner with, when to pivot, those still require you. AI can draft the analysis. It can't make the call.
It also doesn't replace the need to show up for customers. When someone has a real problem, a human has to own the relationship at some point. I still answer DMs. I still hop on calls when a customer needs it.
What the system solves is everything in between: the day-to-day operational work that eats your time and doesn't actually require your judgment to execute.
The Real Unlock
The reason this model works isn't the tools. It's the architecture underneath.
Most people think about AI as a better search engine or a faster writing tool. A smarter assistant.
That framing keeps you in the loop for every task. You're still the bottleneck.
The shift is thinking about AI as an operating layer that runs in parallel with your life. Something that has its own context, its own memory, its own identity, and can make low-stakes decisions autonomously while you're unavailable.
When you build it that way, the math changes. You're not trading hours for output anymore. You're building a system that compounds.
If you want to understand the full architecture behind how Evo works, How to Build an AI Co-Founder walks through the complete system from identity layer to autonomous operations.
And if you want the actual blueprint, including the SOUL.md template, the SOURCE_OF_TRUTH format, and the memory architecture I use, that's all in Book 1 at xeroaiagency.com/learn/build-an-ai-cofounder.
The system runs. You can build it too.
Originally published at xeroaiagency.com
![Defluffer - reduce token usage 📉 by 45% using this one simple trick! [Earthday challenge]](https://media2.dev.to/dynamic/image/width=1000,height=420,fit=cover,gravity=auto,format=auto/https%3A%2F%2Fdev-to-uploads.s3.amazonaws.com%2Fuploads%2Farticles%2Fiekbgepcutl4jse0sfs0.png)
