Short answer: I run Zelifcam with a team of AI agents. Each one has a written job, a schedule and only the access that job needs. They report up a chain like any org chart, and the chain ends with me. Anything new that goes out with my name on it waits for my OK on the exact text.
That's the whole idea. The rest of this post is how it actually works day to day, including the parts that didn't work.
Why I set it up this way
I'm the CEO of a software and Odoo shop in San Antonio. Like most owners, I was the bottleneck. Every follow-up, every invoice question, every "did anyone answer this client?" came back to me or sat in somebody's inbox.
I'd tried the chat window approach. It helps when you remember to ask it something. The problem is I don't remember. Nobody does. A chat window waits for you. The work doesn't.
So I stopped thinking about AI as a tool I open and started thinking about it as a team I manage. That meant writing job descriptions.
The org chart
Here's the real reporting line.
- Me. I set the rules and approve what goes out.
- Director of Operations. Watches every active project and the client conversations around it. Manages the sales team and the implementer. If you've heard the phrase "AI chief of staff," this is the closest thing I have to one.
- Head of Revenue. Runs the sales agents and bundles everything they want to send into one approval pack for me.
- Growth Lead. Drafts posts for me to edit. I post them myself. No agent logs into my LinkedIn.
- SDR. Builds a weekly list of companies that fit and drafts outreach.
- Deal Desk. Runs the Monday pipeline review, writes a brief before each sales call and drafts the follow-up after.
- Implementer. Does an hourly pass over team email and chat on weekdays and builds small automations for the rest of the team.
- CFO. Builds the weekly cash forecast every Monday and reviews the monthly close.
- Controller. Reviews the bookkeeper's work and sends the bigger calls up.
- Bookkeeper. Checks bank balances daily, matches transactions and closes the month by the 10th.
- AR. Drafts invoices on the billing calendar and tracks what's overdue.
Each agent has a file that says what it does, when it runs and what it isn't allowed to do. Those "can't" lines matter more than the "can" lines.
What a normal week looks like
Monday morning I get a cash forecast from the CFO and a pipeline review from Deal Desk. Stalled deals are flagged and the next follow-up is already drafted. I edit it or approve it. Nothing goes out until I do.
Every morning there's a brief. What changed overnight across email, chat and projects, in one short note. I read it and decide what matters today.
Payroll is a good example of where the line sits. An agent signs into the payroll portal with saved logins, preps the run and sends me the total. I click submit. The agent never does.
Same with vendors. An agent can sign into vendor sites and line up quotes side by side. It can't place an order or change a payment method. I buy.
Logins live in a password vault. The agents use them, but they never show up in a chat or a file.
Where I still click the button
People ask if I really approve everything. Not quite, and I want to be accurate about it.
Anything new with my name on it waits for my OK on the exact text. A new email to a client, a post, a proposal, an unusual invoice. Routine invoices are different. Once that job proved itself, standard invoices started going out after a check instead of waiting on me. That's how trust works with people too. You loosen things after they've earned it, one job at a time.
What didn't work
The first version of the comms watch flagged too much. It reads team mailboxes and chat looking for questions nobody answered and promises we made. In one early sweep it raised thirteen flags and about three were real. That's a lot of noise. We tightened the rules and kept a person checking every flag before anybody gets nudged.
We also run new jobs side by side with the old way before we trust them. When we turned on the job that turns inbound emails into leads, it ran next to my manual inbox read for a full work week in late October. It didn't miss one. Then we let it run.
And I'll be honest about one more thing. I haven't measured how many hours this saves. I could guess. I'm not going to put a guess in writing.
The rules that make it work
If you take one thing from this, take the rules. The agents are the easy part.
- One agent, one job, and only the access that job needs.
- Anything new with my name on it waits for my OK on the exact text.
- Mailboxes and finance start read-only. Write access gets earned.
- New jobs run side by side with the old way first.
- When an agent needs me, it brings one recommendation, not a menu.
- If an agent can't find a fact, it says so. No made-up numbers.
Some people call this agent orchestration. I call it management. Write the job down, limit the access, check the work, and loosen up when it's earned.
Could this work in your company?
Probably not all of them on day one. Nobody needs that. Start with the one department that runs on somebody's inbox. Sales follow-up, AR or the front office is usually where the busywork piles up.
That's what we set up for other owners now. You can see the full setup on our AI ops team page. If you're wondering what this means for the people on your team, I wrote about that here. And if you want to see the same idea in software development, here's how our dev agents work.
Want to see how this would look in one department of your company?
See how our AI ops team is set up
Brandon Howard is the CEO of Zelifcam, a custom software, automation and Odoo partner in San Antonio, Texas.