If you’re weighing whether AI can really carry work inside your agency, you’re probably tired of hearing from people who’ve only done it in a demo. Fair enough. So here’s where it stands in mine, including the parts that aren’t working yet.

I run Marketing Heroes. A big share of our delivery runs through offshore seats, and when I started this there were six: two web developers, a designer, an ads specialist, a writer and an editor. All in, about $12,000 a month.

Two of those seats are off now. One operator and a team of AI agents are picking up that work. The rest come off the same way, one lane at a time, and I’m not going to pretend that’s finished.

Key Takeaways

• My agency is moving six offshore seats, about $12,000 a month, to one operator plus AI agents. Two seats are off so far.
• It goes one lane at a time. Nothing comes off until the replacement process has run and been checked.
• The weekly content cycle, lead verification and an AI assistant on inbox and calendar are working today.
• Training the operator takes longer than building the agents, and a person still reads everything before it ships.
• The lesson for other owners: measure first, move one lane, and keep a human approving every change until it’s proven.

Why I started with my own agency

I didn’t want to sell anybody something I hadn’t lived with.

It’s easy to build an impressive demo. It’s a lot harder to run real client work through a system every single week, with real deadlines and real clients who’ll notice if something’s off. If I was going to tell other agency owners this works, I needed to be the one finding out where it breaks.

So my agency is the test bed. Every lane we move, I’m doing first on our own work.

What’s working today

A few lanes are running every week now.

The weekly content cycle. A blog post, social posts for each platform and the graphics get planned, written, checked against the client’s voice and scheduled. The “what are we posting this week?” scramble is gone.

Lead verification. Every lead in a client’s CRM gets read and sorted: real buyers on one side, spam, sales pitches and wrong numbers on the other. Nothing gets re-tagged until a person approves it, and the lead counts we report are finally true.

An assistant on inbox and calendar. It sorts what needs me from what doesn’t and keeps my calendar from eating itself.

A voice agent on the phones and an SDR agent on follow-up. Calls get answered, and leads get worked until they book or say no.

What isn’t working yet

This is the part most AI case studies leave out.

Training the person takes longer than building the agents. The operator running this needs to know the system well enough to spot when it’s wrong. That’s weeks of learning, and you can’t skip it.

The quality check still misses things a person catches. Our automated review has let small errors through more than once. So a person still reads everything before it goes to a client. I’d rather say that out loud than claim the machine is perfect.

Some lanes are slower to move than I planned. Web builds are the next big one, and they’ve got more edge cases than content does. They come off when the replacement is proven, not when the calendar says so.

How we move a seat

Every lane follows the same steps.

First, we clock the work the old way: how long each task takes and how often it happens. Then we build the replacement process, starting from what the team already does. We run it on a small slice of the work while the old way keeps running alongside it. A person approves every change. When the numbers hold up for long enough, that lane moves over for good.

It’s slow on purpose. Moving fast and breaking things is a terrible plan when the things are client deliverables.

What I’d tell another agency owner

Don’t start with the whole agency. Start with one lane where the work repeats every week and you can measure it.

Get your team on one platform with a real business account before you build anything. Write down which client data can go where. Then clock the job, build the replacement, pilot it and keep a human in the loop until it’s earned trust.

And be honest with your team about what’s happening. The work coming off their plate is the point. The people who stay get to do the parts of the job a machine can’t.

Frequently Asked Questions

Which seat should come off first?

The one where the work is the most repeatable and the easiest to check. For us that was content and reporting work, where a weekly rhythm and a clear standard already existed. Judgment-heavy work like strategy comes much later, if ever.

Do you still need people?

Yes. Someone has to run the system, catch what it gets wrong and talk to clients. What changes is how much each person can carry, and how much of their week goes to work a machine could do.

How long does moving one lane take?

Weeks, not a weekend. Building the agents is the quick part. Clocking the work, running the new process alongside the old one and training the person who runs it is what takes the time. Anyone telling you it’s a weekend project hasn’t done it on client work.

If you want to see it running

I take the Head of AI seat inside other agencies now, and I start by showing owners what’s running in mine, warts included. You can read more about me or see how the lane-by-lane loop works.

If you’d rather just talk it through, book a 20-minute fit call.