I Kept Rebuilding the Same Agent for Every Client. Then I Changed the Model.

When I was creating visual assets for Padel Society — renderings of the facility’s different rooms for investor presentations — we hit a consistency problem early.

Each room needed to feel like the same building. Same lighting. Same materials. Same visual signature across the lobby, courts, locker rooms, retail area. But when we started each rendering from scratch, there was no guarantee of that.

So we built a prompt library.

Save every prompt that works. Document the lighting description. Document the material palette. When you move to the next room, start from the foundation you already built — not from zero. After a few rooms, the process was fast, and the results were consistent.

The principle seemed obvious: don’t keep reinventing what you’ve already figured out.

I was doing the opposite in my AI consulting work.

The Old Model (And Why It Has a Ceiling)

For the first year of doing AI implementation work, I operated the way most consultants do.

Client hires me to build something — an investor-targeting agent, a pre-call briefing system, a CRM automation pipeline. I build it. It works. I hand it over. They own the code. They own the system. Engagement ends.

Then another client comes with a nearly identical problem.

I build the same thing again. From scratch.

I worked with a medical office brokerage in Austin — Seth and Steve at Transwestern, cold-calling building owners every day. They needed a morning call sheet that pulled property data before each call: occupancy numbers, Google Maps links, recent lease activity, pre-sorted by priority. Built the agent. It ran well. I handed it over and moved on.

A few months later, another brokerage came to me with basically the same workflow request. Different city, different property type, same fundamental structure.

I rebuilt it.

At some point I started asking: why do I keep giving this away?

The Shift: Keep the IP, License the Access

Now when I build an agent for a client, I keep the intellectual property.

The client gets access to it — not ownership. They can use it. They can’t resell it, reverse-engineer it, or hand it to a competitor. When the engagement ends, the agent stays in my library.

The contracts need to be structured correctly: IP retention clause, license terms, a non-compete that’s time-limited (usually one to two years) and industry-specific rather than a blanket restriction. This isn’t complicated legal work. A good attorney can draft it in an afternoon.

But what changes is everything downstream.

How It Works in Practice

Lucas Siegel, co-founder of Yuna (AI mental health platform, 50,000 users across 155 countries, $30M valuation), asked me on a call whether I’d build him a sales prospecting agent — something similar to an investor-targeting pipeline I’d already built for another project.

My answer: “I might already have that built. You’d just license it from me.”

Pay per 100 introductions, or a flat monthly fee. No new build on my end. He gets access to something that’s already been designed, tested, and refined in a real deployment.

That’s the model.

Every client engagement is now an investment in a library rather than a one-time delivery. The first client pays for the full build — customized to their workflow, their data sources, their specific use case. I keep the IP. If the next client needs something similar, I’m adapting an existing agent, which takes a fraction of the original time.

The Compound Effect

Custom consulting has a fixed ceiling. Every client costs the same amount of your time. You can’t serve 10 clients as efficiently as you serve 1, because each one requires a full custom build from scratch.

The agent library breaks that relationship.

Second client with a similar use case doesn’t pay for a new build — they pay for access to what already exists. Third client, same thing. The marginal cost of serving each new client decreases as the library grows.

Think about it like the difference between a contractor and a software company. The contractor’s revenue scales with hours worked. The software company’s revenue scales with users — the build cost doesn’t multiply with each new customer. The agent library isn’t pure software, but it moves the model in that direction.

There’s still customization, still client relationship work, still implementation. But the underlying assets compound in a way that time-for-money never will.

The Question Worth Asking

If you’re doing any kind of implementation or consulting work right now — AI or otherwise — here’s a question: after every engagement ends, what do you have left?

If the answer is “a good client relationship and a case study,” that’s something. But you’ve also handed over the actual work product. The thing you spent the most time building.

The agent library model starts from a different assumption: the thing you build has value beyond this specific client. Document it. Retain it. Offer the next client access instead of a rebuild.

Over time, you’re not selling custom work. You’re selling access to a growing collection of tested, working systems. Every engagement makes the collection more complete.

That’s a different kind of business than what most consultants are building.

Thanh Pham is the founder of Asian Efficiency and runs the Two Hour Workday program. He helps business owners and consultants use AI to build systems that work when they’re not in the room.