Why experts in the middle make your AI agents worse

Who is better at the job: the expert with the technical skills, or the user with the business context? I asked myself this question for weeks. The answer hurt my ego as a builder, and it changed how I design business systems.

The person with the most context should make the decision, and your job is to give them a tool and guardrails to do it. When we let our clients write their own AI instructions, they outperformed our experts. Experts in the middle add delay and distort the signal from the market. Context beats competence.

I thought my value was my ability to control the machine. I thought clients paid me to be the Wizard of Oz, who pulls levers behind a curtain. I was wrong. The best business system is not a team of wizards. It is a machine so simple that the user can pull the levers.

This is how we stopped managing AI for our clients and became architects. Here is what we learned in three phases.

Phase 1: The spaghetti

We did not start with a clean SaaS platform. In the months before February 2026, Ninjabot launched a product called AI Salesman, an AI agent that handles sales conversations. To test product-market fit, we used n8n, a low-code workflow automation tool. It worked, but it was ugly. We hard-coded business logic into a visual spaghetti of nodes.

This was Level 1 of what I call the Autonomy Spectrum: manual execution. At this level, our developers and I were essential. I was the bottleneck, and I knew it was only the first phase. Many founders get addicted to this phase because it feels like hard work. But it is not a business. It is a high-stress job with overhead.

We proved the concept: the AI could perform as well as an average sales rep. But the operating cost was too high. We spent 80% of our time on maintenance of the wires and only 20% on improvements to the product.

Phase 2: The internal tool

As we grew to dozens of clients, the manual n8n workflows became impossible to manage. We needed to delegate. So we built an internal app to manage the implementations. In its dashboard, we could update vector databases, change temperature settings and prompts, and adjust conversation flows without raw code.

This was Level 2: delegated execution. We sold AI Salesman as a managed service. The client paid a premium retainer, and we gave white-glove support. We had SOPs and a team of experts who managed the technology, so the client did not have to.

But we had walked into a trap. As the experts in the middle, we had created a game of Telephone. Every piece of market feedback went from the client, to the account manager, to the engineer, and then to the AI. The information decayed at every hop. We had the competence but not the context. We optimized for clean code, while the client needed to win in messy commerce.

Phase 3: Clients took the controls

To grow further, we could not hire one account manager for every 10 new clients. The economics did not work, so we removed the middleman. We took our internal tools, gave them a user-friendly design, and gave the logins to the clients. Each client got a text box to write their own system instructions. A system instruction is the standing text that tells an AI agent how to behave in every conversation.

This was Level 3: infinite execution. We were scared. We thought that without our prompt engineering skills, clients would break the agents. We prepared our support team for a flood of "I broke it" tickets.

Instead, the clients started to outperform us. A floor heating company wrote a system instruction that did not follow our prompt engineering best practice. But their AI agent booked sales meetings at a higher rate than the one we had fine-tuned for weeks. (If you want to see what a plain, working instruction looks like, read how to write boring prompts that work.)

The floor heating company won because it knows the details of floor heating inquiries and how its typical clients talk. They know the context. We know the solution. With the right tool, they put their 20 years of experience directly into the system, and our team's expertise did not dilute it.

The three levels of the Autonomy Spectrum

The Autonomy Spectrum describes who does the work and how close the decision sits to the market. Most founders think delegation means hiring people to do tasks. That view is too narrow. Real delegation moves control and intelligence closer to the reality of the market.

The psychology of detachment

The move to Level 3 has an almost Stoic side. Level 2 feeds the ego, because it feels good to be the expert. Level 3 requires humility. You must admit that you are not the smartest person in the room about how people apply your tool. You build the hammer. The carpenter knows better how to swing it.

Code is the ultimate Stoic employee. It has no opinion. It does not try to prove that it knows better than the user. It executes the logic it gets. When you build a Level 3 system, you practice the Stoic discipline of detachment: you separate your personal effort from the result.

How to apply context over competence in your own team

The same rule applies to your internal teams, not only to SaaS products or AI agents. Count the decisions in your company that need a manager's approval. A refund over $50. A blog post. A small change to a sales script.

Each approval step pushes your company back to Level 2. It says: "I trust my judgment (far from the problem) more than your judgment (facing the problem)." A Stoic operator builds Level 3 systems for employees too:

1. List the approval steps. Find each decision that waits for a manager, such as the refund over $50.
2. Write the guardrail. Turn the decision into a rule in the system. For example: "Any support agent can issue a refund up to $100 without approval. Above $100, the system blocks it."
3. Give the tool to the frontline. The person who faces the problem decides at once, inside the rule. You have replaced a manager with a rule.

Guardrails matter even more when an AI agent does the work. Before you give an agent access to live systems, read 5 rules before an AI agent touches your business. To pick the right automation pattern for each rule, use the Automation Architecture Guide.

From operator to architect

I do not want to oversimplify. We did not fire everyone and leave our clients alone. Some clients still want us to manage the whole system, and we still sell onboarding packages. But our role has changed. We no longer hoard the magic like wizards. We are the architects. We build the house, check the wiring, and hand over the keys. Then we say: "You know how to live here better than we do. But if the roof leaks, we are here."

The best delegation is not to find someone to do the work for you. It is to build a system that lets the person with the most context do the work, with no friction.

If you run an agency, a shop or a tech company, ask yourself one question. Do your tools make you feel essential, or do your systems make your users powerful? The first feeds your ego and takes your time. The second grows your revenue and gives your freedom back.

Build the interface. Trust the frontline. Let the code do the heavy lifting.

Build a calmer business,Yuri

Frequently asked questions

Why can a client's own AI agent prompt outperform a prompt written by an AI expert?

The client knows the context: the specific questions their customers ask and how to talk to them. At Ninjabot, a floor heating company wrote its own system instruction for AI Salesman. It did not follow prompt engineering best practice, but it booked sales meetings at a higher rate than the prompt the Ninjabot team had fine-tuned for weeks.

What are the three levels of the Autonomy Spectrum in delegation?

Level 1 is manual execution: you do the work, and you run out of hours. Level 2 is delegated execution: employees do the work through internal tools, SOPs and approval chains, and the signal from the market decays at every hop. Level 3 is infinite execution: the system and its user do the work, with automations, AI guardrails and direct feedback loops.

Why does a managed service model for AI agents stop scaling?

A managed service puts experts between the client and the AI. Every piece of feedback travels from the client to an account manager, to an engineer, and then to the AI, and information decays at every hop. Costs also grow with revenue: you need a new account manager for every group of new clients.

How do you replace a manager's approval with a guardrail?

Write the rule into the system instead of routing the decision to a manager. For example: any support agent can issue a refund up to $100 without approval, and the system blocks any refund above $100. The person who faces the problem decides at once, and the rule sets the limit.