Building Your First AI Agent: Start With Expertise, Not Technology

In my last article, I introduced a simple idea: AI should not replace human expertise. It should help preserve and extend it.

That idea is the foundation of the AGENT Framework.

Now comes the next question I hear from almost everyone: How do I actually build an AI agent?

Most people immediately begin thinking about software, platforms, models, and technical integrations.

I do not. Before building an AI agent, you first need to understand whose expertise you are trying to capture—and what makes that expertise valuable.

Start With the Person Everyone Trusts

Think about the smartest person in your organization.

That may not be the CEO or the person with the highest-ranking title.

It is often the person everyone calls when they have a difficult question.

It might be:

• The compliance officer who knows the regulations inside and out

• The coder who can explain why one modifier is correct and another is not

• The billing manager who understands why a claim was denied

• The physician who applies twenty years of clinical judgment to a complex case

• The operations leader who knows why a process works the way it does

• The experienced employee who remembers the history behind an important policy

That person’s knowledge is incredibly valuable.

Your goal is not to replace them.

Your goal is to capture how they think so their expertise can become more accessible, consistent, and scalable across the organization.

Step One: Define the Agent’s Role

Every effective AI agent begins with a clearly defined role.

Before selecting a technology platform, ask:

• What job will this AI agent perform?

• Who will use it?

• What specific problems should it help solve?

• What information should it be allowed to access?

• What actions should it never take?

• Which decisions must always remain with a human?

Broad descriptions usually lead to broad—and often unreliable—results.

For example, instead of saying:

Build me a compliance AI.

Define the role more precisely:

Build an AI assistant that helps coders locate relevant Medicare documentation requirements, summarizes the applicable guidance, and reminds users that final coding decisions require qualified human review.

That is an entirely different level of clarity. The second description defines the user, the purpose, the information the agent should retrieve, and the boundary of its authority. A well-defined AI role should make it clear not only what the agent does, but also what it does not do.

Step Two: Capture the Knowledge Experts Use

AI does not become useful simply because it has access to the internet. It becomes useful when it has access to the right knowledge.

Start by identifying and collecting the information your experts rely on every day. This may include:

• Policies and procedures

• Regulatory guidance

• Coding references

• Training materials

• Standard operating procedures

• Frequently asked questions

• Audit findings

• Corrective action plans

• Lessons learned

• Internal best practices

• Decision trees and escalation pathways

• Examples of common errors and how to correct them

These resources become the knowledge foundation for your AI agent. However, collecting documents is only the beginning. Organizations must also determine whether those documents are current, accurate, complete, and approved for use. An AI agent built on outdated or conflicting information can make errors more quickly and consistently than a person ever could. Trusted AI requires trusted knowledge.

Step Three: Teach the Agent How Experts Think

Knowledge alone is not enough.

Experience matters.

Experts do not simply locate a regulation and repeat it.

They interpret the guidance in context.

They recognize exceptions.

They identify missing information.

They ask follow-up questions.

They compare the current situation with past cases.

They understand risk.

They know when an issue should be escalated.

These decision-making patterns are often more valuable than the documents themselves.

To begin capturing them, ask your experts questions such as:

• What do you look for first when reviewing this issue?

• What information is commonly missing?

• Which details would change your recommendation?

• What warning signs indicate higher risk?

• What exceptions should users know about?

• When would you stop and ask for additional documentation?

• When should the issue be escalated to another expert?

• What mistakes do inexperienced employees make most often?

The answers help transform information into usable expertise. This is where an AI agent begins to reflect not only what an expert knows, but also how that expert reasons.

Step Four: Design the Right Human Checkpoints

One of the biggest mistakes organizations make is assuming that an AI agent should make the final decision. In many situations, it should not. AI should support decisions, not automatically assume accountability for them. Human oversight protects quality, reduces risk, and builds trust. It is especially important when an AI-assisted decision could affect patient care, regulatory compliance, reimbursement, employment, privacy, or other high-impact outcomes.

Before deploying an AI agent, define:

• Which outputs require human review

• Who is qualified to perform that review

• How users should verify the information provided

• When the agent must disclose uncertainty

• When an issue must be escalated

• How corrections and feedback will be documented

• Who remains accountable for the final decision

The best AI systems do not eliminate experts.

They make experts more effective.

Step Five: Start With a Narrow, Valuable Use Case

Your first AI agent does not need to solve every problem in the organization.

In fact, it should not. Start with one clearly defined use case that is valuable, repeatable, and manageable.

A good first use case often has several characteristics:

• Employees ask the same questions repeatedly

• The answers rely on a defined set of trusted resources

• The task consumes meaningful staff time

• The current process is inconsistent or difficult to navigate

• The output can be reviewed by a qualified person

• The risks and limitations can be clearly defined

Examples might include:

• Helping employees locate an internal policy

• Summarizing applicable documentation requirements

• Guiding staff through a standardized intake process

• Identifying missing information before human review

• Answering routine questions based on approved training materials

• Directing users to the appropriate escalation pathway

Starting narrowly allows the organization to test the quality of the knowledge, evaluate the agent’s responses, identify risks, and improve the process before expanding.

Your Assignment This Week

Choose one expert role in your organization.

Then answer the following questions:

• What questions does this person answer every day?

• What documents and resources do they use?

• What knowledge exists only in their head?

• What patterns do they recognize that others may miss?

• What mistakes do new employees commonly make?

• Which decisions require their professional judgment?

• What issues do they immediately escalate?

• If this person retired tomorrow, what expertise would your organization lose?

Write down your answers. Then select one recurring question or process that could become the foundation for a narrowly defined AI assistant. Congratulations.

You have just taken the first step toward building an AI agent based on expertise instead of technology.

Looking Ahead

In the next edition of The AI Readiness Journal, we will explore one of the biggest reasons AI initiatives fail: Poor knowledge management.

We will examine why organizing and governing information is often more important than selecting the right AI platform—and how trusted knowledge becomes the fuel that powers trustworthy AI.

Because AI does not create expertise. It amplifies the expertise you choose to share.