Know what AI should change inside your business.

AI looks like a person inside a computer. It talks like a person, answers questions, writes, plans, analyzes, and explains. The output feels human enough that it is easy to assume there is a mind behind it.

There is not.

At the most basic level, a language model predicts what should come next based on patterns it learned during training and the context it has right now. Then it predicts what should come next again. And again. It keeps doing that until it produces the response you see.

That prediction is incredibly useful. It can research a market, interpret a document, organize information, write software, compare options, and help work through difficult problems. But prediction is different from thought.

AI does not share your purpose. It does not understand the consequences of a bad business decision. It does not know which parts of your company matter unless you give it that context. The person using the system still decides what matters.

Think about a ball falling through a peg board. Drop a ball from the top of a peg board. Every time it hits a peg, it can move in different directions. You know the possible paths. You can understand the probability of where balls are likely to land. But you do not know exactly where one individual ball will end up until it gets there.

A language model works in a similar way. It looks at the context it has, predicts the next token, adds that token to the context, and predicts again. That is why adding context changes the answer. It is also why AI can solve a difficult problem one moment and make a basic mistake the next. The answer is being generated through probability. It is not being retrieved from a perfect internal record of truth.

Traditional software is usually deterministic. Two plus two equals four. Run that calculation again tomorrow and the answer is still four.

A quarter is probabilistic. You know it will land on heads or tails, but you do not know which one you will get on the next flip.

AI is probabilistic. Give a model the same request multiple times and you can get different wording, different reasoning, and sometimes different conclusions. That does not make AI useless or unreliable by definition. It means you have to build differently around it.

When the same input should always produce the same output, regular software is usually the better tool. AI becomes valuable when the work involves language, interpretation, uncertainty, incomplete information, or many possible correct answers. That distinction is going to matter throughout the rest of this workshop.

A language model generates a response. An agent puts that model inside a process.

Imagine I ask you to flip a quarter until you get heads three times. You do not know how many flips it will take. You flip the quarter, inspect the result, keep track of what happened, and decide whether you need to flip again. You continue until the condition is met.

An agent works in a similar way. The model is still predicting, but now the surrounding system gives it tools, memory, instructions, checks, and another chance to act. It can gather information, use a tool, inspect what happened, correct a mistake, and continue working toward a goal. Agents do not magically make AI think. They put prediction inside a workflow.

Imagine hiring someone who has read almost everything available in their field. They can explain difficult ideas, see connections quickly, and bring an enormous amount of general knowledge into the company. They have also never worked for you.

They do not know your customers. They do not know why your team follows one process instead of another. They have not lived through the mistakes that taught your experienced employees when the written rule should be ignored. And they can sound extremely confident while being completely wrong.

That is a useful way to think about AI. The model brings general knowledge. Your company brings experience.

Some of the most valuable knowledge in your company has never been written down. It lives in the salesperson who recognizes a bad lead within thirty seconds. It lives in the operations manager who knows which missing field does not matter and which one will create a problem three weeks later. It lives in the employee who knows that one customer always needs an extra phone call.

The model does not arrive knowing any of that. Your job is to combine what the model knows with what your company has learned. You are teaching the PhD graduate how your company actually works.

Think about your most experienced employee.

What do they know that a new hire would never find in the employee handbook?

Write a few examples down. That experience is part of the context an AI system needs before it can help with the same work.

You now have a workflow and a possible place for AI. Before building anything, decide whether improving that workflow actually matters. Start with the business result. Does solving the problem increase revenue? Does it lower costs? Does it remove a bottleneck that is preventing the company from doing either?

Put numbers around the opportunity using your actual business. A workflow that saves five minutes once a month probably does not deserve the same attention as one preventing your sales team from responding to qualified leads. Technology is not the reason to invest. The result is.

I usually look at revenue first because growth gives the business more room to hire, invest, and improve. Cost reduction eventually reaches a floor. The strongest opportunities do both. A faster lead response can increase conversion while reducing manual follow-up. A better internal knowledge system can reduce time spent searching while helping the same team serve more customers. Start with the business result. Then decide what deserves to be built.

AI makes work move faster. If the process works, that can create enormous leverage. If the process is broken, you create faster confusion.

Before putting AI into a workflow, ask whether another employee could follow the current process and reach a consistent result.

• Can they find the right information?
• Do they understand the steps?
• Do they know when the work should continue and when it should stop?
• Can they tell what a good result looks like?

If every employee gives you a different answer, you have a process problem before you have an AI opportunity. The process does not need to be perfect. It needs to be clear enough that another person can understand the job and how success is judged.

Think about how you train a new employee. You explain the job, then you show them what good looks like. You give them a completed proposal. You show them a good customer response. You show them a lead the company accepted and one it rejected. When they make a mistake, you correct it and explain why.

AI needs the same thing. Instructions tell the model what you want. Examples show it what you mean. Good examples should include successful work, common edge cases, and failures. The failures are especially valuable because they expose boundaries your experienced employees follow without consciously thinking about them. Without examples, the model has to guess what your company considers good.

A weak instruction says:

Write a proposal for this customer.

An experienced employee would need much more than that.

• Who is the customer?
• What problem are they trying to solve?
• What did they tell us during discovery?
• What services do we offer?
• What have we learned from similar customers?
• What should never be promised?
• What makes an opportunity a bad fit?
• What does a good proposal look like here?

That is context. The model does not know what you left inside your head. Better context does not make prediction disappear. It gives the model a better path through the problem. Think back to the peg board. You are still dropping the ball through a probabilistic system. Context changes the path available to it.

A useful instruction answers three things.

You do not need a collection of magical prompt tricks. The model needs to understand the job.

You already know the model is probabilistic. So do not ask it to be responsible for things regular software can verify. If a field is required, code can check whether it exists. If a dollar amount must fall within a defined limit, code can check the number. If a customer must meet a fixed qualification rule, code can enforce the rule. If an output needs to follow a specific format, validate the format before anything happens next.

Let AI handle the uncertain work. Let deterministic systems enforce what must be true. That is how you take advantage of prediction without making the entire workflow probabilistic.

Human review is part of the architecture. An internal research assistant that prepares a first draft for an employee carries different risk than an agent speaking directly to customers, touching money, making promises, or changing important records. As the consequences increase, so should the safeguards.

High-risk workflows need stronger tests, narrower authority, fallback paths, and clear rules for when the system stops and asks a person. Your employees already know where exceptions happen and which mistakes matter most. Use that judgment when you design the workflow.

And when a person corrects the system, keep the correction. Save the failed output. Record what should have happened. Preserve why the original result was wrong. The next version should begin with what the company already learned.

Go back to the workflow you mapped. First, define the business result you want to change. Raise revenue or lower costs? Then inspect the process itself.

• Could a new employee follow it?
• What examples would you give them?
• What context would they need?
• Which parts can be checked with a fixed rule?
• Where does a person need to remain responsible?

If you cannot answer those questions yet, do that work before you add AI.

Your employees already hold the experience the AI needs. They know your customers, exceptions, relationships, processes, and the consequences hidden behind the official rules. Removing those people also removes the experience required to build the system well.

The bigger opportunity is to give them more capability. AI handles more of the repetitive research, organization, translation, drafting, and preparation around the work. Your people spend more time applying judgment, helping customers, improving the process, and making the decisions that require them. The human stays at the center. AI makes them more capable.

A person can describe what they need in normal language. AI can translate that intent into forms other systems understand. That can mean research, analysis, database queries, software instructions, visual concepts, structured data, or the first draft of work that previously required specialized technical knowledge to begin.

That changes who gets to participate. A business leader does not need to become a programmer before helping define the internal tool the company needs. An employee does not need to become a data analyst before asking useful questions about company data. A person with twenty years of industry experience does not need to become a software engineer before helping build a solution to a problem they understand better than most engineers ever will. The expert stays close to the problem. AI helps translate their intent into action.

Think back to the employee from Part One. The person who knows things that were never written in the handbook. AI gives you a way to extend what that person knows without forcing them to personally perform every step forever. Their examples can improve the system. Their corrections can become context. Their judgment can define when an agent continues and when it stops. Their experience becomes more useful because more work can benefit from it.

That changes the question leadership should ask. Instead of starting with:

How many people can AI replace?

Start with:

How much more can these people accomplish?

A salesperson can spend less time researching an account and more time talking to the right customer. A provider can spend less time on administration and more time treating patients. An experienced operator can supervise more work without manually performing every step. A domain expert can help build software without waiting for someone else to fully understand the problem first. That is the opportunity.

More capability does not remove responsibility. AI can research more information, prepare more options, and carry more of the work. It still does not know what kind of company you want to build. It does not know which customer relationship matters enough to make an exception. It does not know what tradeoff you are willing to accept or what promise you refuse to make.

Those decisions belong to people. The system carries more of the work. People remain responsible for what the work means.

Think about the workflow you have been working through.

If the AI handles more of the research, preparation, organization, or execution, what does that allow the person doing the job to spend more time on?

Do not only measure the labor you remove. Look at the capacity you create.