​Agentic AI Is Exposing The Process Behind The Process

This voice experience is generated by AI. Learn more.This voice experience is generated by AI. Learn more.By Masa Maruyama, CEO, Saison Technology International.

gettyAs I talk with companies about agentic AI, I have noticed that the conversation often begins with a process.

A company wants an AI agent to process an order, resolve a customer request or decide what to do when something does not go as expected. On paper, the process may look straightforward. There is a workflow, systems that support it and people responsible for each step.

But once the conversation gets more specific, something interesting happens.​

Someone will say, “Normally, we do this, unless…”

The interesting part is what comes after “unless.”

Most established companies have documented processes. They have workflows, policies, approval structures and systems that have been refined over many years.

But documentation does not always capture everything that makes a process work.

People interpret information. They recognize unusual circumstances. They know when an exception is reasonable and when it should be escalated. An experienced employee may notice that something looks correct in a system but does not make sense given what is happening with a particular customer or supplier.

Over time, this judgment becomes part of how the organization operates. Some of it is written down. Much of it becomes part of how people do their jobs.

As a result, a gap can emerge between the process a company thinks it has and the process it actually runs.

For years, that gap could remain relatively invisible because people fill it naturally. Agentic AI makes it easier to see.

Traditional automation worked particularly well when the rules were clear: If this happens, do that. Move information from one system to another. Trigger an approval when a threshold is reached.

Generative AI expanded what AI could do with information, from answering questions to creating and summarizing content. In many of those uses, however, a person still decides what happens next.

Agentic AI is being asked to go further. Companies are exploring agents that can interpret information and take actions across business systems. Doing that well requires context, the surrounding knowledge, history and judgment that tells a person (or an agent) what a situation actually calls for, not just what the documented process says to do.

Consider something as common as processing a customer order. The documented process might say that an order can proceed when inventory is available, the customer is within the appropriate credit limit, and the required approvals are in place.

An experienced employee may know there is more to it.

A long-standing customer may have a temporary issue that should be reviewed rather than automatically rejected. An unusually large order may deserve another look even though it meets the formal criteria. A situation that appears routine may need to be handled differently because of something the team knows about the customer or the order.

None of this means the documented process is wrong. It means the real process contains judgment.

People can make these adjustments without necessarily thinking of them as rules. They may pause before taking the next step—checking on something or asking a colleague. Over time, these small decisions simply become part of how the work gets done.

An AI agent cannot reliably derive all of that from the workflow alone.

Trying to capture every exception and convert every piece of institutional knowledge into a rule is unrealistic. It may not be desirable either.

When companies begin defining what an agent should do, they have to look more closely at how people actually make decisions along the way. Practices that once happened naturally suddenly have to be examined.

But making those practices visible raises another question: Should the agent learn to do the same thing?

Some practices may reflect valuable business judgment that the company depends on. Others may be workarounds for outdated policies or habits that made sense years ago but have never been revisited.

That distinction matters. Instructing an agent to reproduce how work gets done today could mean passing along not only what works well, but also practices that should have changed long ago.

Before instructing an agent to follow an existing practice, companies should ask why that practice exists.

Much of the discussion around agentic AI focuses on how much work agents will eventually be able to perform. But deciding what an agent should do can also force companies to look more closely at how the work is being done today.

What comes after “unless” can reveal a great deal about how a process actually works.

Companies can teach an agent the existing process. The question is how much context that process depends on, and whether the agent should learn to hold that context.​​

Forbes Business Council is the foremost growth and networking organization for business owners and leaders. Do I qualify?