​Enterprise AI Needs A New Change Management Model

Trust Has To Work In Both Directions

The first challenge is that we generally see two types of users: people who over-trust AI and people who under-trust it.

Over-trust is already visible. People are sharing data they once might not even have sent through an API. Sensitive financial, HR or other classified information can get pushed into AI systems without enough thought about the implications.

The same problem exists with agents. If an agent is making a decision and a human is technically part of the escalation path, how carefully is that human actually reviewing the recommendation? Are they thoroughly reading it or simply skimming the lines and approving it notionally?

Under-trust is the opposite problem. A leader or employee may simply decide, “The agent cannot work for this use case,” when it actually could.

Sometimes the same person exhibits both behaviors depending on the use case. So the problem is ultimately about bringing the appropriate amount of trust into AI while maintaining the governance and guardrails required for each workflow.

Start With The Outcome And Route According To Risk

The Trust stage begins with ownership.

An agentic AI use case needs a sponsor or stakeholder who owns the workflow. That could be a CFO, CIO, line-of-business leader or process owner.

From there, the organization has to determine whether the use case is worth solving. What is the ROI? What are the unit economics? Does it make sense commercially? And is it technically feasible given the limitations agents still have today? Only then should an organization determine how the agent should behave.

That is the purpose of Route.

Unlike traditional software, an agent is not simply executing instructions and changing the state of data. It is mimicking human behavior or decision-making based on the signals available to it. Therefore, not every decision should be treated the same.

Invoice reconciliation is a good example. A small-value transaction might move through a faster path with significant agent augmentation. A high-value transaction or something with substantial compliance implications should be routed on a path with more validation and human review.

The point is to understand the impact of each decision and route it accordingly. Work should be classified by complexity and consequence so that human oversight becomes more intensive as the potential impact increases.

Adoption Starts While You Are Building

The next stage is Adoption.

We have all seen organizations subscribe to powerful AI tools only to discover that relatively few people are using them to their full extent. Agentic systems face the same challenge.

You should not build everything, run a massive classroom training session and expect everybody to start using the system on day one. Instead, users should become co-builders.

We involve process stakeholders throughout development. Validation does not happen only during a traditional UAT phase. It happens throughout the process. Process owners and execution teams contribute to what is being built, creating significantly more ownership.

Organizations can also develop local champions inside business units who understand the agentic workflow and help take it to other team members. Adoption, in other words, should happen through measured, simple, small packets of enablement throughout the value chain.

And adoption becomes much easier when it connects back to the outcome. If a user sees the system is saving several hours of work, they will invest more into using it. If a business unit sees clear ROI, it becomes much easier to justify the next investment.

Measure Business Outcomes, Not AI Activity

That brings us to Calibrate.

Because agentic systems are probabilistic by design, they have to be constantly measured and improved. But the important metrics are not necessarily tokens consumed or hours spent utilizing an AI system.

Of relevance is whether the workflow is producing the intended outcome. Broadly, we look at business value from the perspectives of efficiency, cost savings and revenue generation.

That could mean hours of human effort saved. It could mean enabling an outbound team to work with more potential customers. It could mean handling significantly more support volume without adding equivalent human capacity.

We are already seeing organizations set goals such as automating 40% to 60% of support interactions. In many cases, the goal is to scale the business itself, and AI allows the organization to address more volume with roughly the same capacity.

Usage alone does not tell you that. Business outcomes do.

Build For Continuous Evolution

Finally, agentic systems have to Evolve.

Enterprises need the ability to upgrade to newer models, change the underlying model entirely and modify behavior dynamically. They cannot become locked into hardbound code that makes adaptation difficult.

Humans also have to understand that these systems will constantly disrupt themselves. New models can create quality advantages, price advantages or both. Teams therefore need to continually calibrate what they have built and feed those learnings into the next evolution.

One workflow may also reveal multiple ancillary workflows that can be automated next. TRACE is designed so that the capabilities developed in one workflow can become reusable foundations rather than requiring every new project to start from scratch.

That is ultimately what sustainable enterprise AI adoption requires. Moving at AI speed does not mean abandoning governance in pursuit of velocity. It means building a system where velocity, measurable outcomes and trust reinforce one another, and where the enterprise can keep adapting as the technology beneath it continues to change.​

Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?