Change Management For Agentic Data Modernization

Reimagining What Data-Driven Means

For years, businesses have used dashboards, reports and analytics to help people make decisions. With agentic AI, the data must also make sense to machines. Take an overdue invoice. An agent needs to know whether the customer has a payment extension or an ongoing dispute before taking action. But what if finance and sales interpret the same customer status differently? That requires business agreement, not just better technology. Data transparency now means understanding why an agent made a decision, not just where the information came from.

The Five Agentic Change Moves

I find the Prosci ADKAR Model useful for thinking about this transition. Awareness, Desire, Knowledge, Ability and Reinforcement are familiar concepts in change management, but agentic AI gives each of them a different practical meaning.

Awareness: Explain What Is Actually Changing

I would not start by telling employees how many tasks an agent can automate. I would start with what their working day might look like afterward. A finance professional who currently spends hours matching transactions may eventually spend more time investigating exceptions and reviewing decisions initiated by an agent.

That is a significant shift in responsibility. Leaders need to explain it clearly, including what employees will still own and where their judgment becomes more important.

Desire: Give People A Role In The Change

One reason employees may hesitate to adopt automation is that they understand parts of the process that the implementation team may have missed. A finance specialist knows when a payment reminder could create a customer problem. Someone in procurement may recognize a supplier transaction that needs an additional check.

These are not minor details. They can determine whether an automated action is appropriate. I would bring those employees into the design process and let their knowledge shape the agent’s operating boundaries. A culture of human-agent collaboration cannot develop simply because management introduces AI tools. People need to see that their experience continues to matter and that they have a say in how the work changes.

Knowledge: Teach People What To Delegate

Most AI training focuses on tools and prompts, but working with agents requires more judgment. Employees need to know what an agent can access, change and handle independently. Identifying duplicate invoices is one thing; changing a supplier’s banking details is another. Both involve automation, but the risks are different. Knowing when to delegate matters as much as knowing how to use AI.

Ability: Gradually Provide Autonomy

I’d prefer agents proposing actions, with employees making the ultimate decisions. This gives teams an opportunity to see where the recommendations work, where they fall short and which business rules need attention.

Once an agent demonstrates consistent performance, routine tasks with limited consequences can be automated more fully. This is where human in the loop (HITL) and human on the loop (HOTL) become relevant. HITL usually means people are getting support for their decisions. HOTL allows agents to operate within agreed boundaries while people monitor their performance and intervene when necessary.

We should not assume that asking for human approval every time is the safest option. NIST noted in August that excessive approval requests can lead to consent fatigue. When people are constantly asked to approve routine actions, there is a risk they stop reviewing those requests carefully.

The better approach is to match oversight to the potential consequences of each action.

Reinforcement: Find Out Whether The Work Improved

Once an agent is deployed, it is tempting to measure success by how many tasks it completes or how frequently employees use it.

I would pay more attention to what happens around the agent.

Are employees still repeating the same checks? Are exceptions handled faster? Can employees understand why an agent made a particular decision? By watching how people use the system in their day-to-day work, we get a better view of what is working and what still needs to be dealt with. Their feedback, especially if they do not agree with or overrule an agent recommendation, can often highlight problems that might have otherwise gone unnoticed. They may reveal problems that performance dashboards do not capture.​

My Personal Lesson: Nearly Removing Checks

While working on an invoice automation project, I noticed that the finance team was still checking transactions in Excel, even though reconciliation was working well. I went over a few cases with them and found out they were looking at special payment arrangements and exceptions we missed in development. We took out those checks, but we updated the workflow to include them.

That experience showed me that manual work is not always unnecessary. Sometimes it has business knowledge that automation can easily miss.

The Leadership Question Has Changed

The challenge is no longer whether AI agents can make decisions but which decisions we should trust them to make. Automation can do some things, but if you are dealing with money, customers or regulations, you need a human touch. Leaders need to make these choices based on business risk, not just technical capability.

Better data and capable agents are only part of the picture. What matters is how confidently people can work with them, knowing when to trust their actions and when to step in. For me, that is what change management in agentic data modernization comes down to.​

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