Traditional change management assumed a bounded transformation with a defined end state. Agentic AI has no end state. Here is what CIOs and CTOs need to rethink before the next deployment.

Tim Keary published a piece in Forbes yesterday on PwC's agentic scaffolding launch. He spoke with Rima Safari, a partner in PwC's US data, analytics and AI practice, and with Chai Atreya, CPO and CTO of ActiveCampaign and a former member of the Amazon Alexa team. I contributed a perspective on change management. What follows is my longer argument, which the format of a news article does not have space for.

What 95% Failure Looks Like from the Inside

Most enterprise AI pilots fail not because the technology does not work. They fail because the organization around the technology was not built to absorb continuous change. The tools get deployed. The workflows do not follow. The return on investment never materializes.

This pattern is not new. It is the same pattern that derailed enterprise resource planning rollouts in the 1990s, customer relationship management deployments in the 2000s, and digital transformation programs in the 2010s. The technology arrived before the organization was ready to use it. What is different now is the speed of the cycle and the nature of what is being deployed.

Earlier transformation waves had a finish line. You implemented a system, went live, entered stabilization. The project had a defined scope and a defined end. Change management existed to get people across that finish line.

Agentic AI has no finish line.

The Structural Flaw in Every Change Management Playbook

The traditional change management model was built on three assumptions. First, that leadership can define a target state before the rollout begins. Second, that resistance is a temporary condition that recedes as familiarity grows. Third, that there is a point at which the change is complete and normal operations resume.

All three assumptions break down with agentic AI. An agent deployed in underwriting today will behave differently in six months because it has processed more cases, encountered more exceptions, and been retrained on outcomes. The target state after deployment is not the same target state that was designed before deployment. The gap between what was planned and what is running widens continuously.

That is not a bug. That is the point. An agent that does not learn is just expensive automation. But an agent that learns means the humans working alongside it have to keep learning too, and that learning cannot be scheduled into a training calendar.

From Programs to Capabilities: Atreya's Framing Is the Right One

Chai Atreya makes a distinction in Keary's piece that deserves to be pulled out and examined separately. He argues that AI is forcing enterprises to move from program-based transformations to capability-based transformations. Under the old model, a leader built a transformation program, defined a roadmap, rolled out the system, and increased adoption. That approach is, in his words, too slow and too brittle for the AI era.

He is right, and the reason connects directly to the change management problem. A program has a budget, a timeline, and a sponsor. When the program ends, the capability it was supposed to build either exists or it does not. Agentic AI does not work that way because the capability keeps changing shape after the program closes. You cannot fund a transformation program for something that has no end state.

What Atreya describes as the requirement, reusable capabilities, context layers, governance models, agent infrastructure, and evaluation loops, is essentially a description of an operating model, not a project plan. That distinction matters more than it sounds. Operating models are funded and staffed on an ongoing basis. Programs are funded to close.

What PwC's Scaffolding Approach Gets Right, and Where It Stops

Rima Safari describes agentic scaffolding as a tool for identifying the bottlenecks that prevent scaling and for supporting the evolution of roles that agentic deployment requires. The application layer lets teams visually simulate workflows, including steps, handoffs, exceptions, validators, evidence requirements, and data flows, before writing production code. The simulation can be stress tested. Governance controls for each agent get refined before go-live rather than after incident.

This is a meaningful shift from the historical pattern of deploying first and discovering failure modes in production. PwC reports that a Fortune 500 insurer using the tool cut quote-to-bill time by 50% to 80% depending on case complexity. That outcome requires not just a better tool but a workforce that was prepared for what the tool would change in their daily work.

Where the scaffolding conversation gets incomplete is the implicit assumption that visual simulation is a one-time event before launch. In agentic deployments the simulation needs to be a recurring discipline. The agents evolve. The simulation environment should evolve with them. Safari's point about role evolution is correct but understates the frequency at which that evolution will need to happen.

Change Management Has to Become the Operating System, Not the Launch Checklist

The enterprises making real progress on agentic AI have reorganized around a different principle. Change management is not what happens after the implementation is ready. It is the infrastructure the implementation is built on top of.

In practice this means three concrete shifts:

Continuous role redefinition rather than retraining events. When an agent takes over intake and quoting, the underwriter's job changes. Not once, but incrementally over months. Organizations that handle this well have standing processes for redefining roles in response to agent capability changes. They do not wait for the annual review cycle.

Governance built for drift, not just for go-live. Most AI governance frameworks are designed around the launch decision: is this safe to deploy? The harder question is whether the agent running today is behaving consistently with the agent that was approved six months ago. That requires monitoring infrastructure and human review processes that did not exist in traditional change programs.

Cross-functional ownership from day one. Atreya's point that product, engineering, design, data, legal, security, and business teams need to work together around a clear outcome is well made. Agentic workflows do not respect departmental boundaries. Change management programs owned by IT and handed to the business at go-live fail because the business was never at the table when the agent's behavior was being shaped.

The compressed timelines are not a temporary condition of the current AI race. They are the new operating environment.

The Question CIOs Are Not Asking Often Enough

Most AI readiness assessments ask: do we have the data, the infrastructure, and the governance to deploy agents? The harder question is whether the organization has the capacity to keep up with what those agents will become after deployment.

A well-built agentic system will outpace a static workforce. The technology advantage erodes if the people operating alongside the agents cannot adapt at the same rate the agents learn. That is not a technology problem. It is an organizational design problem, and it requires a different answer than any change management playbook written before 2024.