The hardest part of an agentic AI programme is not the technology. The technology, given sufficient investment and engineering talent, is solvable. The hardest part is the workforce: the twenty thousand people whose daily work will be changed by the agents the programme deploys, most of whom were not consulted about the change, some of whom will find parts of their role automated away, and all of whom will be asked to trust a system they cannot fully understand. An agent programme that ignores this reality will succeed technically and fail institutionally.
The workforce change problem
The change management challenge for agentic AI is structurally different from previous enterprise technology waves. When a company deployed a new ERP, the change was to the interface and the workflow; the work itself remained recognisably human. When a company deployed a copilot for document drafting, the change was to the starting point of a task; the human still judged, edited, and owned the output. When a company deploys an agent that autonomously executes a multi-step business process — sourcing a supplier, resolving a customer complaint, reviewing a contract — the change is to the nature of the work: some of what was human work is now machine work, and the human's role shifts from doing to overseeing.
Oversight is not a lesser role; it is a different one. But it requires different skills — the ability to recognise when the agent is wrong, to understand the boundaries of its authority, to intervene effectively when it needs to stop — and different psychological relationships to the work. A customer service representative who spent ten years building expertise in resolving complex complaints does not automatically become a skilled agent supervisor; that transition requires deliberate investment in training and role redesign.
The communication architecture
The single most common change management failure in agentic AI programmes is the communication gap between the programme team and the workforce. The programme team knows what the agent does, what its limitations are, and why it was deployed. The workforce knows that an agent is now doing something it used to do, and nothing else. This information asymmetry produces the full spectrum of change-management pathologies: resistance driven by fear of job loss, over-trust driven by a misunderstanding of the agent's capabilities, and under-reporting of agent failures driven by a belief that admitting the agent was wrong reflects badly on the person supervising it.
The communication architecture for an agent deployment must answer four questions for every affected employee: What does this agent do? What does it not do, and where does it need my judgment? What happens when it makes a mistake, and what is my role in that? And — critically — what does my role look like in a world where this agent is running? The last question is the one most programmes avoid answering, because the honest answer is sometimes "your role is smaller," and programme teams are understandably reluctant to say that. The alternative — leaving the question unanswered — is always worse.
"People do not resist change. They resist being changed without being heard." — Kurt Lewin's insight, adapted for every enterprise transformation since, remains as true for agentic AI as for anything that preceded it.
Building trust in agents
Trust in an agent is not binary; it is calibrated. A workforce that trusts an agent appropriately — neither over-trusting it into accepting its outputs without scrutiny nor under-trusting it into overriding it on every action — is a workforce that has been trained to understand the agent's specific strengths and failure modes. This is not a general AI literacy programme; it is domain-specific training that tells a claims adjuster, for example, that this agent reliably extracts structured data from unstructured claims documents but consistently fails when a claim contains an unusual legal instrument it has not encountered in its training distribution.
The evals lead plays an important role here. The eval library is not only a governance instrument; it is a communication instrument. Publishing the eval results to the workforce — here is what this agent is tested on, here is its pass rate, here is the scenario where it most often fails — converts an opaque system into a legible one. Legibility does not eliminate the discomfort of working alongside an agent; it reduces the discomfort that comes from working alongside something unknown.
Union and works council engagement
In jurisdictions where employee representative bodies have legal consultation rights over significant changes to working conditions — which, in the European Union under the EU AI Act's workforce monitoring provisions, includes many agentic AI deployments — engagement with unions and works councils is not optional. It is a legal requirement. The programme team that discovers this after the deployment is live faces a significantly more difficult remediation than the programme team that structures the consultation process before the deployment begins.
Even where there is no legal requirement, early engagement with employee representatives produces better outcomes. Representatives often have a more accurate picture of how the work actually operates than the programme team's process documentation suggests, and their involvement in the deployment design process — not as gatekeepers but as subject-matter experts — consistently produces agent designs that are more durable in practice.
Reskilling as programme design
Reskilling is most effective when it is designed into the agent deployment, not bolted on afterward. An agent deployment that simultaneously creates a new oversight role, builds a training programme for that role, and provides a twelve-month transition pathway for the people moving into it is a change management programme. An agent deployment that deploys the agent and then asks HR to figure out the people implications is a technology project that will eventually become a change management problem.
The McKinsey State of AI research consistently finds that workforce investment is one of the strongest predictors of AI programme success — not AI literacy training in the abstract, but specific capability building tied to specific role changes. The programme team that designs reskilling as part of the deployment specification, rather than as a downstream HR activity, is the programme team that has understood what change management for agentic AI actually requires.