ADKAR’s Knowledge phase is about knowing how to change. What do I need to learn to succeed in the new way of working?

For traditional change initiatives, that’s straightforward: learn the new software, the new process, the new role. For AI adoption, it’s trickier. The obvious answer—“learn to use the AI tools”—masks a deeper question: what knowledge actually matters when agents do the execution?

What to Deprioritize

Syntax and boilerplate.
When agents generate code, the marginal value of memorizing language syntax drops. That doesn’t mean developers shouldn’t know the language—they need to read, evaluate, and debug. But drill-and-practice on syntax is the wrong focus. Knowledge that was core (how to write a loop, how to structure a component) becomes less central when the agent does the first draft.

Mechanical process steps.
”Follow these 12 steps to complete the workflow” is knowledge that agents can encode. What humans need is judgment: when to deviate, when to double-check, when to escalate. Training that emphasizes rote procedure wastes time. Training that emphasizes decision points and edge cases builds the right knowledge.

Tool-specific minutiae.
Which button to click, which menu to open—that knowledge is fragile. UIs change. Tools get replaced. Focus knowledge on concepts: what the tool does, when to use it, what good output looks like. The mechanics can be discovered.

What to Prioritize

1. Prompt design and iteration.
How do you get good output from an agent? Clarity, context, examples, constraints. People need to learn how to compose effective prompts, how to iterate when output is off, and how to structure multi-turn interactions. This is learnable, and it’s high leverage.

2. Evaluation and validation.
How do you know when agent output is good enough? When is it wrong? When is it subtly wrong? Knowledge here includes: what to look for, how to test, how to spot hallucinations and logical errors. This is judgment, not procedure—and it’s essential.

3. Orchestration and handoff.
When do you use the agent? When do you step in? When do you hand off to another agent or a human? Knowledge of the overall workflow—the agentic pattern—matters more than knowledge of any single tool. People need mental models of delegation, supervision, and handoff.

4. Failure modes and recovery.
What happens when the agent gets stuck? When it produces garbage? When it’s confident but wrong? Knowledge of failure modes and recovery paths reduces anxiety and improves outcomes. “When X happens, try Y. If that doesn’t work, do Z.”

Knowledge as Judgment, Not Procedure

The shift is from knowledge-as-memorization to knowledge-as-judgment. We’re not teaching people to do what the agent does. We’re teaching them to work with the agent: direct it, evaluate it, correct it, and own the outcome. That’s a different curriculum—and one that stays relevant as the tools evolve.