Part 4 of the six-part series From Course Factory to Capability Function.
Previously in this series: Stop Treating Every Performance Problem as a Training Problem
Every L&D team is now being asked some version of the same question: what should AI do, and what should stay with people? Most teams answer it tool by tool, which produces a scattered result: a chatbot here, an auto-generated module there, no coherent view of the work.
Answer it work by work instead. Take the inventory from Step 0 (installment 2): the plain-language list of what your team actually does. Then put each activity through one decision: who or what should do this part?
Five categories cover nearly everything.
Machines do the repeatable and verifiable. Report generation, scheduling, tagging content, routing requests, compiling completion data. The work has clear inputs, clear outputs, and a way to check the result. If no check exists, build one before you automate.
AI copilots assist the judgment-heavy work. Drafting a first version of content, summarizing research, generating practice scenarios, suggesting a diagnosis. The AI proposes; a person decides. The rule is simple: AI never ships to a learner without a human verifying it. The verification step is part of the workflow, not an afterthought.
AI agents handle bounded multi-step work with clear guardrails. A new-hire onboarding sequence that adapts to role and progress. A coaching practice environment with defined scenarios. Agents work when the boundaries are explicit: what the agent may do, what it must escalate, and where the audit trail lives. If you cannot write those three down, the work is not ready for an agent.
Humans own what depends on judgment and trust. Diagnosing a performance problem with a defensive stakeholder. Deciding what good looks like in ambiguous work. Handling anything where being wrong carries real cost. These do not get automated. They get protected, because they are the work that justifies the function’s existence.
Escalated humans catch what the other four push up. Every agent and automation workflow needs a named person who handles the exceptions: the edge cases, the failures, the situations the guardrails did not see coming. Escalation is a designed role, not a hope. If no one owns the escalations, the workflow is not finished.
Two mistakes show up everywhere. The first is automating the visible work and ignoring the invisible kind: the relationship management and judgment calls that never appear in a process map. The second is letting the tool decide the allocation. A vendor will tell you their platform can do everything. Your inventory tells you what should be done by whom. Start from the work, not the product.
Revisit the allocation twice a year. Capabilities change fast; an activity that needed human judgment last year may be verifiable today. The decision is never finished. But it should always be deliberate, written down, and owned by someone.
Questions this article answers
How should a leader decide whether a task belongs with AI or a person?
Break the role into specific tasks. Check what each task requires, how the result can be verified, and what happens when it fails. Allocate the work only after you have identified a person responsible for exceptions. Try the work allocation decision tool on one task.
Who is accountable when an AI agent encounters an exception?
A named human owner must be able to handle exceptions, challenge the output, and decide when to stop or escalate. The AI Work Allocation Review also examines accountability and the verification plan.
