Delegating a task to AI changes the work of the person responsible for it.
An assistant can help me prepare a meeting brief. I bring the question, review the response, ask for revisions, and decide how to use it. An agentic workflow can take on more of the sequence: find relevant material, check several sources, assemble the brief, place it where I need it, and report what it could not complete.
That sounds like a difference in convenience. For a leader, it is also a difference in responsibility. Someone now has to specify what counts as completion, which sources are appropriate, what may be changed, and what should happen when the expected path breaks.
The names are not reliable enough to answer those questions. Products described as assistants may use tools and perform actions. Systems called agents may follow tightly prescribed workflows. I care more about the delegation: what can this system do without a person directing every step?
Anthropic's Building Effective Agents makes a useful architectural distinction between predefined workflows and agents that dynamically direct their processes and tool use. Its advice to begin with the simplest adequate approach is relevant here. A fixed sequence of known rules may be better served by conventional automation. More autonomy needs a reason. Source
My practical test starts with the work itself. Does it mainly require an answer, interpretation, or creative exchange that a person will use? An assistant may be sufficient. Does it require pursuing a defined outcome through several steps, responding to what happens along the way? An agentic approach may be worth examining.
Then comes the harder question: can we define a boundary within which that pursuit is useful and acceptable?
Consider a proposed intake workflow. A request arrives with a missing business owner and an ambiguous description. The system could identify the missing field, compare the request with the existing catalog, and prepare a clarification for the requestor. Those are distinct responsibilities. Approving the underlying investment is another responsibility entirely.
In a course design exercise, I explored this kind of separation for AI intake and triage. The useful design question was where routine coordination could be reduced while preserving human judgment for substantive risk and investment decisions. It was a proposed design, not evidence that a deployed agent had achieved the hoped-for result.
The exercise also makes a broader point: a process rarely belongs wholly to either a person or an agent. Responsibility can differ from step to step. A system may retrieve information independently, propose a classification, prepare a record for review, and stop before making a commitment.
Before it acts, it needs explicit permissions. Reading a record, editing a record, sending a message, and approving a request are separate capabilities. Access should be limited to the task, and consequential boundaries should be enforced through the surrounding system. A sentence asking an agent to be careful cannot carry the entire control burden.
It also needs a usable definition of quality. A meeting brief should identify its sources, separate confirmed facts from inference, and expose missing information. An intake assessment should apply the right criteria and route uncertain cases appropriately. Fluency does not establish that either job was done well.
Testing should include incomplete inputs, conflicting records, unavailable tools, duplicate requests, and plausible attempts to redirect the system outside its purpose. The question is how it behaves when conditions depart from the demonstration. A useful agent should sometimes stop, ask, or return a partial result with its limits made clear.
That creates work for people. Someone has to receive exceptions, understand them, and resolve them within the time the workflow requires. If every unusual case accumulates in an unattended queue, the organization has automated the easy steps and left the process incomplete.
Human review needs design as well. Reviewers need the evidence, context, time, and authority to challenge the recommendation. A person clicking approve under pressure may add very little protection. For important decisions, it can be useful to compare an independent human assessment with the AI recommendation before assuming agreement demonstrates quality.
Completion must also be visible. Producing a draft does not establish that it was saved. Preparing an update does not establish that the receiving system accepted it. A reliable workflow distinguishes attempted actions from confirmed results and reports where the chain ended.
Finally, delegation should change deliberately over time. A system may start by preparing work for approval. If evidence supports it, a specific class of action can later be permitted within defined limits. New conditions, recurring failures, or broken controls may justify reducing that authority again.
For leaders, this makes the assistant-or-agent question concrete. Choose one task and describe what the person currently does between receiving it and declaring it complete. Which parts are ready to be delegated, and which unresolved judgments would you otherwise be asking the system to make for you?