Agent, Automation, or Human? A Decision Boundary Matrix
The vendor pitch says put an agent on everything. That is how you get an expensive, probabilistic model doing work a five-line rule did perfectly, and an agent quietly owning a decision no one should have delegated. The fix is to stop labeling the whole process and start classifying each step.
Do not choose one tool for a process. Classify each step. Known path with structured inputs and stable rules goes to automation. Known objective with a variable path and a recoverable outcome goes to an agent. Anything irreversible, legally exposed, or non-delegable stays with a human. When the path is variable but the outcome is irreversible, split the step: the agent proposes, a human authorizes, automation executes. Two questions decide most of it. Is the outcome recoverable if it is wrong, and is the decision one you are actually allowed to delegate?
Most of the agent-versus-automation debate is miscast, because it argues about whole processes. Real processes are a sequence of steps, and the right answer usually changes from one step to the next. So I classify the step, not the process.
Here is the matrix I use.
| Step condition | Default actor |
|---|---|
| Known path, structured inputs, stable rules | Automation |
| Known objective, variable path, recoverable outcome | Agent |
| Irreversible, legally exposed, or non-delegable authority | Human |
| Variable path but irreversible outcome | Agent proposes, human authorizes, automation executes |
That last row is the one people skip, and it is where a lot of good designs live. The agent does the hard part, working out what to do when the path is not fixed. A human owns the moment of consequence. Automation does the mechanical execution once the human says go. You get the agent's flexibility without handing it an irreversible decision.
Two tests carry most of the weight
Recoverability. If the step is wrong, can you catch and undo it before real damage, and does that stay true at volume? A single reversible slip is fine. Ten thousand correlated slips you notice a week late are not. Recoverability is what separates the agent row from the human row.
Delegability. This is the one the vendor pitches ignore. Even if an agent can make the decision, are you allowed to delegate it? Some decisions carry authority that has to sit with a named person for legal, regulatory, or accountability reasons, regardless of how capable the model is. Capability is not permission. If the decision is non-delegable, it is a human step, full stop.
A note on cost
Running a probabilistic model on work a deterministic rule handles is a bad trade. You pay more per run, you add latency, and you introduce a class of quiet failure the rule never had. If the path is known and the rules are stable, an agent is the expensive, less reliable choice. Reach for the agent when the path actually varies, not to look current. The Pilot Trap essay covers the full cost of misaligning tool to task at production scale.
Are AI agents just glorified automation?
For the known-path steps, honestly, yes, and that is fine, because a rule is the better tool there anyway. The agent earns its place only on the steps where the path genuinely varies and the objective is clear but the route to it is not. If a vendor is selling you an agent for work that is really a stable rule, they are selling you cost and risk you do not need.
I made a version of this argument in a Diginomica piece on agentic AI as a non-uniform solution. The matrix here is the more precise version of that argument, with the hybrid row that the original piece did not have room to develop.
How to use this in practice
Map the process step by step. Put each step in a row of the matrix. Wire the steps together with explicit hand-offs, because the seam between an agent step and a human step is where these processes break in production. I cover what that failure looks like in the Pilot Trap.
Support triage is the everyday shape of it: automate intake and routing, let an agent draft the resolution, split the refund-over-threshold step so the agent proposes and a human authorizes, then drop back to automation to update the systems. One process, every row of the matrix, and not a single "put an agent on it" decision in sight.
For the governance layer that sits on top of this, the Augmented Marketing Decision Architecture applies the same three-question logic to marketing decisions specifically: reversibility, exposure, and frequency, sorted into AI-Primary, Collaborative, and Human-Primary zones.
When should I use an AI agent instead of workflow automation?
Use an AI agent when a step has a known objective but a variable path and a recoverable outcome, such as resolving an unstructured request where the route to the answer changes each time. Use workflow automation when the path is known, the inputs are structured, and the rules are stable, because a deterministic rule is cheaper, faster, and more reliable there. Classify each step of the process separately rather than picking one tool for the whole thing.
When should a human make the decision instead of an AI agent?
A human should own any step that is irreversible, legally exposed, or carries authority that cannot be delegated, regardless of whether an agent is technically capable of deciding. When the path is variable but the outcome is irreversible, split the step: the agent proposes, a human authorizes, and automation executes. The deciding questions are whether the outcome is recoverable if wrong and whether the decision is one you are actually permitted to delegate.
Are most AI agents just glorified workflow automation?
For steps with a known path and stable rules, effectively yes, and a deterministic rule is the better tool there. An agent earns its place only on steps where the objective is clear but the path genuinely varies. Putting an agent on stable, rules-based work adds cost, latency, and a new class of quiet failure for no benefit. The honest test is whether the path actually varies.