When not to use an agent
Agents suit some complex and variable work, not every process. The decision depends on uncertainty, reversibility, observability, the cost of error and the capacity for supervision.
When a technological capability becomes possible, there is a temptation to turn it into a universal answer. If an agent can use tools and complete tasks, why not give it any process?
Because autonomy and usefulness are not synonyms. An agent adds flexibility precisely because it makes decisions we have not encoded one by one. That flexibility is valuable when context varies and rules would be hard to maintain. It also introduces variability where we may already have had a predictable solution.
Four options before choosing
For any use case, compare at least four responses:
- Do not automate. The decision is infrequent, sensitive or depends on a human relationship that is part of the value.
- Assist. AI prepares information or a proposal, but a person drives every step.
- Automate with rules. The workflow is stable, structured and can be expressed deterministically.
- Delegate to an agent. The objective is bounded, but the path requires interpreting context, choosing tools and handling exceptions.
An agent is not the highest level of every option. It is a different architecture.
Five decision questions
How much uncertainty is there? If inputs and rules are stable, determinism usually wins. If the work involves natural language, dispersed information and variable paths, an agent may add value.
Is the outcome reversible? Classifying a request for review can be corrected. Executing a payment or deleting information requires much stronger controls.
Can we observe quality? If nobody can define a good outcome, we cannot evaluate the agent. Automation does not resolve ambiguity; it hides it at greater speed.
What is the cost of error? A small, detectable failure permits more autonomy. Legal, financial, reputational or physical harm demands boundaries, approvals or a different solution.
Is there capacity for supervision? Human approval is useless if the person receives too much volume, lacks context or routinely clicks accept. Supervision must also be designed.
Signs that an agent is the wrong choice
I would not use an agent when:
- A simple rule resolves the complete process.
- The action is irreversible and there is no reliable prior control.
- The required data cannot be used with authorisation.
- The system cannot record what it did or how it reached the outcome.
- The organisation has no owner for exceptions.
- The objective changes with every person and nobody agrees to define it.
- The benefit depends solely on removing a human interaction that the user values.
Start with less autonomy
There are many designs between doing nothing and delegating everything. An agent can propose a sequence without executing it, act only on test data, require approval for certain tools or work within cost and time limits.
Autonomy should be earned through evidence. First we observe recommendations, then reversible actions and finally—if quality and controls justify it—more complete processes.
The best system may be hybrid
A robust workflow combines rules for what is known, models for interpreting ambiguity and people for decisions that require accountability or contextual judgement. Designing those boundaries usually adds more value than trying to turn the entire process into a single agent.