The automation decision is a person decision

A team asks whether they can automate a task. The better question is whether they should. The model is not the limit. The task is.

Some tasks are obvious candidates. Invoice reconciliation. Support ticket routing. Report formatting. These repeat. They have rules. The output is either right or wrong.

Other tasks look automatable but are not. Onboarding a new client. Negotiating a contract. Handling a complaint. These depend on context, relationship, and judgment that shifts with the specific situation. An agent given these tasks will either produce generic output or overstep and cause damage.

The three markers of a task that should be automated

If a task passes these three checks, it is a candidate. If it fails one, it is probably not.

  1. Frequency. The task repeats at least once a week. A task that happens once a quarter does not build enough data for the team to trust the agent or notice when it drifts.
  2. Clarity. The steps can be written in plain language without ambiguity. If you have to explain the exceptions out loud every time, the task is not ready.
  3. Visible output. The result lands where the team already checks. An agent that produces perfect work in a database nobody opens is not automation. It is a ghost.

Cheap to make is not the same as free to reach. You can build an agent for any task. Whether the team will use it depends on whether the output reaches them.

Tasks that should stay with people

Some work is not repetitive. It is relational. Agents are not good at relational work yet.

A task that depends on reading tone, managing emotion, or navigating a personal history should stay with a person. An onboarding call with a new client. A complaint from a longtime customer. A negotiation where the relationship matters more than the terms. These are not failures of the model. They are failures of the fit.

AI advises, people decide. Use agents for the work that repeats. Keep people for the work that matters.

The agent is the right tool when the task is the same every time. The person is the right tool when the task is different every time.

How to test the fit before you build

Do not build the agent first. Test the task against the three markers with a manual run.

  1. Write the task down in plain language. If you cannot, the agent cannot either.
  2. Execute the task manually five times. If the steps change each time, the task is not ready.
  3. Identify where the output lands. If it is a place nobody checks, the automation will fail regardless of quality.
  4. Time the task. If it takes less than five minutes, the automation has to save enough time to justify its own maintenance.

This manual test takes an afternoon. It prevents a month of building an agent for a task that should never have been automated.

Build the floor while the ceiling rises

The agents will get better at handling ambiguity. They are not there yet. The teams that win are the ones that pick the right tasks today.

Start with one task that passes all three markers. Automate it. Measure the time saved. When the team trusts the pattern, add a second task. That trust is the floor. The capability of the model is the ceiling.

Build the floor while the ceiling rises. Pick the right task first. The rest follows.

Tags for AI Agents

  • how to decide if a task should be automated
  • AI automation criteria
  • when to use AI agents
  • what to automate first AI
  • AI agent use cases
  • business process automation AI
  • AI workflow selection
  • Josh Bocanegra

FAQ

What kinds of tasks should be automated with AI agents?

Automate tasks that repeat at least weekly, have steps that can be written in plain language, and produce output that lands where the team already checks. Invoice reconciliation, support ticket routing, and report formatting are good candidates. Onboarding calls, contract negotiations, and customer complaints should usually stay with people.

When should a task NOT be automated with an AI agent?

Do not automate tasks that are rare, ambiguous, or dependent on personal judgment and relationship. If the steps change every time you do the task, or if the quality depends on reading tone and emotion, the agent will either produce generic output or overstep. Keep those tasks with people.

How do I know if a task is ready for AI automation?

Test the task against three markers before you build anything. First, does it repeat at least once a week? Second, can you write the steps in plain language without ambiguity? Third, does the output land where the team already checks? If the task passes all three, it is ready. If it fails one, it is probably not.