In this guide
Divide tasks, not jobs
The common question is "which jobs could AI do?" It is usually not the most useful level. Most roles mix work that needs judgement with routine preparation. Dividing whole jobs either hands too much to an agent or too little.
The useful level is the task.
In short
Look at each task and ask whether it is well defined, repeated often and easy to check. Those answers point to candidates for delegated work. Separately, the stakes and whether a mistake can be undone decide how much control it needs, whatever the other answers.
Questions to find candidates
- Is it well defined? Clear inputs, a clear result and known rules make a task easier to delegate. Ambiguous situations suit a person.
- Is it repeated often? Frequent, similar tasks can be candidates for delegated support.
- Is it easy to check? If a person can quickly confirm the result is right, an agent can prepare the work and a person can review it.
These questions help identify candidates for delegated work. They do not prove a task needs an AI agent. Where the steps are predictable, a conventional workflow may be enough. Agents may be useful where the work varies and the next step depends on the case, provided suitable controls are in place.
The limits come first
Two further questions are not part of a score. They set the controls:
- What are the stakes?
- Can a mistake be undone?
High-stakes or hard-to-undo work needs a person to review and approve before it goes ahead, whatever the other answers. Frequency or clear instructions do not remove that requirement.
Give agents the preparation. Keep people on the judgement.
What usually suits each
Often suits delegated AI support: gathering information from several systems, checking records against a policy, drafting routine replies, keeping a process moving between steps. Give an agent only the data access and action authority the task needs, and apply the organisation's privacy, security and recordkeeping requirements.
Usually stays with people: decisions with real consequences, situations that do not fit the rules, conversations that need trust or tact, and decisions a person must make and own.
These are tendencies, not rules. Your own answers, and the controls the stakes require, are what count.
Start small and keep the record
Begin with one task that is a clear candidate and low risk, set the review level to match the stakes, and watch the record. The division can change as confidence grows, but that is always the organisation's decision, not the agent's.
What is agentic software? explains how people and agents share one system under the same rules.
Questions
Frequently asked.
Start with tasks, not jobs. Most roles mix judgement-heavy work and routine preparation. Dividing at the task level keeps people on the parts that need them.
Not necessarily. The questions identify candidates for delegated work. If the steps are predictable, a conventional workflow may be enough. Agents may be useful where the work varies and the next step depends on the case, provided suitable controls are in place.
Yes. As a team gains confidence in how an agent performs a task, and the record shows it working well, the level of review can be adjusted. The organisation decides, not the agent.
Written by
Mare Pomana
Plumb Digital