AI agents are not equally appropriate for every task. This guide helps a reader decide where an agent might assist by looking at repeatability, data sensitivity, impact, reversibility, and available human review. The examples are illustrative categories, not measured deployments or promised productivity gains.
Use this five-question screen
- Is the task repeated with a clear input and expected output?
- Can a person verify the result against a source?
- Would a mistake be easy to reverse?
- Does the task involve personal, confidential, financial, health, or legal information?
- Can the system ask for approval before sending or changing anything?
Illustrative task matrix
| Task pattern | Possible support | Boundary to keep |
|---|---|---|
| Organizing public meeting notes | Suggest headings and action items | Human checks names, dates, decisions, and recipients |
| Finding an internal policy passage | Retrieve likely source text | Reviewer opens the cited policy and confirms it is current |
| Drafting a routine response | Prepare a proposed draft | Person checks accuracy, tone, and authorization before sending |
| Changing access or approving a payment | Explain steps or collect required information | Keep the actual approval and irreversible action with an authorized person |
Start with a reversible pilot
Choose one low-impact task. Define normal, ambiguous, and out-of-scope inputs, then record the expected action and stop condition. Begin with read-only permissions where possible. Review errors, data exposure, and time spent correcting output before considering a broader role.
When not to automate
Do not delegate decisions that require accountable professional judgment, informed consent, or a legally authorized person. Do not let an agent send external messages, change records, or make purchases without clear scope and review appropriate to the consequences.
Apply the matrix to a real task boundary
Use the matrix to decide how much authority a candidate use case should receive—not to claim that an agent will improve it. Start by writing one task in the form “When approved input arrives, prepare specific output for named reviewer.” If the input, owner, or acceptable result cannot be described, keep the process manual until it can.
| Task pattern | Safer starting role | Reason to stop or add review |
|---|---|---|
| Find a record in an approved knowledge base | Read-only retrieval with source links | Permission mismatch, missing citations, or conflicting records |
| Classify an incoming request | Suggest a category and mark uncertain cases | Misclassification changes eligibility, priority, or a customer outcome |
| Draft an internal summary | Prepare a draft for its owner | Sensitive material is outside the approved data path |
| Send a message, change a record, or make a purchase | Prepare a proposed action; require an explicit approval | The action is difficult to reverse or affects someone else |
For a low-risk pilot, record the task owner, permitted data, read/write scopes, approval point, stop phrase or disable control, and manual fallback. Review a sample of normal and ambiguous cases with a person who understands the work. If the workflow creates more correction, access, or exception work than the original process, do not widen its permissions.
This is a use-case suitability screen. For the separate capability vocabulary, see the agent-versus-assistant checklist; for implementation controls, use the approval and rollback guide.
Related reading
For a capability-based distinction between agents and assistants, see the comparison checklist. For workflow design at team level, read the governance-first workflow guide.
About the author
Femica Maydinda Harend writes practical technology guides for AILooma, with emphasis on reader-verifiable decisions and clear limits.
Editorial note
Examples are illustrative, not evidence of actual deployments. Apply organizational policy and current product documentation to any real workflow.