People, automation or AI: which does your task need?

Start with the result, not the tool

A recurring task doesn't automatically need AI. Identify the result you need, the inputs you have and the decisions involved, then choose a method you can explain, review and maintain.

Most workflows combine all three: automation moves information between systems, AI prepares an interpretation or draft, and people approve decisions and resolve exceptions. The real question is which responsibility belongs where, and the answer starts with choosing the right first task for AI.

When should a person do the work?

Customer commitments, conflicting instructions and unfamiliar exceptions need a clear owner. A specialist may need to ask a follow-up question, apply policy or decide whether work should proceed.

For example, a refund request outside the usual terms needs a person to review the circumstances. Preparing the request can be assisted; approving the exception follows your authorization process.

Before adding tools, check whether better instructions or more capacity would solve the problem. Often a trained support team with a clear playbook beats a new system.

When is automation enough?

When the action follows stable conditions, a simple automated step is usually enough: notifying a team after a form submission, copying an approved field between systems or creating a task when a record changes status.

Document the trigger, the fields used and what happens when information is missing. Never treat incomplete records as successful work silently. Someone needs to see failures and be able to fix them.

When does AI help?

Requests, documents and conversations arrive in many forms. AI can summarize material, suggest a category, extract candidate information or draft a response for review.

Define which sources it may use and which questions it must leave unanswered. A suggested category or extracted field is an output to check against your quality standard. Plan for unclear inputs, missing information and plausible-looking mistakes, the same failure modes listed in the OWASP Top 10 for LLM applications (opens in a new tab).

Document-heavy processes are a common fit, because AI-assisted extraction with human validation handles messy inputs well. Customer-facing drafts are another, as long as support replies are drafted from approved knowledge and escalation rules are clear.

Example: one inquiry workflow, three roles

Say your team receives inquiries by email and web form:

1. 1Automation collects each inquiry in one queue and notifies the responsible team.
2. 2AI suggests a topic and prepares a short summary.
3. 3A person checks the request, resolves uncertainty and approves the reply or next step.
4. 4Reporting records completions, corrections and exceptions so the process improves.

This is a planning example, not a client system. The right design depends on your channels, permissions, data and responsibilities.

Checklist: which method fits your task?

• Can you describe the input and the accepted result clearly?
• Are the rules stable, or does the task need interpretation?
• Who approves decisions and customer commitments?
• Which sources and systems may be used?
• What happens when an input is missing or uncertain?
• Can the reviewer spot a wrong result?
• How much review and correction time is acceptable?
• Who maintains the process after the pilot?

If these answers are unclear, a short AI readiness and workflow assessment will surface them. Measure from input to accepted result, including review and correction, before deciding a workflow is worth keeping.

Frequently asked questions