Should You Automate It? A Decision Checklist

Just because something can be automated does not mean it should be. The best automation removes the administrative work around people without removing people from moments that require care, trust, judgment, or accountability.

This checklist helps you decide what AI can safely assist with, where a person should remain in control, and when an automation should not be built at all.

What is the human test for automation?

If the interaction involves emotion, trust, judgment, or a relationship, slow down before removing the person.

Ask these five questions first:

1. Emotion: Could the person be upset, afraid, embarrassed, grieving, or angry?
2. Trust: Does the outcome depend on feeling heard, respected, or treated fairly?
3. Judgment: Are the rules incomplete, the facts disputed, or the exceptions important?
4. Relationship: Could a poor interaction damage a valuable customer, employee, patient, or partner relationship?
5. Accountability: Does a person need to explain, own, or defend the decision?

Review this process using the human test: [describe process]. Identify where emotion, trust, judgment, relationships, or accountability matter. Separate the administrative work AI could assist with from the moments a person should lead. Explain any uncertainty instead of assuming it is safe to automate.

A fast response is not automatically a good experience. Sometimes the most efficient design is to help a person respond better, not replace the response.

How do you assist the person instead of replacing the moment?

Use AI to prepare, organize, and surface information while a person handles the important interaction or decision.

1. Employee conversation: AI can organize notes, suggest questions, and help a manager prepare. The manager should have the conversation.
2. Unhappy customer: AI can flag sentiment, collect the history, and draft options. A person should protect the relationship and choose the response.
3. Hiring: AI can summarize approved information or help schedule interviews. People should review evidence, challenge bias, and make accountable decisions.
4. Sensitive decisions: AI may support the work, but it should not quietly become the final decision-maker.
5. Ownership: Keep a clear owner who can correct, override, explain, and stop the process.

Redesign this process so AI assists the person without taking over the human moment: [describe process]. Show what AI can prepare, what a person must review, who makes the final decision, what the person should see, and how someone can correct or appeal the outcome.

What are the red flags that mean you should pause?

One red flag does not always mean no, but it does mean you need stronger controls, better data, or a human-led process.

1. A mistake could affect someone''s job, money, health, safety, legal rights, or access to an important service.
2. The process uses private, confidential, regulated, or highly personal information.
3. Historical data may contain bias or reflect decisions your business would not make today.
4. The rules are unclear, exceptions are common, or experienced employees regularly use judgment.
5. The system cannot show why it produced a result or where the information came from.
6. The action is public, difficult to reverse, or likely to create a promise your business must keep.
7. No one has been assigned to monitor failures, complaints, corrections, or changing conditions.

Act as a cautious reviewer. Find the risks in this proposed automation: [describe it]. Check for high-stakes impact, private data, bias, unclear rules, missing explanations, difficult-to-reverse actions, weak monitoring, and lack of ownership. Rate each risk as low, medium, or high and recommend proceed, redesign, pilot only, or do not automate.

Where are the green lights for a first automation?

Good first automations tend to be repetitive, predictable, low risk, reversible, and easy to measure. Look for this pattern:

1. The task happens often and follows a stable process.
2. The input and desired output are easy to describe.
3. A mistake is easy to notice, correct, and reverse.
4. A person can approve the result before it affects a customer, employee, or important record.
5. Success can be measured through time saved, errors reduced, faster service, or another useful result.
6. There is a clear owner, an exception path, and a stop button.

Evaluate this task as a possible low-risk automation: [describe task]. Score it from 1 to 5 for frequency, predictability, reversibility, human review, measurability, and clear ownership. Explain each score, identify what is missing, and recommend the smallest safe pilot.

How do you design a human-in-the-loop process?

Human review should be a real control, not a box someone clicks without enough time or information. Build these seven parts:

1. Trigger: What starts the process?
2. AI assist: What can AI classify, summarize, extract, or draft?
3. Evidence: What sources and confidence information will the reviewer see?
4. Human decision: Who approves, changes, rejects, or escalates the result?
5. Action: What happens only after approval?
6. Record: What decision, source, change, and reason should be logged?
7. Stop condition: What error, complaint, exception, or threshold pauses the automation?

Map this process as Trigger, AI Assist, Evidence, Human Decision, Action, Record, and Stop Condition: [describe process]. Make the review meaningful by showing what the person needs to see, how much time they need, what they can change, and when the task must be escalated.

Run the decision checklist

Use this before choosing a tool or connecting a live system. Answer yes, no, or not yet:

1. Can we clearly explain the current process and the problem we are solving?
2. Are the rules stable enough to describe, including common exceptions?
3. Is the data accurate, permitted, and appropriate for this use?
4. Can a person review, correct, override, and appeal the result?
5. Can we reverse the action and stop the automation quickly?
6. Do we know what success and failure will look like?
7. Is there a named owner who will monitor performance and complaints?

Walk me through this automation decision one question at a time: [describe idea]. Ask about the current process, rules, exceptions, data, human review, reversibility, success measures, failure signals, and ownership. At the end, recommend automate, assist only, pilot with controls, not yet, or keep human, and explain why.

Make the decision

Your goal is not to automate the most work. It is to remove low-value friction while protecting the parts of the experience where people make the difference.

• Automate when the task is predictable, low risk, reversible, measurable, and owned.
• Assist only when AI can prepare the work but a person should lead the interaction or decision.
• Not yet when the data, rules, controls, or ownership are not ready.
• Keep it human when trust, care, accountability, or high-stakes judgment is the work.