Don’t Automate a Bad Process

Automation feels like progress.

A task disappears.

A message sends itself.

A system updates automatically.

Information moves without human intervention.

But automation has an uncomfortable property:

It magnifies process design.

If the process is good, automation can create leverage.

If the process is bad, automation can create mistakes faster.

Start With Elimination

Before asking:

“How can we automate this?”

Ask:

“Should we be doing this at all?”

Businesses accumulate processes.

Reports nobody reads.

Approvals nobody needs.

Meetings that exist because they always existed.

Duplicate data entry.

Notifications nobody acts on.

Automation is wasted on unnecessary work.

First:

Eliminate.

Then Simplify

Suppose a customer onboarding process requires information to be copied manually across five systems.

The instinct may be to automate the copying.

But why are five systems involved?

Could the process be redesigned around two?

Could the information be captured correctly once?

Simplification makes future automation easier and safer.

So the sequence becomes:

Eliminate → Simplify

Then Assist

Before allowing AI to perform the entire process, use it to assist.

Let AI prepare the output.

Keep the human in control.

For example:

AI drafts customer onboarding email.

Human reviews and sends.

AI prepares project status summary.

Project manager verifies it.

AI identifies likely FAQ categories.

Customer service lead approves them.

This stage teaches you where the system fails.

That information is essential before automation.

Now the sequence becomes:

Eliminate → Simplify → Assist

Then Automate

Only after a workflow becomes reliable should you ask whether AI should act with less intervention.

Good automation candidates tend to be:

Frequent.

Stable.

Low-risk.

Rule-bound.

Easy to monitor.

Reversible when something goes wrong.

Examples might include:

Formatting internal information.

Routing approved documents.

Generating low-risk internal summaries.

Updating routine records after verified events.

Automation deserves more caution when:

Customers could be harmed.

Money moves.

Contracts change.

Employees are affected.

Sensitive information is involved.

Errors are difficult to reverse.

Design Exceptions Before Automation

Ask:

When should the automation stop?

For example:

A customer message contains a legal threat.

A transaction exceeds a threshold.

Required information is missing.

AI expresses low confidence.

A customer requests a human.

The output conflicts with approved policy.

These are exception rules.

A mature automated workflow does not merely know what to do.

It knows when not to proceed.

Measure What Happens

Automation should produce evidence.

Track:

Time saved.

Error rate.

Rework.

Customer impact.

Exceptions.

Human interventions.

Cost.

Quality.

If the automation saves 30 minutes but creates 45 minutes of correction work, you have not created leverage.

You have moved the work.

Don't Automate Relationships Just Because You Can

Some interactions are valuable precisely because a person is involved.

A difficult customer conversation.

Negotiation.

Employee feedback.

Complex discovery.

A major sale.

Service recovery after a serious failure.

Technology can prepare humans for those conversations.

It does not necessarily need to replace them.

Automation Is a Permission Decision

This is one of the most useful ways to think about automation.

Automation is not simply a technical question.

It is a decision about permission.

What is this system authorized to do without asking us?

That framing immediately improves the conversation.

Instead of:

“Can AI send this?”

Ask:

“Under what circumstances should AI be authorized to send this without human review?”

The second question forces you to think about risk, exceptions, accountability, and consequences.

That is the difference between automating a task and designing an operating system.

So before automating your next process:

Eliminate what does not matter.

Simplify what remains.

Use AI to assist.

Observe failures.

Define exceptions.

Assign ownership.

Measure results.

Then automate only what has earned the right to run without constant supervision.