How I decide what to automate with AI — and what to keep human

Businesses are getting better at asking, “Can AI automate this?”

I find another question more useful: should it?

Drafting, sorting, summarising and monitoring can now be accelerated in ways that were impractical for many small teams only a few years ago. I use AI across the businesses I work on, and I see enormous value in that increased capacity.

But technical possibility is only one part of the decision.

What matters more is understanding what a task is for, what happens when the output is wrong, how much context it requires and who remains responsible for the result.

A task can be automated and still be a poor candidate for automation.

That distinction has become increasingly important to the way I work.

“Can we automate this?” is the wrong first question

When a team looks at a task and immediately asks whether AI can do it, the conversation starts from the tool.

I prefer to start from the task.

What is it trying to achieve?

How often does it happen?

What happens downstream when the result is wrong?

Take a reply to an unhappy customer. A model can prepare a draft in seconds, and that draft may be perfectly well written.

Whether it should be sent without a person reviewing it is a completely different question.

The answer depends on what happened, who the customer is, what the relationship is worth and how much discretion the situation requires.

Technical possibility and a good automation decision are not the same thing.

I try to keep them separate.

Frequency and cost of error

The first filter I tend to use has two dimensions:

How often does the task happen?

and:

What does a mistake cost?

When a task happens frequently and the cost of an error is low, it is usually a strong candidate for automation.

Tagging incoming enquiries, organising information, preparing first drafts, producing routine summaries or detecting unusual movements in data all fit naturally into this category.

The task repeats often enough for automation to matter, and a wrong result is usually easy to identify and correct.

When frequency is high but the cost of error is also high, I prefer automation with human review.

Customer complaints, quotations, sensitive client communication, financial exceptions or operational decisions can benefit from faster preparation without removing the person responsible for the final output.

Low-frequency, low-risk tasks are different.

If something happens twice a year and takes little time to complete, building a dedicated automation may create more complexity than it removes. A general-purpose AI assistant may be enough.

Then there are low-frequency, high-consequence decisions.

These are usually the ones where I want AI supporting the research and analysis while human judgement remains firmly in control.

I do not treat this as a rigid framework. Plenty of tasks sit somewhere between those categories, and the boundaries move as technology improves and businesses learn.

But frequency and cost of error remove a surprising number of bad automation ideas before anyone spends time building them.

Fix the process before you automate it

Automation speeds up whatever process you give it, including a broken one.

If lead qualification is weak, automating the flow does not suddenly make the criteria better.

If customer support has no clear procedure, adding AI can distribute that confusion more efficiently.

Bad reporting automated every morning is still bad reporting.

Unnecessary approval steps can survive perfectly well inside an automated workflow.

So before I automate something, I want to understand whether the underlying process already produces a useful outcome when people run it manually.

If it does, automation can remove friction and extend it.

If it does not, I would rather fix the process first.

That can feel slower at the beginning, but it usually prevents us from automating the wrong thing.

Automate the noticing. Keep the deciding.

This is the principle I come back to most.

Automation is particularly useful for observation:

monitoring, detecting, sorting, summarising, flagging and preparing information.

People who understand the business handle the other half:

interpreting what a signal means, deciding what to do, setting priorities, changing direction and accepting the consequences.

Advertising gives a clear example.

In our agency, automated monitoring can flag unusual movements in spend, cost per lead or campaign delivery. That means someone does not need to open every advertising account simply to discover whether something changed.

But detecting a change and understanding it are different jobs.

A number may have moved because of seasonality.

A promotion may have ended.

A landing page may have stopped working correctly.

A product may be unavailable.

The client may have changed priorities.

Something may have happened in the market that is not visible inside the advertising dashboard.

The alert is useful because it directs attention.

The decision still benefits from someone who understands the wider business.

That is the distinction I like:

automate the observation, keep consequential decisions with someone who understands the context.

Some decisions are expensive because errors compound

Some mistakes cost what they cost and end there.

Others keep producing consequences.

A pricing decision can affect margin across every sale until it is changed.

A poor hire can affect a team long after the decision is made.

Entering the wrong market can absorb capital, time and management attention.

A major budget reallocation, a change to an offer or a badly judged message to an important client can create effects that take months to unwind.

For these decisions, I still use AI.

It can research a market, organise information, summarise alternatives, challenge assumptions and surface possibilities I may not have considered.

That can improve the decision considerably.

What I do not want to automate simply because it is technically possible is the responsibility for making the call.

Generating an answer is becoming cheaper.

Owning the result is not.

The more expensive or difficult to reverse a decision is, the more context and human judgement I want around it.

Context decides what can be automated

The same task can be a good automation candidate in one business and a poor one in another.

An automated response to an enquiry may work perfectly well for a high-volume e-commerce business selling standardised products.

The same approach may damage the experience of a luxury business where customers expect individual attention, or a service company where a single enquiry can represent a significant contract.

The difference is context.

Margins.

Customer expectations.

Operating model.

Team capabilities.

Risk tolerance.

Brand positioning.

Historical data.

Strategic priorities.

AI can produce an output.

Whether that output is useful depends on information that lives inside the business.

That is why I am cautious about copying automation decisions from one company to another, even when the businesses appear similar from the outside.

A note on building products

I use the same logic when working on digital products.

In projects such as Budget Buddy and Salon Voice, AI helps reduce the distance between an idea and a working version.

The harder decisions remain product decisions: who the user is, which problem deserves attention, what should be simplified and what standard the final experience needs to meet.

The technology accelerates the work.

It does not remove the need to decide what is worth building.

The questions I ask before automating something

When a task looks like a candidate for automation, I usually come back to a few questions.

How often does this happen?

What is the cost if the output is wrong?

Does the process already work today?

Can someone review the result quickly and cheaply?

How much business context does the decision require?

And who remains responsible if the output turns out to be wrong?

When the answers point toward high frequency, low cost of error, a working process and inexpensive review, automation becomes attractive.

When they point in the opposite direction, I am more interested in using AI to prepare information than in allowing it to make the decision.

Most real tasks sit somewhere between those two extremes.

The questions help me decide where.

More automation is not the objective

I like AI in business for practical reasons.

It removes repetitive work, gives teams more capacity and puts better information in front of people sooner.

I expect to automate more across the businesses I work on.

But more automation is not the objective.

The goal is to remove friction where it adds little value, preserve judgement where the consequences matter, and give the people making decisions better information before they make them.

The best systems, in my experience, automate repetitive observation while leaving consequential decisions with someone who understands the context.

Automate the noticing. Keep the deciding.