How to Find the Best Business Processes to Automate with AI

Most advice about AI automation starts with the tool.

Use this platform. Connect that app. Build an agent. Add a prompt. Watch a video of somebody automating an entire department while you are still trying to get the weekly report out before lunch.

That is backwards.

The first question is not "What can AI do?" It is:

Which job is costing me time, happening often enough to matter, and structured enough to improve?

That is where the useful automation is hiding.

I have spent a lot of time working in the uncomfortable middle ground where the workload is too big for one person, but the budget is too small for a team. Before AI, I built systems to triage competing demands, decide what mattered and get repeatable work moving without relying on heroic effort every week.

AI is extremely useful in that environment. But only if you give it a proper job.

Here is the process I use to find the jobs worth automating.

Start with the work you repeat, not the work you hate

The task you hate most is not automatically the best automation opportunity.

A task might be annoying because it is difficult, ambiguous or emotionally draining. That does not necessarily make it easy to automate. The best first candidates are usually more boring:

• They happen every week or every month.
• They follow roughly the same sequence.
• They begin with information you already have.
• They produce a repeatable output.
• You can tell whether the result is good enough.

A weekly marketing report is a good example. You gather numbers, compare them with the previous period, look for unusual changes, explain what probably happened and decide what to do next.

The work is not identical every time, so a basic rule-based automation cannot handle all of it. But the shape of the job is stable enough for AI to help with the investigation, explanation and first draft.

That is a much better starting point than asking AI to "do marketing".

Make a list of everything you do twice

You do not need a complicated process-mapping workshop. For one week, keep a running list of tasks that you do more than once.

Include the small things. The copy-and-paste between your CRM and spreadsheet. The monthly request for performance numbers. The lead you meant to follow up with. The meeting notes you never turned into content. The same explanation you write to a client for the fourth time.

At the end of the week, group the list into five types of work:

1. Collecting - finding information from different places.
2. Checking - looking for errors, changes or missing details.
3. Deciding - choosing what deserves attention.
4. Producing - turning information into a report, email or plan.
5. Moving - updating a system or passing work to the next person.

AI is particularly useful when a process includes collecting, checking, deciding or producing language-based outputs. Rules-based automation is often better when the job is simply moving data from A to B.

The distinction matters. Not every repetitive task needs AI.

Score each process on five things

Once you have a list, score every process from 1 to 5 against these five questions.

1. How often does it happen?

A task that takes 20 minutes once a year is probably not your first automation target. A task that takes 20 minutes every day is different.

Frequency creates leverage. Even a modest improvement becomes meaningful when it is repeated 250 times a year.

2. How much time does it consume?

Estimate the total time, including the hidden parts:

• finding the information
• switching between tools
• checking your work
• correcting mistakes
• waiting for someone else
• remembering what happens next

People often count only the visible task. The real cost includes all the friction around it.

3. What happens if it is late or wrong?

A process that affects revenue, customer experience or important decisions deserves more attention than one that is merely inconvenient.

This is also where you need to be honest about risk. A process with a serious downside should not be fully automated just because it is repetitive. It may be better suited to an AI-assisted workflow with a human approval step.

4. How much judgement does it require?

Give a low score to work that is completely predictable and a high score to work that involves messy information, interpretation or prioritisation.

Low-judgement tasks are often ideal for normal automation. Higher-judgement tasks are where AI can add value - as long as a person remains responsible for the decision.

5. Can you recognise a good result?

This is the question people skip.

If you cannot describe what a good output looks like, you cannot reliably improve or review it. "Write something useful" is not a quality standard. "Summarise the three biggest changes, cite the underlying numbers and recommend one action for this week" is much better.

A process is a stronger automation candidate when success can be checked.

Use this simple prioritisation formula

You can turn the scores into a rough priority without pretending the result is scientific:

Priority = frequency × time cost × business impact × clarity of output ÷ implementation effort

You do not need perfect numbers. The point is to compare opportunities consistently.

For example:

Process	Frequency	Time cost	Impact	Output clarity	Effort	Priority
Weekly marketing report	5	4	4	4	2	High
One-off event planning	1	4	4	3	4	Medium
Daily social captions	5	2	2	3	1	Medium
Lead follow-up	4	4	5	4	3	Very high

The exact scores are less important than the conversation they force you to have.

If lead follow-up is inconsistent and directly affects revenue, it probably deserves attention before generating more social posts.

Choose the right type of automation

There are three useful categories.

Automate the movement

Use traditional automation when the instruction is predictable:

• When a form is submitted, create a CRM record.
• When an invoice is paid, send a receipt.
• When a task is marked complete, notify the next person.

There is no prize for adding AI to a job that a simple rule can handle reliably.

Assist the judgement

Use AI when the process involves reading, classifying, comparing or drafting:

• Read an enquiry and identify what the person needs.
• Review a week of numbers and flag unusual changes.
• Turn meeting notes into possible content angles.
• Compare a lead with your ideal customer criteria.

These workflows should usually produce a recommendation or draft for a person to review.

Delegate a contained job

A more advanced workflow can give AI responsibility for a defined sequence, provided the boundaries are clear:

• check new leads each morning
• score them against agreed criteria
• draft a suitable follow-up
• flag anything unusual
• report what happened

The job should have a clear start, finish, output and escalation rule. "Run the marketing" is not a contained job. "Review new enquiries, draft replies and escalate anything involving pricing or complaints" is.

Do not automate a broken process

Automation makes a process faster. It does not automatically make it better.

If your lead data is incomplete, automating lead follow-up may simply help you send the wrong messages more quickly. If nobody agrees which marketing numbers matter, an automated report will create a polished argument rather than a useful decision. If the approval process is unclear, an AI content workflow will produce more drafts for nobody to approve.

Before automating, ask:

• Is the goal clear?
• Does someone own the process?
• Is the necessary information available?
• Are the rules understood?
• Is there a way to check the result?
• What happens when the workflow is wrong or unsure?

If the answer to several of these is no, fix the process first. That is not wasted work. It is the preparation that makes automation safe and useful.

Your first workflow should usually be smaller than you think

A common mistake is trying to automate the entire journey at once.

Instead of "automate lead generation", start with:

"Every morning, review yesterday's enquiries, identify which ones need a response, draft replies in my voice and show me anything involving a complaint, discount request or unusual question."

That is specific enough to build, test and improve.

You can add CRM updates, follow-up sequences and reporting later. The first version only needs to remove a clear bottleneck without creating a new one.

The same approach works for reporting:

"Every Monday, compare this week's marketing numbers with the previous four weeks, flag meaningful changes, explain what evidence supports each conclusion and give me three decisions to consider."

That is a real job. It has a schedule, inputs, an output and a review point.

The best automation target is usually hiding in the handoff

Look closely at the moments where work gets stuck:

• a lead arrives but nobody owns the next step
• numbers are available but nobody interprets them
• a meeting produces ideas but no content gets made
• a customer question gets answered repeatedly by different people
• a campaign is planned but the assets and deadlines are scattered
• a report is delivered but no decision follows it

These handoffs are often more valuable than the task itself. The problem is not always that the work takes too long. Sometimes it is that the next action is unclear.

A good AI workflow can close that gap by turning information into a recommendation, draft or assigned next step.

Start with one job, then measure it

Pick one process and run a small pilot for two weeks.

Measure:

• time saved
• number of human reviews required
• errors or corrections
• work that moved faster
• decisions made that would otherwise have waited
• whether the process actually felt easier

Do not measure only how quickly AI produced an output. A fast draft that creates more checking work is not a time saving.

The goal is not to use AI everywhere. The goal is to create more capacity for the work that needs your judgement.

Find your best starting point

If you are stretched across marketing, sales, reporting and admin, it is hard to choose the first process objectively. Everything feels urgent, and the loudest task usually wins.

That is why I built the free AI Readiness assessment. It asks fifteen questions, gives you an AI Readiness Score and points towards the areas where automation could save the most time.

Get your AI Readiness Score →

Or, if you already know the job you want to improve, browse the practical workflows for reporting, lead follow-up, campaign planning, content repurposing and more.

Browse the AI marketing workflows →

The right first automation is not the most impressive one.

It is the job that happens often, costs more than it should, has a clear definition of "good" and gives you back attention you can use somewhere better.