Before you automate a process with AI, draw the process

AI can automate increasingly complicated work. But before you give it the work, there's a surprisingly useful thing you can do with a piece of paper: work out what the work actually is.

There is a question I hear increasingly often: Where could we use AI in our business? It's understandable.

But I think there's a better question to ask first: Where is the work? Where are people copying information?

Checking things? Waiting? Chasing?

Making repetitive decisions? Moving information between systems? Writing the same kind of response again and again?

Fixing mistakes? Asking somebody else what happens next? That's where I'd start.

Because if you start with the AI, there's a good chance you'll find somewhere to put it. That doesn't mean you've improved anything.

Automation can make a bad process faster

Imagine this process. A customer completes a form. Someone receives an email.

They copy the information into a spreadsheet. Someone else checks the spreadsheet. They copy some of the information into the CRM.

A manager receives a message asking them to approve it. They reply. Someone goes back into the CRM.

Then another person sends the customer an email. You could add AI to several parts of that. AI could read the form.

AI could write the email. AI could update the spreadsheet. AI could message the manager.

Brilliant. Except we haven't asked why the spreadsheet exists. Or why the information is being copied twice.

Or why approval happens through a message. Or whether every case actually needs approval. We've automated the existing process without asking whether the existing process makes sense.

This is why I start with the process

Long before AI became part of every technology conversation, businesses were dealing with exactly this problem. Systems grow. Processes evolve.

Someone introduces a workaround. Another system gets added. An employee leaves.

A spreadsheet appears. A new requirement gets bolted on. Five years later, everyone follows a process nobody would deliberately design from scratch.

That's normal. But it means automation should begin with understanding. Not technology.

Draw it

It doesn't need specialist software. It doesn't need a consultancy workshop. You can genuinely start with a piece of paper.

Take one repeated piece of work. Then map six things: TRIGGER → INFORMATION → DECISION → ACTION → CHECK → NEXT ACTION

That's it. This is one of the simplest ways I know to begin seeing where AI and automation might actually help.

1. Trigger: what starts the work?

Every process starts somewhere. A customer submits a form. An email arrives.

An invoice becomes overdue. A new lead enters the CRM. An order is placed.

A document is uploaded. A meeting finishes. A date is reached.

Someone asks a question. Write down the trigger. This matters because good automation needs to know when to begin.

If the process only starts because someone remembers to check something every Tuesday, that's useful information too. You've already found a manual step.

2. Information: what do we need to know?

Next ask: What information does somebody need before they can continue? Perhaps it's:

the customer's details, their previous orders, the contents of an email,

the current price, an account balance, a policy,

a document, stock availability, or information from another system.

Then ask where each piece lives. This is often where things get interesting. Because the answer might be:

Some of it is in the CRM, some is in the spreadsheet, and Sarah normally asks Tom for the rest. Now we've learned something important. The problem may not yet be AI.

It may be information.

3. Decision: what needs to be worked out?

This is where AI becomes particularly interesting. Traditional automation is excellent when the decision is something like: If invoice is more than 30 days overdue, send reminder.

Clear rule. Easy. But many business decisions aren't quite like that.

They involve interpreting something. What is this customer actually asking? Which category does this enquiry belong to?

Does this document contain the information we need? What is the appropriate response? Which information is relevant?

Does something look unusual? AI can potentially help with those less rigid decisions. But first you need to identify that there is a decision.

And ideally understand how a good person currently makes it.

4. Action: what happens because of the decision?

Now something changes. Send an email. Update the CRM.

Create a task. Generate a document. Request approval.

Issue a refund. Book an appointment. Move a file.

Notify someone. This is where the distinction between AI advising and AI acting becomes important. Perhaps the AI should recommend the action.

Perhaps it should prepare it. Perhaps it can perform it automatically. Those are separate decisions.

5. Check: how do we know it worked?

This is the bit that gets forgotten constantly. The email was sent. Did it arrive?

The CRM was updated. Was it updated correctly? The payment was requested.

Did it succeed? The document was generated. Does it contain everything it should?

The customer was contacted. Did they reply? A process shouldn't simply perform an action and assume success.

Somebody, or something, needs to check the result.

6. Next action: what happens now?

This is where many automations stop too early. Imagine automating a sales follow-up. The system sends an email.

Done. Except the business process isn't done. What if the prospect replies?

What if they don't? What if they ask for pricing? What if they say not now?

What if they ask to be contacted in three months? The real process continues. So ask:

What happens next? And then, very often, you've found the trigger for another loop through the same framework.

Let's take a normal business example

Imagine a company receives enquiries through its website. Trigger A new website enquiry arrives.

Information We need: the enquiry,

the person's details, what they're interested in, relevant product or service information,

possibly previous contact history. Decision Is this:

a genuine sales enquiry, an existing customer, support,

spam, something urgent, or something we don't handle?

Action Route it appropriately. Create or update the CRM record.

Prepare a response. Assign someone if necessary. Check

Did the record update? Was the enquiry routed correctly? Did the response send?

Next action If they reply, continue the conversation. If they don't, perhaps follow up after a certain period.

Already, we have a much clearer picture of where technology could help.

Now decide which parts need intelligence

This is the next useful step. Not everything in that process needs AI. A form being submitted doesn't require intelligence.

That's a trigger. Creating a CRM record probably doesn't require intelligence either. Normal automation can do that.

Understanding what the person is asking might benefit from AI. Writing a suitable response might benefit from AI. Waiting three days definitely doesn't need AI.

Checking whether they replied may simply be automation. This matters because businesses can easily end up using AI for things ordinary software already does perfectly well. AI should be used where intelligence adds something.

AI and automation aren't interchangeable

This distinction is going to become increasingly important. Traditional automation is very good at: When X happens, do Y.

AI is useful when something in the middle requires interpretation. For example: When an email arrives → understand what it's about → decide where it belongs → take the appropriate action.

The first and last parts may be ordinary automation. The middle may be AI. Real business systems increasingly combine both.

That's usually much more sensible than trying to make AI responsible for everything.

Look for the handoffs

One of the best places to find wasted effort is between steps. Someone downloads a file from System A and uploads it to System B. Someone copies an address from an email into the CRM.

Someone sends another employee a message saying: Can you approve this? Someone checks one system and then updates another.

Someone receives a report and manually creates five tasks from it. Those are handoffs. And handoffs are often excellent automation candidates.

Not because they're difficult. Because they're repetitive.

Look for waiting

Waiting is another useful clue. The process stops because: someone needs approval,

someone hasn't replied, a document is missing, a payment hasn't arrived,

a customer hasn't responded, or somebody hasn't noticed there's something to do. Automation is very good at remembering.

A system can check again. Send a reminder. Escalate after a period.

Create a task. Move something to the next stage. Humans don't need to spend their working day remembering what the software could remember for them.

Look for copying

If somebody regularly copies information from one place to another, ask why. Sometimes there is a perfectly good reason. Often there isn't.

Copying creates: work, delay,

and opportunities for mistakes. Before asking AI to perform the copying, ask whether the systems could simply share the information. The best automated step is sometimes the step you remove entirely.

Look for repeated judgement

This is where AI can add something genuinely new. Suppose an employee receives 100 enquiries and repeatedly decides: sales,

support, billing, spam,

other. That decision isn't difficult for a human. But it does require understanding language.

Historically, automating it reliably may have been awkward. AI changes that. The same applies to:

reading documents, extracting meaning, classifying requests,

summarising information, comparing text, drafting responses,

and spotting patterns. These are places where AI can sit inside an otherwise ordinary automated process.

Look for the person who knows all the exceptions

Almost every process has one. You draw your lovely diagram. Then someone says:

Yes, except when... Listen to them. The exceptions are often where an AI implementation succeeds or fails.

Ask: How often does that happen? How do you recognise it?

What do you do differently? Could the system recognise it? Should it simply send those cases to a person?

You don't have to automate every branch. A system that handles 70% of routine cases well and sends 30% to a human can still remove a huge amount of work.

Don't automate something just because people hate doing it

This sounds odd, but it's important. People may hate a task because the process itself is ridiculous. Suppose someone spends an hour every Friday reconciling two spreadsheets.

You could build an AI agent to perform the reconciliation. Or you could discover that both spreadsheets are generated from the same underlying information and one of them shouldn't exist. The second solution is better.

The objective isn't: Make AI do the annoying task. It's:

Why does this annoying task exist?

Remove before you automate

I'd actually add another step before the framework. Once you've drawn the process, look at every box and ask: Does this step need to exist?

Then: Could these two steps become one? Then:

Could a normal system handle this automatically? Then: Does this part actually require AI?

That's the order I'd use. REMOVE → SIMPLIFY → AUTOMATE → ADD AI Not:

ADD AI → hope.

The best AI opportunity might be halfway through the process

Businesses sometimes look for entire jobs to automate. I think that's often too big a unit. Look for pieces.

Perhaps AI can read incoming requests. A person still makes the important decision. Automation completes the administrative work afterwards.

Or AI prepares everything for the person. The person approves. Then the system completes the remaining steps.

That can be enormously valuable without pretending the whole job has disappeared.

Draw the failure path too

Most process diagrams show what happens when everything works. Real businesses also need: What happens when it doesn't?

The customer isn't found. Two records match. The information conflicts.

The AI isn't confident. The API fails. The payment doesn't go through.

The document is unreadable. The request falls outside the rules. Where does it go?

Who needs to know? Can the process retry? Does it stop?

Failure isn't an edge case to think about after launch. It's part of the process.

Then draw the human

Where does a person genuinely add value? Perhaps they make the final decision. Handle unusual cases.

Approve high-value actions. Deal with emotional customers. Review something legally sensitive.

Apply judgement where the information is ambiguous. Don't leave a human in every step simply because that's how the process works today. But don't remove them from an important decision simply because the AI can technically make one.

Design the human into the process deliberately.

This is where my infrastructure background keeps reappearing

Before web development became the main part of my career, I worked in infrastructure and business environments where building something wasn't simply about the technology. You needed processes. Ownership.

Handoffs. Approvals. Systems.

Exceptions. People needed to know what happened next. AI feels new.

A lot of the implementation problem really isn't. We're still trying to make information and work move through an organisation properly. We just have a much more capable new component available to us.

A simple AI process audit

Take one process in your business this week. Not the biggest one. Choose something annoying and repeated.

Draw: TRIGGER What starts it?

↓ INFORMATION What does somebody need to know?

↓ DECISION What needs to be worked out?

↓ ACTION What happens?

↓ CHECK How do we know it worked?

↓ NEXT ACTION What happens afterwards?

Then go through it again. Mark every place where someone: copies

checks waits chases

searches rewrites re-enters

approves or says: "It depends."

Those are the places I'd investigate. Not all of them need AI. That's the point.

You may discover three different solutions

After doing this, you might find: A process problem Remove or change the step.

An automation problem Connect systems or create a rule. An AI problem

Use AI where interpretation, language, reasoning or flexible decision-making genuinely helps. Often, a good solution contains all three.

This is why I don't start with an AI product

There are thousands of AI tools now. You could spend every week trying another one. But a tool isn't a strategy.

I don't particularly care whether a business has adopted the newest AI platform. I care whether: something takes less time,

customers get a better experience, mistakes fall, information moves properly,

people stop doing pointless administration, or the business can do something it couldn't do before. That's the outcome.

AI is one possible component.

Before you automate the process, understand it

The temptation with AI is to jump straight to what is newly possible. I understand why. The technology is extraordinary.

But the best implementation often starts somewhere much less exciting. A process. A pen.

A piece of paper. And somebody who actually does the work. Ask them to show you what happens.

Not what the procedure says should happen. What actually happens. Follow the information.

Find the decisions. Find the copying. Find the waiting.

Find the exceptions. Find the unnecessary steps. Then decide what to remove.

What to simplify. What ordinary automation can handle. And finally, where AI adds something we couldn't easily do before.

Because making a bad process faster isn't transformation. Sometimes the smartest thing you can do with AI is work out which parts don't need AI at all.

Where to go next

• Don't start with AI. Start with the work. A broader way to find worthwhile AI opportunities across your business.
• Why your AI demo worked and your AI implementation didn't The gap between an impressive prototype and a system that survives the real process.
• The important question isn't what AI can do. It's what you should let it do. Once you've mapped the process, decide how much authority the AI actually needs.
• AI is getting smarter. Your business data might be what holds it back. Your process depends on information. AI needs to know which information it can trust.

Related: Don't start with AI. Start with the work..