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

If your first question is “Where can we use AI?”, you've already chosen the solution before you've properly understood the problem. There's a better place to begin.

There's a conversation happening in businesses everywhere. Someone comes back from a conference. Or watches a demo.

Or sees what another company is doing. And asks: We need to be doing more with AI. Where can we use it?

I understand the question. AI can now do things that would have sounded ridiculous a few years ago. Of course businesses want to know how they can use it.

But I think there's a problem with starting there. You've already chosen the technology. Now you're wandering around the business looking for somewhere to put it.

I'd turn the question around. Where is the work? Then let's see what could make it better.

Find the friction

Don't start by looking for an AI use case. Look for friction. Where are people:

copying information, searching for things, chasing people,

checking things, rewriting things, re-entering data,

waiting for approvals, switching between systems, producing the same kind of document repeatedly,

reading large amounts of information to find a small answer, or doing something manually because "that's just how we've always done it"? That's where I'd look first.

Not because all of those things need AI. They don't. Because those are the places where improving the way work moves through the business can create actual value.

"Where can we use AI?" is too broad

Almost anywhere. That's increasingly the problem. AI can help with:

sales, marketing, customer service,

operations, research, analysis,

software development, administration, finance,

documentation, reporting, training,

and internal knowledge. So asking where AI can be used doesn't narrow things down very much. A better question is:

Where is there a problem worth solving? That gives us somewhere useful to start.

Sometimes the answer won't be AI

This is important. Imagine someone spends an hour every Friday moving information from one spreadsheet into another. You could use AI.

But perhaps the spreadsheets should simply be connected. Or perhaps one of them shouldn't exist. Imagine staff repeatedly ask the same internal question.

You could build an AI assistant. But perhaps the information is just badly organised. Imagine every sales enquiry needs a manager's approval.

You could use AI to prepare the approval request. But perhaps 80% of those enquiries shouldn't require approval at all. Technology can improve a process.

It can also help us avoid asking why the process is ridiculous in the first place.

I don't think businesses need an "AI use case"

They need a business case. This sounds like a small distinction. It isn't.

An AI use case asks: What could AI do? A business case asks:

What gets better if we change this? Does it: save time?

reduce mistakes? respond to customers faster? increase capacity?

remove repetitive work? improve consistency? make information easier to access?

allow the company to offer something new? reduce a cost? help somebody make a better decision?

That's what I care about. The AI comes afterwards.

Start with the annoying stuff

Some of the best opportunities aren't particularly glamorous. Ask a team: What do you waste time doing every week?

You'll get much more useful answers than if you ask: How would you like to use artificial intelligence? You'll hear things like:

I spend ages finding the right information before I reply. We have to copy every enquiry into another system. Someone checks these manually every morning.

I spend Friday afternoon putting this report together. Customers keep asking the same five questions. We have three systems and none of them quite agree.

Every time this happens I have to ask John what to do. That's where the opportunity is.

Look for repeated movement

A lot of business work isn't really creating anything. It's moving things. Information arrives here.

Someone reads it. Copies part of it there. Checks something somewhere else.

Asks someone a question. Updates another system. Sends a response.

Creates a task. Waits. Checks again.

None of those individual steps necessarily takes very long. Together, they can consume an enormous amount of a working week. This is why I think workflow is becoming such an important way to think about AI.

The opportunity isn't always making one task faster. It's reducing how much human effort is required to move the work from beginning to end.

Look for repeated interpretation

This is where AI adds something that ordinary automation often struggled with. Businesses are full of information that isn't neatly structured. Emails.

Documents. Conversations. Forms.

Notes. Images. Customer requests.

Reports. Someone has traditionally needed to read those things and work out what they mean. AI can increasingly help with that interpretation.

What is this customer asking? Which information matters in this document? Which category does this belong to?

What has changed? What needs attention? What should happen next?

That ability can sit inside a much larger process. And that's often where AI becomes genuinely useful.

Look for knowledge trapped in people's heads

This one is enormous. Ask: Who knows how this works?

If everybody points to one person, investigate. Businesses accumulate enormous amounts of undocumented knowledge. Which supplier to use.

How to handle a particular customer. Why one system works differently. What to do when something unusual happens.

Where a particular document lives. Which report can actually be trusted. AI can potentially make organisational knowledge much easier to access.

But first you need to recognise that the knowledge exists. And ideally capture it before the person who knows everything goes on holiday.

Look for waiting

Some processes are slow not because the work takes a long time, but because the work spends most of its life waiting. Waiting for someone to notice it. Waiting for approval.

Waiting for information. Waiting for a reply. Waiting for somebody to remember.

Waiting for the next person to pick it up. AI and automation can sometimes remove huge amounts of elapsed time without making any individual task dramatically faster. That can matter to customers far more than saving three minutes of employee time.

Look for checking

Checking is another interesting category. People check: documents,

orders, prices, stock,

accounts, reports, applications,

forms, websites, competitor information,

customer records, and whether something happened when it was supposed to. Some checks are mechanical.

Normal automation can handle them. Some require interpretation. AI may help.

Some carry enough risk that a person should still perform them. The important thing is to understand which kind you're dealing with.

Look for work that has grown around the software

This is something I've seen throughout my career. A company buys software to support a process. The software doesn't quite match the process.

So people create workarounds. A spreadsheet. An export.

A manual check. A shared folder. An email chain.

A second system. A note saying: Remember to change this afterwards.

Eventually, the workaround becomes part of the job. AI may now make some of those gaps easier to bridge. But again, don't assume the workaround needs automating.

First ask why it exists.

Then measure it

This is where vague AI ambition becomes something useful. Take one piece of work and work out: How often does it happen?

How long does it take? How many people touch it? How long does it wait between steps?

How often does something go wrong? How much checking is required? What happens when there's a mistake?

What is the business consequence? Now you have a baseline. And suddenly you can have a sensible conversation about whether changing the process is worth doing.

Time saved isn't the only value

This matters because AI projects often get reduced to: How many hours will this save? That's useful.

But it isn't everything. Imagine a customer enquiry currently takes two days to answer because it sits in three different queues. An improved workflow might only remove ten minutes of actual employee effort.

But if the customer receives a useful response in an hour instead of two days, that's a very different kind of value. Other improvements might be: fewer errors,

better consistency, faster response, greater capacity,

better visibility, less dependence on one person, fewer dropped tasks,

or being able to offer something that wasn't practical before. Measure what matters to the process.

Then work out what kind of problem you've actually found

Once you understand the work, you can choose the technology. You may have found a process problem. The process itself needs changing.

You may have found an information problem. People can't easily find or trust what they need. You may have found an integration problem.

Systems aren't sharing information. You may have found an automation problem. The rules are clear, but humans are still performing the steps.

You may have found an AI problem. Something requires interpretation, language, reasoning or flexible handling that traditional automation couldn't easily provide. Or you may have found several at once.

That's normal.

AI is one layer of the solution

This is why I get uncomfortable when every business problem is suddenly described as an AI problem. A useful AI system may still depend on: a database,

an API, normal automation, business rules,

permissions, a user interface, email,

a CRM, a website, and a human approval step.

The AI might only perform one part. That's fine. The objective isn't to maximise the amount of AI in the architecture.

It's to build the best system.

Start with the outcome, not the product

This also protects businesses from chasing tools. A new AI product launches. Everyone gets excited.

Six months later there's another one. If your strategy begins with the product, you're constantly starting again. Instead, define the outcome.

We want every genuine sales enquiry acknowledged within ten minutes. We want staff to find the correct internal policy without searching five folders. We want invoices checked without someone manually opening every PDF.

We want this weekly report produced without three hours of copying and pasting. Now the technology can change underneath the objective. That's much more durable.

The newest AI isn't necessarily the right AI

Sometimes the task is simple. If a cheaper, smaller system performs it reliably, use that. If ordinary automation can do it, use ordinary automation.

If a spreadsheet formula solves it, use the spreadsheet formula. If removing a step solves it, remove the step. There is no prize for putting the most sophisticated AI into the most places.

A good implementation can be technologically boring. That's often a compliment.

Workflows are a better unit than jobs

I also think this helps with the constant question: Which jobs will AI replace? A job is usually a collection of very different tasks.

Take someone in customer service. Their day might include: reading messages,

understanding the issue, looking up information, making decisions,

writing responses, updating records, handling angry customers,

escalating unusual cases, and coordinating with colleagues. AI may be excellent at some of those things.

Less appropriate for others. So instead of asking whether the job can be automated, map the work inside it. That's much more useful for businesses and much less likely to produce ridiculous conclusions.

The human should stay where the human adds something

I don't think the objective is to remove people from every process. It's to stop using people as expensive connectors between systems. If someone's role in a step is:

copy this field, paste it there, download this,

upload it here, check whether this changed, send the same reminder,

move this into that folder, then technology should probably be doing more. If the person is:

handling an unusual situation, making a high-consequence decision, understanding a sensitive customer,

applying experience, negotiating, creating trust,

or deciding between genuinely ambiguous options, that's different. Design around where human involvement actually adds value.

This isn't really a new philosophy

Before AI, I worked across sales, infrastructure, processes and then around 15 years in web and digital. The technologies have changed enormously. The underlying question hasn't.

What are we trying to make work better? A website wasn't useful because a business "had a website". A CRM isn't useful because a company "implemented a CRM".

Automation isn't useful because a process "uses automation". And AI isn't useful because the company can say it's using AI. The value comes from what changes afterwards.

AI makes this discipline more important

There's a strange effect when technology becomes easier to use. It becomes easier to use it unnecessarily. AI can produce something impressive in minutes.

You can build prototypes quickly. Connect systems. Create agents.

Generate content. Analyse enormous amounts of information. That's extraordinary.

But it means the cost of asking "could we?" has fallen dramatically. So the value of asking "should we?" goes up.

A simple place to begin

If I were looking at a business that wanted to use more AI, I wouldn't begin by creating a list of AI tools. I'd ask people across the company one question: What do you do repeatedly that feels like a waste of your time?

Then I'd collect the answers. For each one, I'd ask: What triggers it?

What information is needed? What decision gets made? What happens next?

Where does it get stuck? What goes wrong? What requires a person?

How often does it happen? What would improve if we changed it? Only then would I ask:

What technology belongs here?

You might end up using a lot of AI

That's entirely possible. Once you look at the work properly, you may find excellent opportunities everywhere. AI might:

read incoming information, understand requests, retrieve knowledge,

prepare responses, analyse documents, support decisions,

operate software, write code, create reports,

or coordinate parts of a workflow. But now each use has a reason to exist. It's attached to a problem.

An outcome. A process. A measure of success.

That's very different from adopting AI because everybody else appears to be doing it.

Don't start with AI

Start with the customer waiting too long. Start with the employee spending Friday afternoon copying information. Start with the spreadsheet nobody trusts.

Start with the inbox everybody forgets to check. Start with the process that requires five people when nobody knows why. Start with the information nobody can find.

Start with the report that takes three hours to produce and ten minutes to read. Start with the thing everybody complains about. Start with the work.

Then ask what should change. Maybe the answer is AI. Maybe it's automation.

Maybe it's better integration. Maybe it's a simpler process. Maybe you should stop doing the thing altogether.

That's not being less ambitious about AI. It's how you make AI useful.

Where to go next

• Before you automate a process with AI, draw the process Use the Process Map to break a workflow into Trigger, Information, Decision, Action, Check and Next Action.
• Why your AI demo worked and your AI implementation didn't Proving AI can do a task is different from building something a business can depend on.
• AI is getting smarter. Your business data might be what holds it back. Once you have found the right work, make sure the AI can access information it can trust.
• The important question isn't what AI can do. It's what you should let it do. Capability isn't authority. Decide where AI acts and where a human stays involved.

Related: Before you automate a process with AI, draw the process.