One of the easiest mistakes to make with AI is to confuse technical capability with business value.
A system does something that previously required a person.
It analyses a document in seconds.
It generates working software from a description.
It handles a customer conversation.
It examines thousands of pieces of information and produces an answer almost immediately.
The demonstration can be extraordinary.
But businesses don’t get paid for impressive demonstrations.
They get paid when something changes economically.
Revenue increases.
Costs fall.
Margins improve.
Customers receive something better.
Risks reduce.
Capacity increases.
Decisions improve.
A constraint on growth disappears.
This creates an important distinction when evaluating AI opportunities.
How impressive is the technology?
and
How valuable is the problem?
are completely different questions.
A technically sophisticated AI system applied to an economically insignificant problem can still produce an insignificant result.
Meanwhile, a relatively mundane application of AI to something economically important can create enormous value.
So if businesses want to find their best AI opportunities, I think the starting point has to be economics.
Technology doesn’t determine the size of the opportunity#
Imagine two potential AI projects.
The first automates 80% of an administrative process.
The technology works extremely well. What previously required a person to read information, interpret it and enter something into another system can largely happen automatically.
It looks like an excellent AI use case.
Suppose the entire process currently costs the company £40,000 a year.
Even if the technology eliminated the cost completely — which it almost certainly wouldn’t — the theoretical opportunity is still only £40,000.
Now consider something much less dramatic.
A business generates £10 million of revenue through a sales process.
AI doesn’t automate the sales team.
It doesn’t replace the CRM.
Customers don’t interact with a chatbot.
Instead, it helps the sales team identify which opportunities deserve attention and gives them better information before important conversations.
Suppose the result is a 1% improvement in the economic performance of that £10 million revenue stream.
That small improvement could be worth substantially more than automating most of the administrative process.
These numbers are deliberately illustrative, but the principle matters.
In the first example, AI transforms a small economic surface.
In the second, it makes a small improvement to a large one.
The size of an AI opportunity is determined less by the percentage of a task the technology can perform than by the economic value of the thing being improved.
That changes where we should look.
Follow where value is created#
Every business has an economic architecture.
Money enters in particular ways.
Costs accumulate in particular places.
Certain activities create margin while others consume it.
Some customers are substantially more valuable than others.
Some decisions have enormous downstream consequences.
Some constraints prevent the organisation from producing more.
Some mistakes are cheap. Others are extraordinarily expensive.
Before looking for AI opportunities, I’d want to understand that architecture.
Where does revenue actually come from?
What determines whether a customer buys?
What determines how much they spend?
Why do customers leave?
Where does margin disappear?
Which activities consume disproportionate amounts of expensive people’s time?
What prevents the company from serving more customers?
Where do delays have financial consequences?
Where does poor information produce expensive outcomes?
Where would improving something by 1% actually matter?
Those questions begin to reveal the economic landscape of the company.
AI can then be evaluated against it.
This is quite different from walking department by department asking what could be automated.
We’re looking for economic leverage.
Look for large economic surfaces#
Some parts of a business are simply larger than others.
A process might involve £50,000 of annual cost.
Another might influence £20 million of revenue.
One decision might happen ten times a year.
Another might happen 100,000 times.
One activity might affect a handful of customers.
Another touches every customer the company has.
This doesn’t mean the largest number always contains the best AI opportunity.
But it does mean scale matters.
Consider customer retention.
Suppose a subscription business loses a meaningful proportion of its customers every year.
AI might potentially help identify customers showing early signs of leaving, understand the reasons behind dissatisfaction or help employees determine the best intervention.
None of that needs to eliminate the people responsible for customer relationships.
If better information allowed the business to retain even a small additional percentage of a sufficiently valuable customer base, the economics could be compelling.
The same logic might apply to pricing.
Or conversion.
Or fraud.
Or inventory.
Or product recommendations.
Or operational errors.
Or purchasing decisions.
The interesting characteristic isn’t that these are “AI use cases”.
It’s that small changes can act across large economic surfaces.
That’s where leverage appears.
The best AI opportunities may not produce the biggest percentage improvements. They may produce small improvements to the biggest numbers.
Cost is only one part of the map#
Cost reduction naturally attracts attention because it’s relatively easy to calculate.
If 20 people spend five hours a week performing a task and AI can reliably reduce that work, we can estimate the value of the time released.
That’s useful.
But focusing only on cost can hide much larger opportunities.
Revenue deserves the same attention.
Could better intelligence increase conversion?
Could salespeople identify better opportunities?
Could customers discover products more effectively?
Could greater personalisation increase how much they buy?
Could a company economically serve a customer segment that previously wasn’t viable?
Margin matters too.
Could AI improve pricing decisions?
Reduce waste?
Improve purchasing?
Help employees recognise where profitable and unprofitable work differs?
Then there is capacity.
Imagine a specialist team whose expertise limits how much business a company can accept.
If AI allows those people to handle 30% more work without degrading quality, the value isn’t necessarily a 30% reduction in headcount.
The company might choose to keep every person and sell more.
The same technology can create completely different economic outcomes depending on where the constraint sits.
That’s why beginning with “How much time could this save?” can be too narrow.
Sometimes time is the value.
Sometimes releasing time creates the capacity through which much greater value becomes possible.
Decisions can carry enormous leverage#
Some opportunities are harder to see because they aren’t traditional processes at all.
They’re decisions.
Which lead should a salesperson pursue?
What price should we offer?
How much inventory should we buy?
Which customer requires intervention?
Which supplier should we use?
Which product should we develop?
Where should we allocate marketing budget?
When should we take a risk?
These decisions might take minutes.
The economic consequences can last months or years.
That creates an unusual opportunity for AI.
Suppose a decision currently takes ten minutes and AI reduces it to five.
We could calculate five minutes of saved labour and call that the benefit.
But that might completely miss the point.
If better information also makes the decision slightly more accurate, the value of the improved outcome could dwarf the value of the saved time.
This is one reason I think businesses should be careful about evaluating AI primarily through productivity.
Productivity measures the economics of performing the activity.
Sometimes the much larger opportunity lies in improving what happens because of the activity.
The cost of making a decision and the value of making a better decision can be radically different.
Constraints reveal hidden value#
I’d also look for places where the business has adapted itself around scarcity.
Perhaps a company only provides detailed analysis to its largest customers because doing it for everyone would require too many people.
Perhaps salespeople only research the biggest prospects.
Perhaps managers receive monthly reports because producing them daily would be impractical.
Perhaps an expert reviews only unusual cases because there isn’t enough of their time to review everything.
Perhaps internal software doesn’t get built for smaller problems because development capacity is reserved for larger projects.
These constraints create invisible boundaries around what a business considers possible.
And because they’ve often existed for years, people stop noticing them.
They simply become:
how we operate.
AI can make some of those assumptions worth revisiting.
If an activity becomes ten times faster, dramatically cheaper or available to people who previously lacked the necessary expertise, the opportunity isn’t necessarily reducing the cost of what already happens.
It might be expanding the activity to places where it previously couldn’t economically exist.
That connects directly to a broader idea I’ve written about before: if useful intelligence becomes less scarce, businesses may discover valuable things they can do that weren’t previously worth doing at all.
The economic value sits not in automation, but in removing a constraint.
Value still has to survive implementation#
Finding a large economic opportunity is only the beginning.
Suppose we identify an AI application that could theoretically create £1 million of annual value.
That doesn’t make it a £1 million business case.
The technology has to work reliably enough.
It needs access to the right information.
It may need integration with existing systems.
Employees may need to change how they work.
Customers may need to behave differently.
There may be security, legal or regulatory considerations.
The cost of implementation matters.
So does ongoing operation.
And if the model makes mistakes, the cost of those mistakes matters too.
The relevant number isn’t therefore:
How much value exists?
It’s closer to:
How much of that value can we realistically capture?
This distinction is important because AI discussions can become strangely detached from implementation.
A presentation identifies a huge theoretical opportunity.
A pilot demonstrates technical capability.
Everyone gets excited.
But the distance between the model can do this and the business reliably captures the value can be substantial.
Economic potential tells us where to investigate.
It doesn’t remove the need to prove that the economics survive contact with the real organisation.
Put value and feasibility together#
This gives us a relatively simple way to think about prioritisation.
For every potential AI opportunity, ask two broad questions.
How valuable would solving this problem be?
And:
How realistically can we solve it?
The first requires business understanding.
The second requires technical and operational understanding.
High feasibility with little economic value may produce an easy project, but not necessarily an important one.
Huge economic value with no realistic technical route may be worth watching, but not investing heavily in today.
The most interesting opportunities combine meaningful economic potential with sufficient feasibility to act.
And because AI capability is moving quickly, feasibility isn’t fixed.
An opportunity can move.
A problem that was too difficult twelve months ago might become viable.
A project that was too expensive might cross an economic threshold as inference costs fall.
A constraint caused by limited context might disappear as models become capable of working with more information.
This is why businesses may eventually need to manage AI opportunities less like a one-off list of projects and more like a changing portfolio.
The economic value may remain relatively stable.
The technical feasibility can move underneath it.
Start with the number you want to move#
Before approving an AI project, I’d ask one question:
Which economic number are we trying to change?
Revenue?
Cost?
Margin?
Conversion?
Retention?
Capacity?
Error rate?
Risk?
Time?
Customer value?
Perhaps the answer isn’t immediately measurable in pounds.
That’s fine. Not every worthwhile investment can be reduced perfectly to a spreadsheet.
But there should still be a credible chain between what the technology does and why the business should care.
AI performs X.
That changes Y.
Changing Y improves Z.
And Z matters economically.
If that chain can’t be explained, I’d be cautious about the project regardless of how impressive the technology looks.
Because over the next few years, businesses are going to encounter an extraordinary number of things AI can do.
The scarce resource won’t be ideas.
It will be management attention, implementation capacity and capital.
Those resources need to be directed towards the opportunities that matter most.
Which means the search shouldn’t begin with the cleverest model or the most impressive demonstration.
It should begin with the business.
Understand where the large numbers are.
Understand what moves them.
Then work backwards to the technology.
The best place to look for AI isn’t where the technology looks most impressive. It’s where a change in capability can move a meaningful economic number.