When people imagine a valuable AI project, they tend to imagine a significant change.
A process becomes almost completely automated.
A task that took hours takes seconds.
A new capability appears that wasn’t previously possible.
An entire category of work changes.
Those opportunities are exciting because the impact is obvious.
But I’m increasingly interested in a less dramatic possibility.
What if AI makes something important only slightly better?
A salesperson becomes marginally better at deciding which opportunity deserves attention.
A company gets slightly better at retaining customers.
A pricing decision improves by a fraction.
An operational team makes slightly fewer mistakes.
A customer finds the right product slightly more often.
An employee saves a few minutes performing an activity that happens thousands of times.
None of these sounds transformational.
But whether an improvement matters depends on what sits underneath it.
A 1% improvement to something economically insignificant is insignificant.
A 1% improvement to something affecting millions of pounds may not be.
This creates a useful way of thinking about AI opportunities.
Don’t just ask how much AI could improve something. Ask how much economic value the improvement acts upon.
Because sometimes the most valuable AI transformation won’t look transformational at all.
Percentages need denominators#
Imagine someone tells you an AI system can improve a particular business outcome by 5%.
Is that exciting?
There isn’t enough information to know.
Five per cent of what?
The same applies when someone says AI can automate 80% of a task, reduce processing time by 50% or improve conversion by 2%.
The percentage tells us the magnitude of the change.
It doesn’t tell us the value.
Suppose an AI system could eliminate 80% of the effort involved in an administrative process costing £20,000 a year.
Even under the unrealistic assumption that every pound of theoretical efficiency became a cash saving, there is a ceiling on the economic opportunity.
Now consider a much less dramatic application.
Imagine a company generates £20 million through a commercial process and AI contributes to a 1% improvement in the economic outcome.
The technology might appear to have done very little.
The business may care considerably more.
These figures are deliberately hypothetical. Real-world outcomes are rarely this clean, and attributing an improvement directly to AI can be difficult.
But the principle is straightforward:
A small percentage of a large economic number can be worth more than a large percentage of a small one.
That sounds obvious when written down.
I’m not sure businesses always behave as though it is.
Frequency can turn tiny improvements into large ones#
Scale doesn’t only come from money.
It can come from repetition.
Imagine an activity that occurs 500,000 times each year.
AI saves three minutes each time.
Three minutes barely sounds worth discussing.
Repeated 500,000 times, it represents 1.5 million minutes.
That’s 25,000 hours.
Again, this doesn’t mean the business has suddenly “saved” the financial equivalent of 25,000 employee hours.
Released time only becomes economic value if the organisation can actually capture it — an important distinction when thinking about AI as more than a cost-cutting technology.
Perhaps fewer resources are required.
Perhaps existing employees can handle greater volumes.
Perhaps service becomes faster.
Perhaps the time simply gets absorbed elsewhere and produces little measurable benefit.
That’s an important distinction.
But frequency has transformed the scale of the opportunity.
The same mechanism appears throughout businesses.
A slightly better classification repeated across millions of records.
A marginally faster response repeated across every customer enquiry.
A small reduction in errors across thousands of orders.
A few minutes removed from an activity performed by hundreds of employees every week.
A small improvement in forecasting repeated across many purchasing decisions.
Individually, each improvement can look almost irrelevant.
Collectively, they can become significant.
AI opportunities should therefore be evaluated at the scale at which their effects accumulate, not merely at the scale at which an individual interaction occurs.
Decisions can compound too#
The same principle applies to decision quality.
I wrote previously about how some of the most economically important decisions inside businesses are almost invisible.
Consider a salesperson deciding which opportunity to pursue next.
AI doesn’t need to double their effectiveness to create value.
Perhaps better information makes their prioritisation slightly better.
One decision probably doesn’t matter.
But the salesperson makes hundreds of them.
The sales team makes thousands.
Across the organisation, those decisions influence how a substantial amount of expensive commercial capacity gets allocated.
A small improvement in prioritisation can therefore propagate into something much larger.
The same could be true of decisions about pricing, inventory, customer intervention, fraud, purchasing, maintenance, scheduling or resource allocation.
This doesn’t mean every marginal improvement can be measured reliably.
It can’t.
Nor does it mean a 2% improvement in some model metric automatically becomes a 2% improvement in a business outcome.
It won’t.
There are usually multiple steps between better analysis and better economics.
But it does mean we shouldn’t dismiss an AI opportunity merely because the improvement appears modest.
The right question is:
What happens when that improvement is repeated across the entire economic surface of the decision?
Revenue makes small improvements interesting#
Revenue provides perhaps the easiest illustration.
Imagine a business with £10 million of annual sales.
It doesn’t need AI to double its revenue for the technology to matter.
Suppose AI helps improve one part of the commercial system.
Lead identification.
Qualification.
Sales research.
Proposal development.
Pricing.
Cross-selling.
Retention.
Customer communication.
Maybe the improvement is tiny.
But commercial systems contain multiplication.
More relevant opportunities can create more conversations.
Slightly better conversations can improve conversion.
Improved conversion can create more customers.
Better retention can mean those customers remain for longer.
Higher customer value can then increase the return on acquiring them in the first place.
This is why apparently small changes in a commercial system can sometimes matter disproportionately.
Not because AI possesses some magical compounding property.
Because business economics already compound.
AI only has to influence one of the variables.
The economic leverage often exists before the AI does. The opportunity is finding where AI can influence it.
That distinction matters.
We’re not looking for technology capable of creating enormous value from nothing.
We’re looking for places where the business already has scale, frequency or leverage and asking whether AI can improve what happens there — the same economic perspective behind starting AI opportunities with the economics of the business.
Margin can make a small number much bigger#
Revenue isn’t the only large economic surface.
Margin can be particularly sensitive to relatively small changes.
Imagine a company with £20 million of revenue and £1 million of operating profit.
A relatively small improvement in pricing, purchasing, waste, utilisation or cost-to-serve could have a disproportionate effect on profit if much of the additional value reaches the bottom line.
Again, the actual economics will depend entirely on the business.
But it highlights why percentages can be misleading.
A 1% improvement in one metric may be irrelevant.
A 1% improvement somewhere else may materially change profitability.
This is one reason generic lists of AI use cases are of limited use.
The same AI capability can have radically different value in two apparently similar companies because their economics are different.
One company might have excess capacity.
Another might be constrained.
One might have strong margins but weak conversion.
Another might convert brilliantly but suffer expensive customer churn.
One might lose value through operational errors.
Another might have excellent operations but poor pricing discipline.
The technology doesn’t know which of those matters most.
The economics of the individual business determine where leverage exists.
Look for bottlenecks#
Sometimes the most valuable number isn’t particularly large.
It’s constrained.
Imagine a company capable of generating more demand than it can fulfil because one specialist team is operating at capacity.
The company doesn’t necessarily need AI to transform the whole organisation.
It needs to relieve the bottleneck.
If AI increases the output of that constrained team by 10%, the impact may flow through the rest of the business.
More work gets completed.
More customers can be served.
More revenue can be accepted.
Fixed infrastructure may be utilised more effectively.
The economic value of improving the bottleneck can therefore exceed the apparent value of the activity itself.
This is another reason I would hesitate before ranking AI opportunities purely according to hours saved.
An hour removed from a non-constrained activity may create very little.
An hour released from the thing preventing the organisation from growing might be extremely valuable.
The technology has saved the same hour.
The economics are completely different.
Small improvements still have to be real#
There is an obvious danger in this argument.
Once we start applying tiny percentage improvements to very large numbers, it becomes incredibly easy to manufacture spectacular business cases.
Take £100 million.
Assume AI improves something by 2%.
Declare a £2 million opportunity.
Put it on a slide.
That’s not economic analysis.
It’s arithmetic with optimistic assumptions.
A credible AI business case needs to examine the entire chain between capability and outcome.
Will the AI actually produce the improvement?
Will employees use it?
Will customers respond as expected?
Does it work reliably enough?
What does implementation cost?
What additional processes are required?
Does improving the intermediate metric actually change the economic outcome?
Can the organisation capture the theoretical value?
How will we know whether the improvement happened?
And what happens when the system is wrong?
The smaller the expected improvement, the more important measurement can become.
If the business case depends on moving conversion from 20% to 21%, we need enough evidence to distinguish genuine improvement from normal variation.
This is where experiments, baselines and careful measurement become important.
The purpose of thinking about small improvements isn’t to justify weak AI projects with enormous spreadsheets.
It’s to make sure we don’t overlook economically important opportunities simply because the technological change looks modest.
Transformation can be quiet#
There’s something else I like about this way of thinking.
It changes what an AI-transformed business might look like.
Perhaps transformation isn’t one enormous system replacing an entire department.
Perhaps it is hundreds of small improvements accumulating across the organisation.
Salespeople know slightly more before customer conversations.
Managers retrieve information more quickly.
Customer-service employees receive better context.
Pricing becomes marginally more intelligent.
Internal knowledge becomes easier to access.
Administrative tasks take slightly less time.
Problems are identified a little earlier.
Decisions become a little more consistent.
None of those makes for a particularly dramatic demonstration.
But businesses themselves are accumulation machines.
Thousands of people perform thousands of activities, make thousands of decisions and interact with thousands of customers.
Small improvements can propagate through those systems.
Over time, the difference between an organisation that continuously finds and captures those improvements and one that doesn’t could become substantial.
This also suggests that waiting for one enormous AI project may be the wrong mental model.
The opportunity may be distributed.
The challenge is finding the improvements worth pursuing.
The difficulty is deciding which small improvements matter#
And this is where the problem becomes more interesting.
If AI can potentially improve hundreds of things inside a business, which ones deserve attention?
A 5% improvement here.
A 1% improvement there.
Ten minutes saved somewhere else.
Better information for another team.
Fewer errors in one process.
More capacity in another.
Individually, many will sound plausible.
But companies have finite capital.
Finite technical resources.
Finite management attention.
Finite tolerance for change.
And finite capacity to implement things properly.
So recognising that small improvements can create enormous value creates another problem:
There may be far more plausible AI opportunities than a business can realistically pursue.
The answer can’t be to do everything.
Nor can it simply be to choose the opportunities with the most impressive technology.
The organisation needs to understand where the economic leverage actually exists, how much of the theoretical improvement can realistically be captured and which opportunities justify consuming scarce implementation capacity.
That is a different problem from discovering AI use cases.
It is a problem of choice.
And as AI becomes capable of improving more and more parts of a business, I suspect making those choices well will become increasingly important.
Look for leverage, not spectacle#
The AI projects that attract the most attention will probably continue to be the dramatic ones.
Entire workflows automated.
New interfaces.
Sophisticated agents.
Capabilities that previously looked impossible.
Some will create enormous value.
But spectacle isn’t an economic measure.
A small improvement can matter when it affects something large.
It can matter when it happens frequently.
It can matter when it improves an important decision.
It can matter when it affects margin.
It can matter when it removes a constraint.
And it can matter when several improvements accumulate through the economics of a business.
This gives leaders a useful question to ask when evaluating AI:
If this gets only slightly better, what is that improvement actually worth?
Sometimes the answer will be almost nothing.
That’s useful.
Sometimes it will be surprisingly large.
That’s where things get interesting.
Because the most valuable AI opportunity in your business may not be the one that changes everything.
It may be the one that changes something important, just enough, thousands of times.