The Distance Between AI Capability and AI Value Is an Implementation Problem

Something strange is happening with AI.

The technology keeps getting better.

The list of things businesses can reliably ask it to do keeps getting longer.

Models can generate increasingly sophisticated content, analyse information, write software, work through complex problems and, increasingly, use tools to take actions.

Yet the economic impact inside many organisations appears considerably less dramatic.

This isn’t because AI has created no value.

It clearly has.

People are using it to save time, increase capacity, analyse information and perform tasks differently.

But there is an increasingly visible gap between what the technology appears capable of doing and the value organisations are actually capturing from it.

McKinsey’s latest global survey illustrates the problem.

Seventy-one per cent of respondents said their organisations were regularly using generative AI in at least one business function.

Yet more than 80 per cent said their organisations weren’t seeing a tangible impact on enterprise-level EBIT from that use.

Those numbers shouldn’t be interpreted too precisely. Self-reported surveys have limitations, and organisation-wide financial impact is a demanding measure for a technology this young.

But the direction is interesting.

AI adoption is spreading faster than measurable enterprise value.

That suggests the important question may be changing.

For the last few years, businesses have reasonably asked:

Can AI actually do this?

Increasingly, another question matters:

Can we turn what AI can do into something that works reliably inside our business?

Because those are not the same thing.

Technical capability is only the beginning#

Consider what happens when a new frontier model is released.

The announcement usually focuses on capability.

The model can write better code.

Follow more complicated instructions.

Reason through harder problems.

Process more information.

Use tools more reliably.

Perform better on benchmarks.

These improvements matter.

In April, OpenAI released GPT-4.1 with substantial improvements in coding, instruction following and the ability to work across very large amounts of context.

In February, Anthropic released Claude 3.7 Sonnet, describing significant improvements in areas including coding and reasoning.

The frontier continues moving.

But imagine tomorrow’s best model becomes another 20 per cent better at a task that matters to your company.

What happens on Monday morning?

Possibly nothing.

The model might possess the capability.

But it still doesn’t automatically know which problem in your organisation deserves attention.

It doesn’t necessarily have access to the relevant information.

It doesn’t understand every part of your operating context.

It isn’t automatically integrated with your systems.

It hasn’t redesigned the workflow around itself.

Employees don’t necessarily know how to use it.

Nobody may own the implementation.

And the organisation may not even have decided how success will be measured.

A capability can exist without an organisation possessing the ability to exploit it.

That distinction is becoming increasingly important.

A demonstration proves possibility, not value#

This is one reason AI demonstrations can be misleading.

You watch a model perform a task that previously appeared impossible.

Perhaps it analyses a complicated document.

Builds an application.

Investigates a problem.

Uses a computer.

Produces a sophisticated piece of work in seconds.

The natural reaction is:

If AI can do that, imagine what it could do inside our business.

That’s a perfectly reasonable reaction.

But there are several steps hidden inside the word could.

Can the model perform the task with our information?

Can it perform it with sufficient reliability?

Can it handle the exceptions that occur in our process?

Can it work with our existing systems?

Can employees use it effectively?

Can we manage the risks?

Can we deploy it at an acceptable cost?

Can we change the surrounding process?

And if we do all of that, will the resulting improvement actually be worth anything?

The demonstration answers one question:

Is this technically possible?

The business case asks another:

Can this capability reliably improve an economic outcome inside our organisation?

The distance between those questions is where much of the difficult work lives.

The bottleneck can move#

This leads to a useful way of thinking about technological progress.

When a technology is immature, the technology itself is often the constraint.

Perhaps the model simply can’t perform the task accurately enough.

No amount of clever implementation fixes that.

You wait.

Then the technology improves.

Eventually it crosses some threshold where the task becomes technically possible.

At that point, the bottleneck moves.

Perhaps the model is now capable enough, but the organisation’s data isn’t accessible.

Solve that and the bottleneck moves again.

Now integration is the problem.

Then workflow design.

Then reliability.

Then employee adoption.

Then management.

Then measurement.

This matters because businesses can continue behaving as though they’re waiting for better AI when the thing preventing value is no longer the AI.

A new model release won’t redesign your sales process.

A larger context window won’t decide who owns AI internally.

Better reasoning won’t clean your data.

An agent won’t automatically know which actions it should be authorised to take.

A lower hallucination rate won’t persuade employees to change how they work.

And another benchmark improvement won’t tell you whether any of this has increased revenue or margin.

As AI capability improves, the constraint on value can migrate from the model into the organisation.

That may be one of the most important shifts for business leaders to recognise.

Implementation is much bigger than integration#

The word implementation can sound disappointingly technical.

Connect the API.

Integrate the systems.

Move the data.

Configure the software.

Deploy it.

Done.

But AI implementation is broader than technical integration.

Imagine a company identifies a promising opportunity to use AI to improve sales qualification.

The model capability might be excellent.

But creating value could require understanding what actually makes an opportunity valuable.

Identifying which information predicts that value.

Making the relevant data accessible.

Connecting information from several systems.

Determining how AI should evaluate it.

Testing whether the recommendations are sufficiently reliable.

Changing how salespeople prioritise their time.

Explaining why the process is changing.

Deciding when employees should override the system.

Monitoring whether people actually use it.

Measuring whether conversion, sales productivity or another economically meaningful outcome improves.

The AI model is one component.

The implementation is the system around it.

This is why I think we should use the word more broadly.

Implementation is the work required to turn technological capability into operational capability.

And operational capability is what gives the technology a chance to create value.

Workflow redesign may matter more than model choice#

There is already evidence pointing in this direction.

McKinsey’s March survey examined a range of organisational practices associated with generative AI and reported that redesigning workflows had the strongest relationship with self-reported EBIT impact among the factors it tested.

Only 21 per cent of respondents whose organisations were using generative AI said they had fundamentally redesigned at least some workflows.

That’s an interesting contrast.

Businesses are adopting AI quickly.

But far fewer appear to be redesigning how work happens around it.

Perhaps that shouldn’t surprise us.

Adding a tool is easy.

Changing an organisation is hard.

You can give employees access to an AI assistant in an afternoon.

Redesigning a customer-service process might require changes to systems, responsibilities, measures, controls, training and management behaviour.

Yet that’s often where the larger opportunity exists.

If an employee uses AI to save ten minutes while performing an unchanged process, some value may have been created.

If the process itself can be redesigned because AI changes what is economically or operationally possible, the value could be much larger.

This distinction becomes particularly important as agents develop.

Last month I wrote about AI moving from producing answers towards taking actions.

The more AI participates in workflows rather than simply assisting individual tasks, the more important the design of those workflows becomes.

The model doesn’t operate in a vacuum.

It operates inside a system of work.

Your business context is part of the technology#

There is another reason frontier capability doesn’t translate automatically into frontier performance.

The model doesn’t know your business particularly well.

It may understand accounting.

It doesn’t know why your finance team treats one particular customer differently.

It may understand sales.

It doesn’t know which characteristics actually distinguish your best customers.

It may understand manufacturing.

It doesn’t know the unusual constraint on one line in your factory.

It may understand contracts.

It doesn’t know why your organisation accepted a particular commercial term three years ago.

It may understand customer service.

It doesn’t know which customer is strategically important despite appearing relatively small in the CRM.

The more valuable the problem, the more likely these details matter.

This creates an interesting inversion.

As general AI becomes more capable, organisation-specific context may become more valuable rather than less.

The intelligence becomes increasingly available.

The proprietary context does not.

A business therefore has to connect general capability with its own information, processes, objectives, constraints and knowledge.

Without that connection, an extraordinary model can still produce an ordinary business result.

Adoption isn’t the soft part#

Then there are the people.

Technology projects have a habit of treating adoption as something that happens after implementation.

Build the system.

Deploy it.

Train everyone.

Move on.

AI makes that separation increasingly difficult.

The way employees interact with AI often determines how useful it becomes.

They need to understand what it is good at.

What it is bad at.

When to trust it.

When to challenge it.

How their role changes.

What they remain responsible for.

And why the organisation wants them to use it in the first place.

If employees don’t trust a system, they may ignore it.

If they trust it too much, they may stop applying appropriate judgement.

If they don’t understand how the output affects the wider process, local productivity improvements may never become organisational value.

And if AI releases capacity but nothing is done with that capacity, the business may struggle to capture any economic benefit at all.

Saving an employee five hours doesn’t automatically save the company five hours’ worth of money.

The capacity has to go somewhere.

Perhaps the employee serves more customers.

Produces more work.

Spends more time selling.

Handles additional volume.

Improves quality.

Works on higher-value activities.

Or perhaps nothing changes.

Productivity becomes value only when the organisation captures the productivity.

That makes adoption part of the economics, not a softer consideration sitting beside it.

Someone eventually has to own the outcome#

There is also an organisational question hiding inside many AI projects.

Who is responsible?

IT might own the technology.

A business function might own the process.

Finance might own the business case.

Legal might own part of the risk.

HR might own elements of employee adoption.

Data teams might own access to information.

Senior leadership might own the strategic priority.

An external provider might build part of the solution.

All of those arrangements can be reasonable.

But somebody still has to own the outcome.

Not the deployment.

The outcome.

If the objective is to increase sales conversion, who is accountable for determining whether conversion increased?

If AI is supposed to release operational capacity, who decides what happens to that capacity?

If it should improve customer experience, who measures whether customers actually experience the difference?

This is where technology projects can become detached from business performance.

The system gets implemented.

Usage gets measured.

People celebrate adoption.

But the original economic objective quietly disappears.

Using AI is not the outcome.

Creating value is.

Measurement changes what gets implemented#

McKinsey’s research also found that tracking well-defined KPIs for generative AI solutions was the adoption and scaling practice most associated with bottom-line impact in its analysis.

That makes intuitive sense.

Measurement forces clarity.

What exactly are we trying to improve?

How would we know?

What happens today?

What should change?

What is the improvement worth?

Over what period?

What else could have caused it?

Those questions can feel inconvenient during an exciting AI project.

They are also the questions that distinguish experimentation from investment.

Without them, a business can accumulate AI activity without knowing whether it is accumulating AI value.

This doesn’t mean every experiment requires a sophisticated financial model.

Early exploration has value.

Learning matters.

Some opportunities need to be tested before their economics can be understood.

But eventually, an organisation has to connect the capability to an outcome.

Otherwise the project risks becoming a demonstration that never quite becomes a business capability.

Better models won’t solve every AI problem#

None of this means model progress is becoming unimportant.

Far from it.

Better models will continue unlocking opportunities that aren’t viable today.

Reliability will improve.

Costs may fall.

Reasoning will improve.

Agents will become more capable.

New modalities will create new applications.

Tasks currently beyond AI will become possible.

But businesses should be careful about assuming the next model release will solve the problems preventing value from the capabilities they already possess.

In some cases it will.

In many others, it won’t.

The more capable the underlying technology becomes, the more important it becomes to diagnose where the actual constraint sits.

Is the model not good enough?

Or is the problem the data?

The context?

The workflow?

The integration?

The economics?

The ownership?

The adoption?

The controls?

The measurement?

These are very different problems.

And they require very different solutions.

Capability is becoming abundant. Implementation isn’t.#

There is an important competitive consequence to all of this.

Frontier AI capability is becoming widely accessible.

OpenAI releases a better model and thousands of businesses can access it.

Anthropic improves Claude and the capability becomes available to organisations around the world.

Google advances Gemini and another frontier capability enters the market.

The technology matters enormously.

But access to it is increasingly shared.

What isn’t shared is the organisation around it.

Your workflows.

Your data.

Your customers.

Your employees.

Your knowledge.

Your operating model.

Your ability to make decisions.

Your willingness to change processes.

Your ability to implement.

Your ability to measure.

Those things vary enormously between businesses.

And that means the competitive gap may increasingly emerge after the model becomes available.

Two companies can have access to exactly the same AI capability.

One experiments with it.

The other identifies where it matters, connects it to the right information, redesigns the surrounding workflow, earns employee trust, integrates it into operations and measures whether the economics improve.

Same model.

Very different outcome.

For the last few years, it has been reasonable for businesses to watch the frontier and ask what AI will become capable of doing next.

That question still matters.

But another is becoming just as important:

What are we becoming capable of doing with the AI that already exists?

Because technological capability and business value are not the same thing.

Between them sits the organisation.

And as AI gets better, the distance between AI capability and AI value is increasingly an implementation problem.