Where AI Actually Creates Business Value

Imagine an organization receives hundreds of documents every week.

Employees open them.

Read them.

Identify relevant information.

Copy that information into another system.

Check it against a policy.

Route the document to the correct team.

Write a summary.

Someone else reviews the summary.

From an AI perspective, there are dozens of exciting things we could build.

But from a business perspective, the interesting questions are much simpler.

Those questions change the AI conversation.

Instead of beginning with a model and searching for a use case, we begin with the work and determine which capability belongs inside it.

Sometimes that capability will be AI.

Sometimes it will not.

It is tempting to talk about AI transforming entire professions.

But organizations do not actually operate in professions.

They operate through thousands of tasks.

A financial analyst may investigate anomalies, prepare reports, reconcile data, interpret performance, communicate findings and advise leadership.

A customer-support agent may retrieve information, understand a customer's problem, navigate policy, communicate a resolution and recognize when an unusual situation requires escalation.

A manager may summarize information, coordinate people, make decisions, resolve ambiguity and exercise judgment.

AI may be extremely useful for some of those tasks and poorly suited to others.

This is increasingly visible in empirical research.

In a large real-world study involving 5,172 customer-support agents, Brynjolfsson, Li and Raymond found that access to a generative AI assistant increased productivity by an average of 15% in that particular setting. Importantly, the effects were not uniform: less experienced and lower-skilled workers experienced substantially greater improvements than the most experienced workers.^[1]

The implication is not that AI improves customer service—or knowledge work generally—by a universal percentage.

It is almost the opposite.

It depends on the task.

The worker.

The workflow.

The information available.

And the capability of the technology relative to the problem being solved.

This becomes even clearer when we look at knowledge work.

Research by Dell'Acqua and colleagues describes a jagged technological frontier: AI can substantially improve performance on some tasks while worsening it on others, even when those tasks appear to require similar levels of difficulty. Their field experiment involved 758 knowledge workers performing realistic consulting tasks.^[2]

That creates a very different implementation problem from traditional software.

With conventional automation, organizations usually define relatively deterministic rules.

If X happens, do Y.

Generative AI operates differently.

Its capability boundary can be difficult to see.

A model may produce an excellent synthesis of a complex document and then confidently make an error on a seemingly straightforward reasoning task.

It may draft a strong first version of a report while inventing a supporting fact.

It may retrieve the right policy but interpret an unusual exception incorrectly.

That means successful AI implementation requires more than identifying something the technology can do.

Organizations also need to understand:

Some of the most promising AI opportunities are surprisingly ordinary.

They begin with information.

An employee knows that an answer exists somewhere inside the organization.

They just cannot find it.

A policy lives in one system.

Process documentation lives somewhere else.

Historical decisions are buried in email.

Technical knowledge sits inside tickets.

Customer information is distributed across applications.

Someone who has worked at the company for ten years knows how everything fits together.

Someone who joined three months ago does not.

This creates what we might call information friction.

People spend time:

Generative AI can be particularly useful here because much organizational knowledge exists as unstructured language rather than clean database rows.

An internal knowledge assistant, for example, might not need to make a decision.

Its value could simply be helping someone find the information required to make one.

That distinction matters.

A useful AI system does not have to replace expertise.

Sometimes its job is to make expertise easier to access.

Traditional automation became extraordinarily good at repetitive procedural work.

Generative AI expands the range of automation into some repetitive cognitive work.

Consider:

These tasks require language and interpretation, which historically made them difficult to automate with conventional rules.

That does not mean the entire workflow should become autonomous.

Often the stronger design is:

Research on AI and management has similarly challenged the idea that organizations must choose neatly between automation and augmentation. Raisch and Krakowski argue that the two are interdependent rather than simple alternatives.^[3]

A system may automate one part of a workflow while augmenting the human performing another.

The interesting design question becomes:

What is the right division of work?

Many organizations have people who become unofficial infrastructure.

Everyone asks them questions.

They know how a process really works.

They remember why a decision was made.

They know which policy applies to an unusual situation.

They know where the right document lives.

They know the exception nobody wrote down.

This expertise is enormously valuable.

It is also difficult to scale.

AI creates an interesting possibility here.

Not replacing the expert.

Distributing access to what the organization already knows.

The customer-support study by Brynjolfsson and colleagues is particularly interesting in this context. The researchers found evidence consistent with AI helping disseminate practices associated with stronger-performing workers, with larger gains accruing to newer and less experienced agents.^[1]

That suggests a potentially important form of organizational value.

AI may sometimes function as a mechanism through which useful patterns, knowledge and practices become more accessible across the workforce.

That is different from simply reducing headcount.

It is closer to reducing the distance between knowing something somewhere in the organization and being able to use it where the work happens.

AI can also help when the constraint is not a lack of information but an excess of it.

A person may need to consider:

Humans remain responsible for the decision, but the cost of processing the information limits the quality or speed of that decision.

This is where augmentation becomes particularly interesting.

Jarrahi's work on human-AI symbiosis in organizational decision-making argues that AI and humans bring different strengths to complex decisions. AI offers computational information-processing capability, while humans remain particularly important where uncertainty, ambiguity and holistic judgment matter.^[4]

That suggests a useful principle:

Use AI to expand the information a human can practically consider—not automatically to remove the human from the decision.

An executive may not need AI to decide what strategy the company should pursue.

But AI might help synthesize thousands of customer comments before that decision.

A procurement professional may still make the supplier decision.

But AI might identify unusual clauses across hundreds of documents.

A manager may still decide how to respond to an operational problem.

But AI might surface patterns that would otherwise require hours of analysis.

The decision remains human.

The information boundary around the decision changes.

This may be one of the most important AI capabilities an organization can develop:

knowing when not to use it.

Imagine employees manually copy customer information from one application into another.

You could build an AI agent that reads the first application, interprets the information and enters it into the second.

Or you could connect the systems with an API.

The API may be cheaper.

More predictable.

More testable.

And considerably easier to govern.

Imagine leadership cannot trust its revenue dashboard because three departments calculate revenue differently.

An AI assistant could answer questions about revenue.

But it would be answering them on top of unresolved definitions.

The organization does not have an AI problem.

It has a data-governance problem.

Or imagine employees spend hours generating a report because the underlying information must be manually consolidated every month.

The solution may be a better data pipeline.

Not generative AI.

Modern technology does not mean using the newest technology for every problem.

AI systems consume context.

That context has to come from somewhere.

If those foundations are fragmented, contradictory, outdated or poorly governed, AI does not magically make them reliable.

It can make the underlying problem harder to see.

A polished conversational interface may create the impression of a unified intelligent system while the information underneath remains inconsistent.

Research on organizational AI capabilities similarly suggests that business value depends on more than access to AI technology. Mikalef and Gupta conceptualize AI capability as a combination of resources and organizational capabilities rather than a technology asset in isolation.^[5]

Other empirical work on machine-learning business value has identified factors such as analytics maturity, platform maturity and top-management support among the conditions associated with value creation.^[6]

The pattern should sound familiar.

The technology matters.

So does everything around it.

AI discussions frequently collapse into one metric:

time saved.

Time matters.

If an employee completes a task in twenty minutes instead of an hour, that can be valuable.

But organizations should ask what happens to the forty minutes.

Business value requires a connection between the productivity improvement and an organizational outcome.

The Brynjolfsson study provides strong evidence that generative AI can improve productivity in an actual workplace, but it also demonstrates why context matters: the measured effect differed substantially according to worker experience and skill.^[1]

The question should never stop at:

Did AI make this task faster?

It should continue:

What did making this task faster allow the business to do better?

There is another assumption worth challenging.

If humans are capable and AI is capable, combining them must produce the best result.

Not necessarily.

Human-AI collaboration introduces its own problems.

People may overtrust AI.

They may ignore correct recommendations.

They may struggle to determine when the system is outside its capability boundary.

They may spend so much time verifying outputs that the efficiency gain disappears.

They may become less attentive because the system usually works.

Raisch and Krakowski describe the relationship between automation and augmentation as a paradox precisely because organizations cannot assume one simple model of human-AI collaboration will always dominate.^[3]

The evidence around the jagged technological frontier makes the operational implication even clearer: AI-assisted performance can vary substantially depending on whether a task sits inside or outside the technology's capability frontier.^[2]

Human oversight is not a design strategy by itself.

Organizations have to decide what the human is actually responsible for.

If the answer is simply "a human will check it," the workflow probably isn't finished being designed.

There is a temptation to respond to AI's potential with an equally large transformation programme.

Enterprise AI strategy.

Organization-wide copilots.

Autonomous operations.

AI everywhere.

There is another approach.

Find one meaningful constraint.

Understand the workflow.

Establish the baseline.

Identify where AI might change the economics or quality of the work.

Build narrowly.

Measure what happens.

Learn where the technology fails.

Improve the surrounding process.

Then expand.

This matters because AI capability is not simply something an organization purchases.

It is something an organization learns how to use.

Research examining AI capability and organizational performance similarly treats AI as a bundle of technical and organizational resources that must be developed and orchestrated.^[5]

Research by Zebec and Indihar Štemberger provides an especially useful extension to this argument. Studying 448 EU organizations using AI, they found that AI adoption contributes to organizational performance through decision-making and business-process performance, with process automation, organizational learning and process innovation acting as significant complementary mechanisms.^[7]

That distinction matters.

Buying or deploying AI is not the same as creating value from it.

Organizations still have to learn how to redesign processes, build knowledge and translate the technology into better ways of working.

The first implementation therefore has another purpose beyond its immediate ROI.

It teaches the organization where its own AI frontier lies.

Before approving an AI initiative, ask:

If those questions cannot be answered, the organization may have an AI experiment.

It does not yet have an AI business case.

There is a final reason not to frame AI purely as a cost-reduction technology.

The most interesting applications may eventually be things organizations simply could not do economically before.

That is where the conversation moves from efficiency to capability.

And capability is ultimately where technology becomes strategically interesting.

The question is no longer:

How many hours can AI remove?

It becomes:

What can this organization now do that it could not reasonably do before?

The organizations that create meaningful value from AI may not be the ones that deploy the most AI.

They may be the ones that become best at distinguishing between:

AI can automate.

AI can augment.

AI can retrieve.

AI can synthesize.

AI can generate.

AI can help people navigate complexity.

Those capabilities are extraordinary.

But none of them is inherently valuable.

Value appears when a capability meets a real organizational constraint.

So perhaps the first question in an AI strategy should not be:

Where should we use AI?

It should be:

Where is the business currently working harder than it should?

Start there.

Then decide what technology belongs in the answer.