AI value starts when the workflow changes, not when the tool is launched

Many organizations are experimenting with AI.

People use ChatGPT. Someone tries Microsoft Copilot. A team creates a prompt library. A manager asks for use cases. A workshop is organized.

Everyone agrees AI is important.

That is a good start.

But it is not yet value.

AI adoption is growing quickly. McKinsey's State of AI research shows that organizations are using AI across more business functions, and that generative AI is becoming part of everyday work.

But adoption does not tell us whether the work is actually getting better.

The more useful question is not:

"Are we using AI?"

It is:

"Is the work becoming faster, clearer, safer or easier to manage?"

That difference matters. Because AI activity is not the same as AI value.

A use case is only the beginning

Most AI conversations start with use cases.

Use AI to summarize meetings. Draft emails. Analyse documents. Prepare reports. Answer internal questions.

There is nothing wrong with these use cases.

But most of them optimize a task, not the workflow.

An employee may write an email in two minutes instead of ten. That saves time.

But did the customer receive an answer sooner? Or did the employee still need to check three systems, ask someone for approval, and correct information that was unclear in the first place?

If the surrounding workflow does not change, the benefit stays small.

That is why the better starting point is not:

"Where can we use AI?"

It is:

"Why does this work take this long in the first place?"

That leads to a very different conversation.

Measure the workflow, not the excitement

The first phase of AI adoption often creates a lot of energy.

People show demos. They share prompts. They discover small tasks that can suddenly be done much faster.

That matters. Experimentation is necessary.

But eventually the organization needs to look beyond activity.

Where is time actually being removed from the process? Which handover has disappeared? Which decision can now be made earlier? Where is rework being reduced? Which information becomes more reliable?

Those are better indicators of AI value than the number of licenses, prompts or pilots.

This is particularly relevant for smaller Aruba-based organizations. Teams are already busy. People often wear multiple hats. There is limited capacity for experiments that look interesting but create another layer of work.

AI should reduce friction in work that already matters. Not create a parallel AI programme next to it.

A concrete example: monthly reporting

Take a monthly reporting process.

A logical first AI use case might be:

"Let AI write the management commentary."

That can save time.

But in many organizations, writing the commentary is not what makes reporting slow.

The real delay sits earlier in the process.

Inputs arrive late. Numbers are copied between spreadsheets. Definitions are interpreted differently. One person knows which file is actually final. Explanations are rewritten every month. And management receives the report after the moment when some decisions should already have been made.

Adding AI to the writing step does not fix that.

The better approach is to look at the reporting workflow first.

Where does it start? Which sources are trusted? Who owns each input? Where does the team wait? Which checks are repeated every month? Which steps exist because the process has simply grown that way over time?

Only then does it become useful to decide where AI can help.

AI might summarize changes, identify missing inputs, compare versions, draft recurring explanations, flag inconsistencies or prepare decision points for management.

Now AI is not only making one task faster. It is helping shorten the reporting cycle.

That is a different level of value.

EY makes a similar distinction in its AI Value Accelerator, connecting AI investments to tangible business outcomes such as productivity, cost improvement, better use of assets and revenue opportunities.

The point is not to prove that AI produced something. The point is to prove that the process improved.

Why working AI pilots still fail

A lot of AI pilots look convincing.

The demo works. The output looks good. The team can clearly see the potential.

And yet six months later, very little has changed.

Reporting around the MIT GenAI Divide study described the same problem: many enterprise AI pilots struggle to produce measurable business value when AI is not integrated into real workflows, data and organizational learning. That coverage is worth reading with some caution around the headline numbers, the underlying methodology has drawn real scrutiny, but the underlying point still holds.

The technology can work while the organization remains unchanged.

That is the real issue.

Someone still needs to own the output. Someone needs to decide when it is good enough. The organization needs to know which data AI may use, where human judgement remains necessary, and what happens when the output is wrong.

If nobody owns the output, the decision and the exception, AI is still sitting next to the process rather than becoming part of it.

This is also why AI governance should not only be discussed at policy level. Frameworks such as ISO/IEC 42001 emphasize clarity around scope, roles, risk management and continual improvement.

Those principles become practical when they are connected to an actual workflow.

Three questions before scaling AI

Before scaling an AI initiative, I would start with three questions.

1. Which workflow are we improving?

Do not start with "AI across the organization."

Start with a workflow where friction is already visible. Look for work that happens frequently, involves several people, depends on information from different places, or creates repeated waiting and rework.

Reporting, onboarding, service requests, compliance checks, internal knowledge search and proposal preparation are obvious examples.

The specific workflow matters more than the number of AI use cases.

2. What should improve?

Be specific.

Turnaround time. Data quality. Manual rework. Preparation time. Decision speed. Consistency. Dependency on one person.

If you cannot describe what should improve, it will be very difficult to prove later that AI created value.

3. Who owns the result?

AI does not remove ownership. It makes ownership more visible.

Who is responsible for the final output? Who decides whether it is good enough? Who owns exceptions? What remains human judgement? What happens when the AI-supported process produces a wrong result?

These questions do not require a large governance programme. But they do require an answer.

This builds on an earlier SciSpecs point: AI enablement should start with the architecture of the work, not with a long list of disconnected AI use cases.

AI value needs more than a tool

AI value appears when several things come together.

You need to understand how the workflow, data, systems and ownership connect. You need to measure whether the work is actually improving. And people need to adopt the new way of working in practice.

This is where many AI initiatives become uncomfortable.

Because the difficult part is often no longer the AI. It is changing the way work is organized.

EY recently went a step further: this summer they created an internal AI Value Realization Office, a dedicated team whose entire job is tracking whether the firm's own AI investments are producing measurable business impact, not just activity.

Organizations do not necessarily need a value realization office to get started. But they do need the discipline behind it.

What are we trying to improve? Who owns it? How will we know whether it worked?

Start small, but measure honestly

You do not need a large AI transformation programme to begin.

Take one workflow that matters. Make the current friction visible. Decide what should improve. Introduce AI where it improves the flow, not simply where it produces an impressive demo. Clarify ownership. Then measure what changed.

If the reporting cycle went from ten days to six, that matters. If a service team can resolve requests without switching between five sources, that matters. If managers receive reliable information before a decision instead of after it, that matters.

If nothing meaningful changed, that matters too. You learned before scaling the wrong approach.

That is the shift organizations need to make.

From AI use cases to workflow improvement. From adoption metrics to business outcomes. From asking where AI can be added to asking how the work should actually be done.

Because the goal is not to become an organization that uses AI.

The goal is to become an organization that works better.

If you want to see where the friction actually sits in your own organization before you scale anything, take the free EA Quickscan, it takes about two minutes and gives you a starting point, not a sales pitch.