‍The hidden costs behind the promise of AI

A global retailer decides to accelerate its international content production with AI. Product descriptions that once took days can now be generated in minutes. Campaign assets can be adapted across markets almost instantly. Translation costs fall, production capacity rises, and the business appears to have found exactly what AI promised: more output, at lower cost, with greater speed.

Then the hidden work begins.

Someone has to check whether the content is actually correct. Someone has to validate terminology market by market. Someone has to fix the data feeding the models. Someone has to connect the AI tools to existing platforms. Engineers have to build integrations and exceptions. Teams create prompts, workflows and controls around systems that were never designed to work together. And when the output is wrong, someone still has to fix it.

The visible cost of AI may be falling. The cost of making AI work may be rising.

That is the hidden cost of AI transformation.

The productivity gap

The first warning sign is already visible in the data.

McKinsey’s latest research shows that 80% of respondents say AI is improving individual productivity, yet only 37% say it is contributing positively to enterprise earnings.

There is a simple reason for this apparent contradiction: improving a task is not the same as improving a system.

An organization does not create value through isolated tasks. Value is created through the entire chain surrounding those tasks: data has to move between systems, decisions have to be made, outputs have to be reviewed, exceptions have to be handled and the final result has to reach a customer or another part of the business.

AI can make one employee dramatically faster while leaving the surrounding process unchanged. The output still has to be reviewed, reformatted, approved, uploaded, localized, reconciled with another system and sometimes corrected manually. What disappears from one step can simply reappear somewhere else.

This is not unique to AI. McKinsey’s research on transformation more broadly found that even organizations considered successful captured, on average, only 67% of the maximum financial value available from their transformations.

The lesson is important: value is rarely lost inside the technology itself. It is lost between the technology and the organization around it.

When access replaces expertise

AI has also created a new illusion: if everyone can use the technology, perhaps everyone can do the work.

A marketer can generate copy. A buyer can analyze data. A project manager can create a translation. An engineer can ask an AI to write code. The barrier to producing an answer has fallen dramatically.

But the barrier to producing a reliable business outcome has not.

Knowing how to ask an AI to produce content is not the same as understanding brand governance, local market expectations, regulatory requirements or the commercial role of that content. Generating code is not the same as designing a robust architecture. Producing a translation is not the same as managing terminology, context, workflows and accountability across dozens of markets.

The danger is not that AI eliminates expertise. It is that organizations begin to hide expertise inside rework.

Specialists are then pulled downstream into validation, correction and exception handling. Engineers spend time repairing integrations rather than creating new capabilities. Operations teams become the safety net for processes that were never properly redesigned.

The organization may therefore appear more automated while quietly accumulating a new layer of invisible labor.

The cost behind the agent

The same phenomenon is emerging with AI agents.

An agent may appear inexpensive when measured by the cost of a single interaction. But tokens are only one component of the economics. There is also inference, infrastructure, integration, monitoring, permissions, data preparation, security, human oversight and the cost of handling the cases where the agent does not behave as expected.

KPMG’s 2026 research found that only 26% of organizations have real-time visibility into AI operating costs, while 35% cite gaps in understanding AI economics, including token and inference pricing, as a significant barrier.

That should change the question companies ask.

The relevant metric is not simply cost per token or even cost per AI task.

It is cost per validated business outcome.

An AI-generated product description that costs almost nothing but requires three rounds of review, manual correction and platform intervention may be considerably more expensive than a controlled process that initially appears less automated.

Conversely, a more structured AI workflow may appear more expensive at the technology layer while producing a significantly lower total cost once the entire process is taken into account.

The same principle applies to agents. An agent that performs one task efficiently but creates five new handoffs can reduce the cost of the task while increasing the cost of the process.

Complexity is where value leaks

This is why AI transformation increasingly resembles what happened in other complex operating environments.

In US logistics, McKinsey has estimated that inefficient “blind handoffs” between companies and systems generate between $65 billion and $95 billion of waste every year.

The underlying problem is not transportation itself. It is what happens between participants.

AI introduces new interfaces between people, models, agents, data, platforms and business processes. Every interface creates the possibility of duplication, delay, error, monitoring and accountability gaps.

And the more AI tools an organization adds without redesigning the underlying architecture, the more difficult it becomes to understand where the value, and the cost, is actually being created.

Because AI makes individual tasks so inexpensive and accessible, it becomes easier to automate more of them without asking whether the resulting process is actually simpler. A company can end up with multiple models, agents, prompts, workflows and monitoring layers operating around an underlying process that has barely changed.

This is the paradox of AI transformation: automation can reduce the cost of individual actions while increasing the complexity of the system.

The real transformation

The answer is not to use less AI.

It is to stop measuring transformation by the number of tasks that have been automated.

The more important question is: if we were designing this operation today, with AI already available, would we design it this way?

That question moves the discussion from tools to operating models.

It forces organizations to reconsider who does what, where expertise belongs, how data flows, which systems should remain, which should disappear, where humans add value, and how AI capabilities should be orchestrated across the entire process.

This is particularly important in environments such as global content, e-commerce and localization, where thousands of individual decisions interact across markets, languages, platforms and channels. The opportunity is not simply to generate more content with AI. It is to redesign the infrastructure through which that content is created, adapted, governed and activated.

The real cost of AI transformation is therefore not the price of the model.

It is the accumulated cost of everything an organization has to build around the model when the transformation stops at automation.

The companies that capture the greatest value from AI will not necessarily be those deploying the most models, agents or tokens. They will be those capable of redesigning the system around them, and making complexity disappear rather than simply moving it somewhere else.

That is where AI transformation becomes more than automation.

It becomes transformation.