The AI Sustainability Paradox:
Why Leadership Means Living with the Tension

Brand Provided | By Franco Amalfi Antonio Vizcaya Abdo Sandra Leyva and Capgemini | Published 3 hours ago | About a 7 minute read

Image: Photo by Berke Citak on Unsplash

What if the biggest sustainability risk of artificial intelligence is not only the energy it consumes, but the work it makes so easy to create?

When people hear “sustainable AI,” the conversation usually zooms in on data centers, electricity, water and compute because AI may feel digital at the point of use, but the infrastructure behind it is deeply physical. That focus has made AI’s environmental footprint much more visible, but it can also make the sustainability question seem as if it begins only after a use case already exists.

In our work on this topic, we have found ourselves going further upstream and asking a different question: if AI is supposed to make organizations more efficient, are we actually eliminating work, or are we simply creating more of it because now we can? That changes the conversation from how efficiently we run AI to what we are asking it to produce in the first place.

Sustainability leaders have spent decades helping organizations do more with less, and while AI appears to promise exactly that, it also changes the economics of creating work. When the cost and friction of producing another analysis, deck, summary, draft or line of code collapse, the natural result is not necessarily less work. It can be more, creating a sustainability paradox in which individual tasks become more efficient while the total amount of work expands.

When Efficiency Creates More Work

When we use AI, each prompt facilitates an additional analysis, another deck, another email, another summary.

Individually, each one may take seconds instead of hours, which is a real productivity gain. But an output is not the same thing as value, and generating it faster does not remove the work that can follow. A report may still trigger review, revision, meetings, storage, decisions and additional work downstream, while a faster workflow may still be a workflow that should have disappeared.

We started calling this work inflation: the expansion of work because the marginal cost of creating it has become so low.

Work inflation can happen even while individual tasks become more efficient. One person may save an hour, while the organization creates ten new things for other people to review, interpret or act on.

This is why measuring AI productivity only through output can be misleading. A more useful measure is whether work disappears, becomes reusable or leads to better decisions rather than simply increasing the volume of activity moving through the organization.

When Work Inflation Becomes Organizational Waste

What looks like work inflation at the individual level becomes organizational waste when it scales across hundreds or thousands of people.

Five teams can solve the same problem independently, pilots can be launched without visibility into what already exists, tools overlap and teams generate similar content for slightly different audiences. A process that should be redesigned or eliminated can simply become faster because AI was layered on top of it.

When those patterns accumulate, the waste is not the use of AI itself. It is the duplication, rework and activity that consume resources without compounding into capability.

Experimentation is also not the problem; some experiments should fail because that is how organizations learn. Waste appears when the learning does not cascade, useful work is not reused, duplication is not reduced and low-value use cases continue because there is no mechanism to stop them.

One useful way to understand why this happens is to separate AI adoption from AI absorption.

Adoption tells us that the technology entered the organization through licenses, users, pilots and use cases.
Absorption tells us whether the organization changed because of it: whether work disappeared, one team’s learning became another team’s starting point, a process improved permanently or a low-value use case stopped.

When adoption moves faster than absorption, AI can sit on top of the existing operating model and accelerate it rather than transform it. The organization may look advanced while quietly accumulating more tools, more outputs and more activity without building much more capability.

That gap is critical for sustainability because every additional workload ultimately draws on compute, electricity, water, infrastructure, specialized chips, capital and human attention. The point is not that every additional AI workload is wasteful, but that the resources behind it should create enough durable value to justify what they consume.

From Footprint to Capability Yield

Once organizational waste is visible, measuring only AI’s footprint is no longer enough because two applications can consume similar resources and create completely different outcomes.

One might improve renewable-energy forecasting, optimize a logistics network or make scarce expertise available to people who previously could not access it. Another might generate content that nobody reads and that disappears into a folder a week later.

What differentiates those use cases is not only what they consume, but what the organization is able to do differently because of them.

We have started using the term Capability Yield to describe the lasting capability created from the compute, energy, water, infrastructure, capital and human attention invested in AI. The word “lasting” matters because the value should survive the individual prompt or output.

Did a person become more capable?
Did a team create knowledge it can reuse?
Did the organization improve a process beyond a single task?
Did an insight transfer?
Did a capability that used to be scarce become accessible to more people?

A high-yield use case compounds over time because the organization retains something useful after the immediate output is gone. A low-yield use case may produce something impressive in the moment but leave very little behind.

We are still working through what Capability Yield should look like as a measurement framework because AI is evolving faster than many of the frameworks around it. A perfect metric is not a prerequisite for recognizing that activity and value are not the same thing. The direction is already useful; connect what AI consumes with what the organization actually becomes capable of doing because of it.

Building the Filter Before the Footprint

For sustainability professionals wondering where to start, the first step does not need to be another framework. It can be much more practical, such as beginning to understand where AI decisions are already being made across technology, operations, legal, risk, security and procurement, and begin bringing these questions into those existing processes. That approach is most useful upstream, when organizations are deciding what deserves to be built and scaled.

Does this work need to exist?
Is AI necessary for the problem?
What type of AI is sufficient?
What will this use case replace rather than simply add?
What becomes reusable?
What capability remains after the task is complete? When should a use case scale, change or stop?

Those questions begin to connect AI adoption with organizational waste, resource consumption and Capability Yield without requiring sustainability to own the AI agenda.

The more the three of us have worked through this topic, the more convinced we have become that sustainable AI will not be solved by looking at any one of those questions in isolation. Its footprint is really important, but so does the work we choose to create.

Productivity matters, but so does whether that productivity compounds across the organization. Adoption matters, but only if the organization develops the ability to absorb what it is adopting.

Sustainability professionals may not own every one of those decisions, but we work at the intersection of many of the systems they affect, which gives us an important opportunity to connect what can otherwise remain separate. Perhaps that is the larger role for sustainability in this next chapter of AI; not to become the owner of the technology, but to help organizations become more deliberate about what they build, what they scale, what they stop and ultimately whether all of that investment is making the organization more capable or simply creating more.