Capability Yield - The Missing Metric for Sustainable AI

The global growth of artificial intelligence represents one of the most profound economic transitions of our time. Generative AI has become a physical reality, demanding massive new investments in electrical grids, supply chains and infrastructure.

As we navigate this transition, sustainability professionals (and the public) have correctly pointed to the need to measure and reduce AI's massive resource inputs; tracking terawatt-hours, accounting for carbon, monitoring water supplies and more. However, the sheer speed of adoption has outpaced our instruments, leaving organizations to do the hard work of building measurement standards in real time while the infrastructure is already running. Getting those footprint metrics right is the bedrock of responsible growth.

But even a perfect accounting of what resources AI consumes cannot measure the operational return on that investment. It only gives us half the conversation.

In the first two parts of this series, we looked beyond the “business-as-usual” environmental footprint to trace the invisible waste driven by human behavior. We explored work inflation (how AI lowers the friction to produce, creating unnecessary outcomes like excess emails, decks, and drafts) and the gap between promoting basic adoption versus ensuring true operational absorption. Both also inadvertently drive a surge in organizational waste.

Because none of this shows up on a traditional sustainability scorecard or report, it typically goes unaddressed. But capturing that lost value starts with a question standard metrics are simply not yet built to answer:

That is why we need to talk about capability yield; a framework measuring how much durable, reusable and transferable value an organization actually generates from the compute it consumes. The environmental footprint is largely dictated by the construction already underway. The yield, however,comes from the choices we make, one workflow and one habit at a time.

Why Most Prompts Yield So Little

Today, most of that potential value is being left on the table. When MIT researchers recently found that roughly 95 percent of enterprise generative-AI pilots produced no measurable return, it spoke volumes about our deployment rather than the technology itself. We are failing to capture value because we are caught in a collision between high-pressure environments and low-context adoption.

Consider today’s reality. Reviewing a year of calendar and inbox data, one of us found 60 hours of work compressed into a 40-hour week before any strategic thinking even began. Into that pressure arrives a tool that removes the friction from production without improving judgment. Before these tools, drafting a deck or a memo took enough effort that you paused to ask whether it was worth it; now the marginal effort is nearly zero, so we simply make more. Because generating something feels like progress, we confuse output with value.

On top of that, most people (even advanced users), are drawing on only a small fraction of what these tools can do. Studies show that more than 85% of employees remain at the early stages of AI use (basic assistance and delegation), while fewer than 10% have reached the higher-value stages of semiautonomous collaboration or autonomous orchestration. We have carried over a habit from every previous enterprise technology, where software was rolled out the same way for everyone, the digital equivalent of issuing the whole company the same “stormtrooper suit”.

But AI does not reward that. It is the first technology of genuine knowledge intimacy. You were chosen for your role over hundreds of other candidates because of your specific experience and judgment, and AI is the first tool that can be tailored around exactly that (less a standard-issue uniform than a custom suit built to your own strengths). Worn generically, it returns generic output. Worn as an extension of a particular person's expertise, it compounds. 

This is where capability yield stops being a concept and becomes a practice…. (and most importantly, where sustainability professionals, so often asked "how" and "what" by colleagues who genuinely want to do the right thing, finally have something concrete to offer). The encouraging part is that the highest-yield habits and the lowest-waste habits turn out to be the same habits. 

But first one critical point that may be counterintuitive: embracing play. Yes, that thing you did effortlessly as a kid that now can feel almost awkward. Learning these tools takes play. The first few weeks are for exploring and a little extra usage while you find your footing is not waste (it is how fluency forms and people deserve that grace). After that, the guidance can start with the simplest habits, like making the tool yours before you make it work. The single largest source of wasted compute is context the model never had. 

Three features that most people skip change this entirely:

None of this is housekeeping. Unlike most technologies, what you build here stays with you and compounds as you grow; it is less a setting than an investment in yourself. It is the difference between an assistant who knows you and one you brief from scratch every single morning.

Once the tool is personalized, the rest of the “capability yield equation” comes down to putting the frameworks from our first two articles into practice. 

Bridging the Global Capability Gap

This is where the social dimension of sustainability becomes most urgent. Guided by the principle of capability yield, AI can serve as a powerful equalizer or deepen existing divides. Everything depends on whether global access becomes genuine capability or stops at consumption. 

The physical infrastructure behind AI is among the most concentrated in the modern economy. North America alone accounted for roughly a third of the AI data-center market in 2025 and of all the capacity under construction worldwide late that year, about three-quarters sat in the United States. Meanwhile, Africa (home to 18 percent of the world's population) holds less than 1 percent of global data-center capacity and a single supercomputer.

According to the International Monetary Fund's AI Readiness Index, advanced economies score more than double what low-income countries do and it compounds even further in practice.

But the trajectory is not fixed, because the talent and the appetite are already there. Sub-Saharan Africa holds the world's largest share of "innovation overperformers," countries that invent far beyond what their resources would predict; the constraint is access and not ability. In one African AI talent network of eleven thousand people, only five percent have reliable access to the computing power their work requires. It is a paradox of simultaneous abundance and scarcity (and the stakes attached to resolving it are incredibly vast). AI could add an estimated $1.2 trillion to Africa's economy by 2030, close to 6% of its GDP (but only if that capability is built locally rather than rented from abroad).

That single condition is what separates the equalizer from the divider. Building local capacity to design models for local problems, in the language and regulations and realities of the place, is capability that compounds. This is capability yield at its largest scale. This is capability yield at the largest scale.

Pointed in that direction, AI becomes the equalizer and the most promising examples already look like capability yield at national scale.

And analysts at the Center for Strategic and International Studies estimate that co-financed regional compute hubs across Africa, ASEAN and Latin America cite that pooling resources can reduce costs by as much as half compared to isolated national builds.

The principle at the macro level mirrors the individual desk. The goal is never just to distribute logins; it is to distribute the power to solve complex, systemic problems. This is terrain one of the authors of this article knows firsthand, from co-founding USAID's Global Development Lab and building technology across more than twenty crisis zones, to launching early human-machine teaming platforms for the Defense Department. The lesson held everywhere: capability that is built and owned locally outlasts capability that is merely delivered.

Moving from Mitigation to Performance

Ultimately, this framework reframes the work itself. Historically, the technology industry and the sustainability profession have approached infrastructure from opposite ends. Today, they are aligning. As developers of these models and strategists guiding corporate sustainability, we see the exact same opportunity. Engineering sustainable AI is not about halting progress. It is about channeling precious compute, energy, and water into our most meaningful work and proving a tangible return on that investment. 

For sustainability leaders, our most valuable contribution is no longer limited to auditing the footprint; it is actively guiding how our organizations use these tools. We don't need to wait for perfect global standards to start teaching our teams how to personalize a model, right-size a query and question the necessity of an output. Our job is not to police participation in a growing system, but to equip people to participate effectively.

Beneath the metrics, there is a deeply human invitation. Learning to use AI well is a compounding investment, quietly shaping who we become in an era of abundant information.

By embracing capability yield, we move sustainability past the defensive posture of simply doing less harm and into the realm of promoting human thriving. We step into an active, collaborative phase where technology developers, sustainability strategists, and enterprise leaders design systems that are genuinely worth the resources they require—systems that leave us, our organizations and our communities definitively more capable.

Note from the authors

Before you go, a quick word about how this piece came together. The four of us come at this from very different angles.

One of us builds and deploys AI systems and has spent a career putting technology to work in the hardest places on earth. One of us guides sustainability strategy inside a global enterprise. One of us has spent years helping companies turn AI ambition into real operational practice. And one of us builds the community of practitioners globally who live these questions every day.

We didn't set out agreeing on much… We each showed up with our own view, started comparing notes and realized we'd been circling the same idea the whole time. The footprint of AI is real and measuring it matters. But it only tells you what AI costs and it never tells you what you actually keep.

That gap is what we wanted to write about. Capability yield is our name for that missing half of the conversation. And we'll be honest with you, we don't think we have it all figured out yet. We're still working out the metrics, in real time, right alongside everyone else.

So please don't read this as the final word on anything, read it as an invitation. We wrote it because we believe the same tool that could deepen every divide we describe could just as easily help close it and the whole thing comes down to small choices, one habit at a time. Those choices belong to you as much as to us.

Thank you for reading.

- Alexis Bonnell, Franco Amalfi, Sandra Leyva, Antonio Vizcaya Abdo

References