In our last article of AI and sustainability, we introduced the concept of “work inflation”: the way AI makes it easier to generate more work than organizations may actually need.
A prompt feels free, so employees create more summaries, decks, emails, drafts and analysis before asking whether that work should exist at all (and which comes at a high environmental cost).
This behavior is scaling across many enterprises, with most not prepared to prevent AI from becoming another unmanaged layer of waste. That is why sustainability exists as a business function in the first place: to put externalities into decision-making, to make resource use visible and to force better questions before costs become embedded.
AI now needs that same treatment.
Without it, we may confuse waste with what looks like innovation: more pilots, more vendors, more tools, more dashboards, more automated workflows, more data moving through systems and more outputs someone has to review, secure, maintain or ignore.
Adoption Is Not the Same as Absorption
Let’s start with the obvious… Most companies are already past the question of whether they will use AI.
McKinsey's 2025 global AI survey found that 88% of respondents say their organizations regularly use AI in at least one business function, up from 78% the year before. And while this sounds like rapid evolution, the same survey found that most organizations remain in experimentation or pilot phases, with only about one-third reporting that they have begun to scale AI programs across the enterprise[1].
In other words, AI is being adopted faster than it is being “absorbed”.
Adoption means people have access to the tool. Absorption means the organization has redesigned the work around it.
Adoption means employees are using AI to draft, summarize, analyze, code, search or support decisions. Absorption means the company knows which of those uses actually improve performance, which create new risks, which duplicate existing work and which should be stopped.
Adoption can happen in weeks. Absorption takes governance, workflow redesign, data readiness, security, procurement discipline, training, measurement and trust.
That is where many companies are exposed. AI is being introduced into organizations that already struggle with fragmented data, disconnected systems, unclear ownership, overloaded teams and competing priorities. In many cases, besides exposing weaknesses faster, it exacerbates them by masking progress with activity (and not the productive one).
The 95% Is A Sustainability Issue and What That Costs the Planet
To put this in perspective, MIT studied enterprise AI in 2025 and found that only about 5% of pilots produced a measurable return; the other 95% went nowhere. Most people read that as a business failure, but it is also an environmental catastrophe. The footprint gets spent whether the project works or not (and for 19 out of 20 projects, it doesn't) [2].
Let’s see what this means for water.
Training a single model the size of GPT-3 used an estimated 700,000 liters of it [3]. Across the industry, researchers estimate AI consumed between 312 and 764 billion liters in 2025, and Morgan Stanley projects AI data centers could pass 1 trillion liters a year by 2028, while nearly half the world's data centers already sit in regions short on water [4].
That is drinking water pulled from farms, rivers, and towns to cool servers and the industry barely tracks it. And that is not it, then we have power…
A single AI rack draws about 120 kW (the electricity of roughly a hundred homes) and keeps drawing a big share of it even while idle [5].
We are spending the planet's water and carbon on AI that produces nothing, most of the time.
The Control Systems Do Not Exist Yet
While we’d like to provide a positive outlook, the uncomfortable truth is that (currently) most organizations are trying to govern AI with management systems built for slower technologies.
Traditional enterprise technology followed a more visible path where a business case was written, a vendor was reviewed, a system was implemented, users were trained and most importantly, changes moved through formal channels.
That process was often slow and frustrating, but it created some points of control. AI behaves drastically differently.
The barrier to experimentation is low and the perceived cost of use is low, the number of potential use cases is enormous and the technology is embedded into tools companies already use and employees can experiment before the organization has even finished writing the policy. That means the control system has to move upstream.
Companies need to know what counts as an AI use case, who approves it, what data can be used, which model is appropriate, which risks require review, how outputs should be validated, what value is expected and when the use case should be retired.
Because the problem is that most governance conversations focus on approval. But mature systems also need phase-out, retirement, substitution and continuous review. Otherwise, companies accumulate old tools, old workflows, old pilots and old assumptions long after the original business case has disappeared. AI is (and will keep) making that accumulation faster. Think about it…
A pilot can launch in weeks.
A workflow can be automated without being redesigned.
A model can become embedded before anyone has defined who is accountable for the output.
A tool can become normal before anyone knows whether it improved the work.
This is why AI governance cannot be limited to risk avoidance. The question is not only, “Is this safe?” The question is also, “Is this worth it?”
Worth the compute. Worth the cost. Worth the attention. Worth the risk. Worth the complexity. Worth the change required to make it useful.
That is where sustainability belongs in the conversation.
Sustainability teams are trained to ask whether resource use is justified by value creation, whether externalities are visible, whether short-term gains create long-term costs and whether systems are being optimized for outcomes rather than activity. Those are exactly the questions AI now requires.
What to Ask Before You Scale
If you’re wondering how to start, here are six that belong between proposing an AI project and scaling it:
These questions sit between an AI idea and an AI deployment, we share them to help leaders decide whether the organization is ready to move from experiment to scale.
What specific problem does this AI initiative address? If the answer is mainly that the organization needs to stay competitive, move faster or keep up with peers, the problem has not been defined clearly enough. Anxiety can create urgency, but it cannot carry an implementation.
How will this capability fit into existing systems, workflows, data environments and decision making processes? A tool that works in isolation may still create friction if people have to work around it, duplicate effort, or manually translate its output into the systems where work actually happens.
Does the organization have the leadership alignment, governance structure, risk involvement, security review, compliance input and change capacity required to absorb this shift? AI introduced into an unprepared organization creates confusion before it creates value.
Who benefits from this initiative, and how will that benefit be measured? The answer should be specific enough to shape KPIs, OKRs, adoption targets, quality standards, and decisions about whether to continue investing.
Do employees understand what the system does, where its limits are, how outputs should be reviewed and who remains accountable for decisions? Trust grows when people understand the role of the technology and the role they are expected to play alongside it.
Does this initiative strengthen or weaken the organization’s long term capacity to operate, adapt, govern and create value responsibly? Speed becomes an advantage only when the organization can maintain coherence as it scales.
Most companies can't answer all six today. That's the work ahead, it should shape how we experiment, not stop us.
Sustainable AI requires organizational capacity
The phrase “responsible AI” often focuses on fairness, privacy, safety and bias. Those remain essential. But sustainable AI needs a broader operating definition.
A company cannot claim to be using AI responsibly if the technology is multiplying low-value work, duplicating tools, expanding unmanaged compute, creating security exposure, overwhelming employees and consuming management capacity faster than it creates value.
That is not responsible adoption, it is resource-intensive disorder.
Sustainable AI requires organizational capacity.
Capacity to decide which use cases matter.
Capacity to integrate tools into real workflows.
Capacity to prevent duplicate systems.
Capacity to measure value honestly.
Capacity to train people by role, not through generic awareness campaigns.
Capacity to govern data access.
Capacity to monitor outputs.
This is the harder work because it is less exciting than launching a new tool and it does not create the same internal buzz as a pilot. But it is the work that determines whether AI becomes useful infrastructure or another layer of enterprise complexity.
Govern where AI is allowed to scale.
Map what resources, risks, teams, and workflows it touches.
Measure whether it creates value proportional to its cost.
Manage what should continue, change, or stop.
That is not bureaucracy for its own sake. It is how companies prevent resource-intensive technologies from becoming unmanaged systems.
Responsibility, Not Restraint
We don't have to resolve the tension between AI's cost and its promise, we have to manage it. And the water and carbon spent on failed pilots and idle clusters show up in no sustainability report. They're invisible, because the company that spent them never had a clear view of what it deployed, what it used, and what it left running. You can't cut what you can't see.
A company running powerful AI on a messy operating model isn't doing sustainable AI, it's doing fragile AI, burning water and carbon no matter how clean its energy contracts look. The rollout won't slow and the footprint will grow before it shrinks.
What we control is whether that footprint goes to AI that helps the planet or AI that fails it (and that comes from the governance, the controls, and the choices leaders make before scale becomes the default). Which is, in the end, where every sustainability question has always been decided.
A Note from the Authors
In this edition of Sustainability Decoded, Sandra Leyva and Antonio Vizcaya Abdo continue the series with Field Decoders Franco Amalfi and Jonathan Solis.
Franco Amalfi is Global Sustainability AI Partner Ecosystem Lead at Capgemini. Across consumer products, finance, government, and technology, he has watched the same pattern repeat: a tool arrives with real promise, the rollout speeds up, and the questions that should have come first get asked too late.
Jonathan Solis is a Program Director at Microsoft, where he leads a $50M-plus portfolio of agentic AI, cloud transformation, cybersecurity and digital-asset programs for Tier-1 financial services clients within Microsoft's Customer Experience & Success organization. His work sits exactly where this article lives: bridging AI deployment with operating-model design and enterprise adoption inside large, highly regulated institutions.
Where our first piece looked at the footprint of the work itself, this one follows that footprint into the organization and asks whether the planet's resources are going to the AI that helps it or the AI that fails.
Sources
McKinsey. The State of AI: Global Survey 2025. QuantumBlack by McKinsey, November 2025. Reports that 88% of organizations regularly use AI in at least one business function (up from 78% a year earlier), with the majority still in experimentation or pilot phases and only about one-third having begun to scale AI across the enterprise. Based on 1,993 respondents across 105 countries.
MIT Project NANDA. The GenAI Divide: State of AI in Business 2025. 2025. Finds that roughly 5% of enterprise AI pilots achieve a measurable return, against an estimated $30–40 billion in spending. Reported in Fortune, August 18, 2025.
Li, Pengfei, et al. (University of California, Riverside / University of Texas at Austin). "Making AI Less Thirsty: Uncovering and Addressing the Secret Water Footprint of AI Models." 2023. Estimates that training GPT-3 in Microsoft's U.S. data centers evaporated roughly 700,000 liters of clean freshwater on-site.
de Vries, Alex. Patterns (Cell Press), December 2025, estimating AI-related water consumption of 312–764 billion liters in 2025; Morgan Stanley projection of more than 1 trillion liters annually by 2028; Arup / IE Insights (2025), noting that nearly half of the world's data centers are located in water-stressed regions.
NVIDIA and Schneider Electric infrastructure data, 2024–2026: a GB200 NVL72 AI rack draws roughly 120 kW, about ten times a conventional server rack (which rarely exceeds 10–15 kW). GPU energy-measurement research (arXiv, 2026) shows accelerators continue drawing substantial power during idle and "execution-idle" periods.