AI’s environmental impact is not just a data-center issue. For enterprise software teams, the bigger question is whether AI should run at all. In this article, George Karapetyan explains why responsible AI governance, restraint, and simpler architectures can reduce risk, improve trust, and lower AI’s environmental footprint.
A colleague asked me recently whether we had thought about the environmental cost of our AI features. It was a fair question, and I had a ready answer, but not the one they expected.
They were anticipating something about carbon offsets, renewable energy pledges, or maybe a metric on tokens consumed per quarter. What I told them instead was this:
“The most impactful thing we do for AI's environmental footprint is decide, deliberately and repeatedly, not to use AI at all.”
That probably sounds like a dodge. It isn't.
For companies building AI into enterprise software, the most expensive AI is often the AI that did not need to run in the first place. It consumes energy, water, compute, attention, and trust without creating proportional value. That is why responsible AI governance has to begin before a model is called, not after the output is generated.
The conversation around AI environmental impact tends to get framed around infrastructure: which cloud provider uses the most renewable energy, how many liters of water a data center evaporates to cool its GPUs, what the AI carbon footprint of a single query is.
These are real numbers worth knowing. A single reasoning-heavy LLM query can emit up to 50 times more CO₂ than a simple one. Training GPT-3 directly evaporated roughly 700,000 liters of freshwater. U.S. data centers consumed 183 TWh of electricity in 2024, and that figure is projected to more than double by 2030.
But fixating on those numbers misses where the real leverage is. The question is not "how do we make our AI usage greener?" The question is "how do we make sure the AI we run is worth running in the first place?"
That distinction changes everything about how you architect a product.
We work in supply chain compliance and ESG risk assessment. The stakes of a wrong AI output are not abstract, they affect supplier relationships, regulatory filings, and the working conditions of people far down global supply chains. So we formalised our approach to AI in thirteen AI Philosophy tenets. They govern how every AI capability on the platform is designed, built, tested, and maintained.
When I look at those tenets through an environmental lens, something becomes clear: responsible AI governance and environmentally conscious AI are not separate agendas. They are the same agenda expressed at different levels.
Take the first tenet: we do not use AI where it is not needed. Before any AI feature goes live, we ask a simple question, is AI actually necessary here? Could a clear rule, a standard search, or a fixed template do the job just as well?
If the answer is yes, we use the simpler approach. This is not just good engineering practice. It is also the single most effective way to reduce the environmental footprint of AI. Every unnecessary AI call we eliminate is a call that consumes no energy, no water, and no compute.
The tenth tenet reinforces this: prefer simple methods over complex ones. If a smaller model solves the problem, we use it. If a deterministic approach works, we reach for that first. Complexity has to justify itself. We ask teams to show that a more elaborate approach is truly needed, not just interesting.
These two principles together act as a constant forcing function against AI sprawl.
The other tenets compound the effect in ways that are less obvious but equally real.
We expose data to AI only progressively and only during the session in which it runs. We share only what is needed for the task at hand, and we broaden context step by step. The environmental implication: shorter context windows mean fewer tokens processed, which means less compute consumed per interaction. This is a privacy principle with an environmental side effect, a good design decision that pays dividends in more than one dimension.
We build AI as a co-worker, not a co-decision-maker. AI suggests; people decide. This means we catch errors earlier, before they propagate into downstream AI calls that try to recover or reprocess bad outputs. Every loop that does not happen, every retry that is not triggered, represents a real reduction in wasted compute.
We keep users informed about what AI can and cannot do. When people understand the limits of a tool, they stop using it for tasks it is not suited for. That is AI footprint reduction through education, unglamorous, but effective.
We maintain the ability to turn AI off entirely, and we ensure core workflows function without it. This is not just resilience engineering. It is a structural constraint that prevents AI from becoming load-bearing in situations where it adds cost without proportionate value.
This is what sustainable AI architecture looks like in practice: not a single optimization, but a system of governance choices that limits unnecessary compute before it happens.
Beyond governance, there is a more direct environmental choice embedded in our platform architecture: we run on hyperscaler cloud infrastructure with a serverless model.
This matters more than it might seem. Hyperscale cloud providers, operating at massive scale with purpose-built efficiency, achieve power usage effectiveness (PUE) figures that most private data centers cannot approach. Serverless architecture means compute scales to zero when it is not in use. We are not running warm servers waiting for work. We pay for what we consume, and what we do not consume does not run.
Building our own data center, even a well-run one, would mean fixed capacity, idle compute, and the full capital and operational burden of cooling infrastructure. Staying on hyperscaler infrastructure lets us free-ride, in the best sense, on the efficiency investments those providers have made at a scale we could never justify independently.
This is not a claim that cloud infrastructure is without environmental cost, the numbers above make clear it is not. It is a claim that the marginal environmental cost of our workloads is lower on well-run hyperscaler infrastructure than it would be on infrastructure we built ourselves.
For enterprise AI teams, that distinction matters. Reducing AI energy consumption is not only about model choice. It is also about whether the surrounding architecture encourages efficient usage, idle capacity, or unnecessary scale.
The framing I keep pushing back against is the one that treats environmental impact as a separate workstream, something to be handled by a sustainability team, disclosed in a report, and offset somewhere downstream.
Our thirteen tenets were written to govern responsible AI: observability, traceability, groundedness, human oversight, resilience, proportionality. But when I map each tenet against its environmental implication, the alignment is nearly perfect. Every principle that makes AI safer and more trustworthy in a compliance context also makes it more resource-efficient. That is not a coincidence. It is what good governance looks like when it is designed from first principles rather than assembled from checklists.
The AI industry is going to keep growing. The energy and water demands are real, and they are not going away. But for companies building on top of AI rather than building AI itself, the most consequential decisions are not about which provider to use or which offset program to buy into. They are about whether to use AI at all, how much of it to use, and whether the architecture encourages restraint or sprawl.
We have chosen restraint, not as a sacrifice, but as a design principle. So far, it has made our AI features more reliable, more auditable, and more trusted by the people who use them. The smaller footprint is a consequence of building things right, not a compromise we made to feel better about the electricity bill.
That, I think, is the answer my colleague was actually looking for.
About author: George Karapetyan is Head of AI and Platform at IntegrityNext. This is part of an ongoing series on putting AI into practice in supply chain sustainability .
Sources:
May 2026
AI is rapidly transforming supply chain sustainability and compliance. Yet many organizations focus too heavily on model performance while overlooking a critical success factor: workflow design. This article explains why structured AI workflows—not just smarter models—are essential for scalable, auditable, and reliable compliance processes.
May 2026
AI is moving beyond experimentation in procurement and supply chain risk management. When connected to structured supplier, product, and risk data, it can help companies scale analysis, improve visibility, and turn sustainability insights into measurable business performance.
April 2026
AI for product compliance is one of the most challenging applications of modern AI. From BOM extraction to REACH and SCIP compliance, the gap between a working demo and a production-ready system is significant.
