What should companies focus on? Florence Carlot, Partner in ADL’s Energy, Utilities & Resources practice, says strategic resilience should be a top priority. Companies should be increasing their flexibility and “portability” in relation to AI, she says. A good option here is designing AI architectures that segment workloads across on-premises, sovereign, and public AI platforms. Another is using architectures and formats that can migrate between clouds, vendors, or regions.

Companies should consider diversifying providers to spread their risk of operational complications should an unexpected point of failure occur. Within this wider supplier base, they might consider negotiating more flexible terms in the event of unforeseen disruptions (think export bans, resource shortages, and demand shocks). Routine stress tests in anticipation of such eventualities are high on Carlot’s should-do list.

Of course, this presupposes a business knows where it currently stands. If not, leadership is flying blind. This makes managing AI-related vulnerabilities almost impossible; the same goes for accounting for AI-linked environmental impacts. The possibility of radical innovation is also greatly reduced. Advances in clean energy and other climate-smart solutions, for example, make a net zero AI ecosystem theoretically possible. But to arrive at such a future, it’s necessary to draw a pathway to there from the present — a difficult challenge for those whose current situation remains opaque.

Interestingly, AI’s analytical power could be used to project backward, so as to redesign its own lower-carbon future, says Lukas Falcke, Assistant Professor of Digital Strategy and Innovation at VU Amsterdam, the Netherlands, and a leading advocate of digital-first solutions. Doing so, however, requires a proactive, opportunity-focused approach to auditing and reporting.

“With a digital-first mindset, it’s almost like the future becomes your starting point, because with AI you can more or less predict what emissions you will create and where, which allows you to work back from there,” Falcke says. “So instead of seeing reporting as compliance, look at it as a data-driven iteration toward a reimagined future.” (For a broader lens on using digital innovation for sustainability, see ADL’s Amplify issue “Charting a Sustainable Future with Digital-First Solutions.”)

For AI’s corporate users, gaining a clear picture of the present is harder than it sounds. As the Blue Shift report points out, obtaining meaningful data from AI operators is difficult. Fewer than 3% of new AI models disclose environmental data, down from 10% in 2023.

Not only is current data scarce, but AI providers tend to aggregate the little data they make available. This makes it tricky for companies to work out the emissions or water footprint linked to specific data centers or particular AI tools, complicating their ability to calculate the impacts of their individual AI usage.

Nevertheless, Carlot counsels against a wait-and-see approach to disclosure. Companies should also expect requirements to report on their AI-related risks and impacts to tighten in the future, she says. “Exactly when is hard to say in some jurisdictions, but there’s no doubt it will happen.”

Carlot outlines three basic measures all companies should undertake. The first and most obvious is mapping their AI touchpoints. What AI tools are they currently using? What’s the respective purpose and importance of these tools? Which vendors are they most reliant on? Do these vendors also operate data centers and, if not, who does? “Every company needs to build a basic AI inventory based on questions like these,” she says. “That creates a baseline for their material impacts along the AI value chain, which they can then use to report back.”

Steps two and three involve putting hard numbers against this inventory and stress-testing them against various scenarios. Once completed, companies should find themselves with the requisite data to set realistic targets for reducing the environmental price tag of their AI use. Carlot says that for many companies, obtaining such data requires prior steps. For example, as part of the contract-setting process with AI vendors, she suggests firms make the provision of environmental data a nonnegotiable criterion.

This demand is entirely reasonable, Carlot explains, because the expense of electricity and water means that AI firms’ usage is carefully tracked. In addition, as more companies request this data, smart AI providers will realize that eco-disclosure can be a competitive differentiator.

Companies should also consider engaging at one level up: the AI industry level. For auditing and decision-making purposes, it is important that site- and product-level information can be compared across various providers. Otherwise, companies will find themselves trying to “measure apples against oranges.”

Ideally, says Carlot, the AI industry would agree on a shared methodology for reporting planetary impacts. With reporting norms for AI firms still taking shape, proponents argue that an industry-led framework could serve as a model for legislators.

Armand Smits, Assistant Professor of Organizational Change and Design at Radboud University, the Netherlands, sees merit in such a position, but recent experience in Europe suggests trust could present a hurdle. Under the umbrella of the Climate Neutral Data Centre Pact (CNDCP), more than 100 data center operators and trade associations came together in 2021 to agree on common environmental targets and measures (see ADL Amplify article “Establishing Collective Environmental Self-Regulation in Fragmented Digital Spaces”). To date, EU legislators have not been “that supportive” of CNDCP’s proposals, Smits says.

“On the plus side, the CNDCP initiative signals acceptance in the AI sector about the demand from regulators for them to report more,” he says. “On the downside, the opposition from NGOs and local governments to AI operators being involved in setting standards shows a high level of distrust toward the industry.”