Two reports recently landed in the same week. The UN University Institute for Water, Environment and Health published a hard-hitting assessment of AI's environmental footprint. Its conclusion: no major AI provider publishes reliable data on the energy, water, or carbon cost of a single query. Days later, the European Commission launched its tech sovereignty package, putting sustainable digital infrastructure at the heart of Europe's AI agenda.

Same message, two sources: how we use AI has consequences. Here's how we're thinking about it at Visma.

The real challenge is blindness

Every AI-and-sustainability conversation hits the same wall: the numbers aren't there. No standardised reporting. No comparable data across providers. No reliable way to know what a single interaction costs in energy or emissions. That gap isn't a technical glitch. It's a governance shortfall, and it matters.

Waiting for perfect data isn't a strategy. The UN report confirms what practitioners already suspected. Once a model is deployed, it's the everyday running of models to answer millions of prompts that dominates AI's operational energy use, 80 to 90 per cent of it. Not the one-off training of large models. Routine, daily use, at scale, right now.

Data centres, hardware, cooling, these are real footprints, no question. But operational load is where organisations can actually move the needle. Every day without an AI usage policy is a day of unmanaged impact. Our own people ask us both sides of the same question: are we moving fast enough on AI, and responsibly enough? That's not a comms problem. It's a fair question, and it deserves a straight answer.

What we control, and what we're doing

Visma runs roughly 550 AI product initiatives group-wide, 200 with real transformational potential. Behind that scale, we've laid the groundwork:

• Internal guidelines for environmentally sustainable AI covering practices such as: provider transparency, the right tool for the right use, disciplined data handling and many more.
• An AI Code of Conduct, a living framework covering which tools get assessed, how customer data is handled, and why AI-based employee monitoring is off the table. Every strategic AI initiative goes through legal review. Our AI Governance Portal gives every team one place to check what's approved. Prompt injection and other AI-specific risks now sit inside our standard security reviews.

What we don't do: approve every tool centrally. That call sits with the teams closest to the work.

This follows the same logic as last year's GreenOps programme: cloud efficiency and emissions reduction are the same move. e-conomic cut cloud emissions 67 per cent in a year. Flex Applications hit 93 per cent lower CO₂ migrating to Azure. Both cut costs at the same time. As token-based billing becomes standard, the same rule holds: less token use means lower cost and lower emissions, together.

The tools already exist. The habits are the work.

None of this waits for regulation, perfect data, or a dedicated sustainability team:

• Choose the right model for the task.
• Write efficient prompts.
• Skip unnecessary multimedia generation.
• Shift workloads to cleaner energy grids where possible.

Responsible AI use isn't about having every answer today. It's about building the right practices ahead of regulation, and staying honest, with stakeholders and with ourselves, about exactly where we stand.

We're on that journey. We're not at the end of it.

‍