What AI Actually Costs Beyond the Subscription, and How to Reduce It
Post 2 of 3 in a series on responsible AI adoption for destination marketing organizations (DMOs) and tourism leaders.
Post 1: You're Not Behind on AI. You're Where You Need to Be to Get It Right.
Post 2: What AI Actually Costs Beyond the Subscription, and How to Reduce It
Post 3: Five Principles for Adopting AI Responsibly in Tourism (Coming Soon!)
A 25-person DMO team using a major AI provider's most powerful model for routine work generates roughly 1,190 kg of CO₂ and consumes 9,520 liters of water in a year. That's about ten round-trip flights between London and Paris worth of emissions, and enough water to fill a backyard pool. None of it shows up on the AI subscription invoice.
The first post in this series argued that DMOs aren't behind on AI adoption. They're behind on responsible adoption. This post brings more awareness to the total costs that are associated with the use of AI. If we talk about 'responsible' adoption, we cannot only talk about the benefits without looking at three categories of impact most DMOs aren't yet tracking. Each category produces material costs. Each has things you can do about them.
Environmental impact
AI's environmental footprint comes in three forms: energy consumption, water use, and greenhouse gas emissions. The mechanics are connected. AI runs on data centers, which need electricity to power the chips and water to cool them. The electricity has its own emissions and water footprint depending on whether the local grid runs on coal, gas, hydro, or renewables.
It's worth noting that AI's environmental impact happens at two distinct moments: training the model (a one-time cost that gets amortized across billions of uses) and inference (the ongoing cost every time someone uses the model). Training is a single massive spike, but inference is the steady tide. For popular AI services deployed at scale, recent research suggests inference accounts for up to 90% of total lifetime energy use. Training gets most of the headlines. Inference is what your organization actually influences with day-to-day choices, and it's where most of the impact lives.
Most of this is invisible at the user level. You pay a flat monthly fee. The energy, the water, and the emissions don't appear anywhere on your bill. They're real, just absorbed by the data center operator and the surrounding community.
The 25-person team scenario above is calculated from the AI Environmental Impact Calculator I built. The inputs that produced those numbers: Europe as the region, OpenAI's most powerful model tier, 30 prompts per day per person. You can replicate the result by entering the same variables, or insert your own organization's numbers.
One variable in particular highlights how a simple action can reduce impact significantly. If that same team uses the same provider's basic model tier instead of the most powerful one, the numbers drop to 98.6 kg of CO₂ and 788 liters of water per year. A 92% reduction across the board.
This is a key area where AI users have autonomy to mitigate their impact. Frontier-tier models, the most powerful options each provider offers, use about 14 to 24 watt-hours per query. A basic-tier model handling routine tasks can use under two watt-hours. The defaults most people select are rarely the most efficient option for the actual job. Most DMO marketing work (drafting copy, summarizing content, basic analysis, customer-facing chatbots) is well-served by basic and mid-tier models. Reaching for frontier reasoning every time is the equivalent of taking a transcontinental flight to visit the next town over.
The first move is to run your team's actual usage through the calculator to estimate your impact. The second move is to match capability to task. Not every prompt needs the most powerful model on the menu. Beyond tier matching, two other levers help. Different providers vary substantially in efficiency, and the gap between providers that publish per-query environmental data and those that don't is itself a useful signal. Stanford's 2025 Foundation Model Transparency Index found that 10 of 13 major AI companies disclosed none of the key environmental metrics. Choosing providers that disclose, and weighing their footprint when capability is comparable, is a powerful decision. The other lever is prompt and output discipline: shorter prompts, tighter responses, and reusing context where possible. These can reduce footprint another 20 to 40% beyond model-tier choice.
These are estimates. Actual impact varies by infrastructure, region, prompt complexity, model updates, and other factors. The calculator is a starting point for understanding scale and relative impact. The point is the order of magnitude of impact and the relative differences between usage choices.
Social impact
Social impact is harder to measure than environmental impact, which is part of why it gets neglected. The categories that matter most for DMOs are workforce displacement, resource competition with host communities, cultural representation, and algorithmic bias in the systems that shape and produce destination content. Each one carries material costs. Some are partly within a DMO's control. Others require advocacy.
Workforce: AI is reshaping European jobs faster than employers are adapting. Littler's 2025 European Employer Survey, covering more than 400 HR professionals across 14 countries, found that 71% of European employers have reassessed job responsibilities due to AI. More than a quarter of European employers have reduced hiring or cut jobs as a direct result of AI deployment.
The hospitality and tourism sector is especially exposed. Eurostat's DESI indicators show that fewer than 15% of EU hospitality employees have the digital skills considered essential for AI-enabled systems, compared with over 40% in IT and finance. Frontline tourism roles like reception, service, content production, and customer support are among the most automatable, and the workforce has the lowest capacity to adapt. A 2026 Brookings analysis found that the U.S. workers most exposed to AI displacement and least equipped to adapt are 86% women, concentrated in clerical and administrative roles in smaller metropolitan areas.
What a DMO can do here is limited but not zero. Naming the most-exposed roles in your organization and planning intentional transition support is a starting point. Investing in upskilling, cross-training, and the kinds of skills AI doesn't replicate well (interpersonal, judgment, context-specific knowledge) is another.
Resource competition with host communities: Data center buildout has become increasingly contested. In Aragon, Spain, a region facing severe water stress, Amazon has been licensed to draw approximately 755,000 cubic meters of water per year for its facilities. In Castilla-La Mancha, a Meta data center is projected to consume 665.4 million liters of drinking water annually. Similar patterns are appearing in tourism-adjacent regions in the United States, with high-profile community pushback in Utah, Arizona, and Texas. These are tourism destinations and tourism-adjacent regions. The infrastructure your AI tools depend on is competing for the same water and energy your visitors and residents need.
A DMO operating in a region affected by data center expansion has standing to advocate. Most DMOs have public commitments around sustainability, community wellbeing, and stewardship of place. AI infrastructure decisions in your region fall inside that scope, even if they aren't usually treated that way.
Cultural representation: Generative AI is dramatically lowering the cost of producing inauthentic cultural content, and this affects many marginalized communities. An Australian Productivity Report found that 75% of Indigenous-style consumer products on the market were non-Indigenous authored, a problem the tourism sector already had and that AI is now exacerbating. AI systems trained on limited data for Indigenous languages have begun hallucinating fabricated words and phrases, presenting them as authoritative and undermining language revitalization efforts. The same dynamics affect representation of Black, Brown, Latino, Asian, LGBTQ+, disability, and religious minority communities, often through stereotyped imagery, reductive descriptions, or outright erasure.
For DMOs working with Indigenous communities, three established frameworks apply directly to AI: OCAP (Ownership, Control, Access, Possession) from the First Nations Information Governance Centre in Canada, the CARE Principles for Indigenous Data Governance, and FPIC (Free, Prior and Informed Consent) under UNDRIP. None of these were created for AI specifically. All of them apply. The broader principles underneath them (consent, accuracy, accountability to the communities being represented) apply to any cultural content AI is generating about people who aren't in the room when AI gets used.
Algorithmic bias: This shows up in two distinct ways for DMOs.
The first is in AI travel-planning systems that concentrate their recommendations heavily. SALT.agency analyzed 80,000 AI travel-planning responses in 2026 and found that the most-recommended countries for general travel were Canada, the UK, Mexico, the US, India, Australia, France, Singapore, Jamaica, and the Bahamas. The pattern: AI tends to recommend destinations and attractions that are easy to talk about, and places with abundant existing content. The same pattern repeats within destinations. AI funnels visitors toward the most-documented sites and neighborhoods (often the ones already most affected by overtourism) while culturally distinctive but less-documented experiences get overlooked. If your destination is in that recommendation set, AI is reinforcing flows that may already be straining places and people. If your destination, or specific places within it, are less well-documented, AI is making them functionally invisible. The algorithm is making the choice, which means destinations that don't actively shape how AI sees them are letting that choice be made by default.
The second is the bias that DMOs introduce themselves when they use AI to produce their own content. AI inherits the biases of its training data. When you ask an AI to generate destination imagery, write visitor personas, or draft marketing copy, the defaults it produces reflect what the training data overrepresents. That tends to be western, white, young, able-bodied, and affluent. This is more directly within DMO control than the recommendation-system problem. Reviewing AI-generated imagery and copy for stereotyping before publication, testing outputs across different prompts to surface the defaults, having community members review content about their communities before it goes out, and being explicit in prompts about the diversity you want represented are all things a DMO can do today. The bias doesn't disappear. The DMO's role in either propagating or counteracting it does become a deliberate choice.
Governance impact
Governance is the impact category most DMOs haven't yet recognized as a category. It shows up as legal liability, reputational damage, regulatory exposure, and decisions made by people without authority to make them.
The clearest tourism example is Air Canada. In 2024, a Canadian tribunal ruled that Air Canada was responsible for incorrect bereavement fare guidance its AI chatbot gave to a customer. The airline's defense, that the chatbot was effectively a separate legal entity, failed in court and in public opinion. Air Canada was ordered to compensate the customer and absorbed reputational damage that lasted longer than the ruling. The case set a precedent: organizations are accountable for what their AI tells customers, regardless of whether a human reviewed the output before it went out.
The pattern repeats across sectors. In the United States, Workday is facing a class action lawsuit over AI-powered screening tools that allegedly rejected applicants over 40 at disproportionate rates. McDonald's exposed thousands of job applicant chat records when its AI hiring chatbot turned out to be protected by a default admin password of 123456.
Regulation is real, has teeth, and is moving. The EU AI Act's transparency requirements for AI-generated content, deepfakes, and chatbots take effect in August 2026, with penalties up to €15 million or 3% of global annual turnover for serious breaches. These rules apply to any organization marketing into the EU, regardless of where the organization is based. Beyond the EU, the regulatory landscape is unsettled. Canada's federal AI legislation collapsed in early 2025. The U.S. has no comprehensive federal framework and a growing patchwork of state laws. DMOs operating across jurisdictions are navigating a moving target. Specific dates and rules will continue to shift. The fact of regulatory exposure won't.
What a DMO can do here starts with three questions.
Who reviews AI outputs before they reach customers, partners, or the public?
Who is accountable when AI produces something harmful?
What gets logged, so you have a record of what AI did and on whose authority?
Most DMOs can't yet answer any of these. Working through them is the foundation of governance that holds up under both legal scrutiny and public reputation.
Why use AI at all?
We just laid out significant costs across three categories. The magnitude of those costs raises a fair question: if AI is this expensive in human and environmental terms, why use it at all?
The only reason to consider using a technology with these costs is if it delivers greater benefits, distributed equitably. AI is a powerful tool with meaningful uses across science, medicine, climate, and access to expertise that wouldn't otherwise reach the people who need it. None of that erases the costs.
So I continue to wrestle with the question of whether we should use it, and I think the wrestling itself is part of responsible AI adoption. For my own consulting practice, the calculator estimates my annual AI usage at 51.7 kg of CO₂ (roughly half a one way flight from San Francisco to Los Angeles) and 266 liters of water (about four eight-minute showers). Those numbers aren't catastrophic. They aren't nothing either. Being able to quantify the impact helps me continue to make informed AI usage decisions going forward.
Here are three positions a person could take. You can oppose AI completely and abstain. You can adopt it without accounting for the impacts. Or you can adopt it as a deliberate practice: counting what it costs, reducing what can be reduced, and advocating at the systems level for the infrastructure and accountability that individual choices can't deliver.
The first is clean. The second is convenient. The third asks you to weigh AI's usefulness and its costs. The next post is about how to operate from that third position when the technology, the costs, and the regulations keep moving. Specific answers age fast. A thoughtful framework for making decisions and governing usage holds up better.
Tyler Robinson is the Founder and Principal Consultant of Tydal Consulting, where he helps organizations adopt AI responsibly and effectively. He brings over a decade of experience as a climate and sustainability strategist, advising places and organizations on how to operate more sustainably. He built the AI Environmental Impact Calculator to help organizations measure the impact of their AI usage.
