Can a sustainable company use AI? How to turn doubt into a responsible decision.
Can a sustainable company use AI? Understand how to evaluate suppliers, emissions, water, and productivity without confusing efficiency with environmental compensation.
A company with environmental commitments can use artificial intelligence. The existence of energy, water, or material consumption does not automatically make a technology incompatible with sustainability. What needs to be evaluated is the relationship between its purpose, the impacts involved, the available alternatives, and the commitments the organization has made.
This also doesn't mean that any application is justified by generating revenue or saving time. Economic gains and environmental impacts are different dimensions: both can be relevant to a decision, but one doesn't automatically negate the other.
The starting point is to treat AI as an activity that uses infrastructure and requires criteria commensurate with its importance to the business. Not as an immaterial exception simply because it's digital, nor as a problem that must necessarily override all other environmental priorities of the company.
ESG is not a single "green" company label.“
ESG encompasses environmental, social, and governance dimensions. It is not a universal certification with a single list of permitted or prohibited technologies. It is necessary to distinguish between voluntary commitments, disclosure standards, certifications, market assessments, and obligations that are actually applicable to the organization.
B Corp certification, for example, follows B Lab standards, with requirements in areas such as climate action, environmental management, human rights, and governance. The organization itself states that the requirements are adjusted according to characteristics such as size, sector, and activity. It is not a declaration of a total absence of impact.
Therefore, the use of AI should not be interpreted as automatic approval or disapproval of a certified company. The real question is whether the organization understands the relevant impacts of adoption, considers risks, and manages its choices in a manner consistent with the criteria it follows.
This distinction avoids an impossible expectation of technological purity. Corporate sustainability involves decisions between real alternatives, not just the abstract choice between producing and not producing impact.
Materiality: not every company needs to start with the same problem.
An organization that occasionally uses AI to support administrative tasks is in a different situation than one that offers large-scale content generation or maintains agents continuously running processes.
However, call volume shouldn't be the only criterion. Task type, infrastructure, location, and the importance of the service to the business also matter. The goal is to identify where analysis can change a decision, rather than requiring the same level of documentation for every use.
The GHG Protocol recommends an initial assessment of emission sources to guide data collection and deepen the work in the most relevant areas. This approach helps to avoid both omission and disproportionate expenditure of resources on insignificant portions of the inventory.
For an industrial company, for example, investigating AI should not be a reason to postpone clearly relevant improvements in energy, materials, or logistics. For a business whose activity depends heavily on computing, digital services may deserve much greater attention. These examples do not establish a sectoral rule: they show why priorities need to arise from an assessment of the business, not from the public visibility of a technology.
Where do emissions from AI services factor into the calculation?
The GHG Protocol includes, in Category 1 of Scope 3, emissions associated with the production of acquired goods and services that do not belong to other specific categories. Applying this logic, external AI and computing services can be included in the evaluation of contracted services. The classification depends on the contractual relationship and the boundaries of the inventory.
It is not advisable to treat every situation as identical. Contracting access to a service, operating one's own equipment, and leasing infrastructure represent different relationships with assets and energy used. The classification must follow the adopted methodology, avoiding double counting within the inventory.
The distinction between scopes organizes accounting responsibilities: Scope 1 deals with direct emissions; Scope 2, with purchased energy; and Scope 3, with other emissions from the value chain. The same activity may appear in different scopes of different companies, without this necessarily representing an error.
The practical conclusion is limited, but important: Outsourcing your computing to a third party doesn't make your emissions nonexistent. . It also does not authorize attributing the supplier's entire footprint to the client.
Water needs its own analysis.
Water is not just an extension of the carbon bill. Its origin, availability, quality, and local conditions alter the meaning of its consumption.
GRI 303 guides the identification of water impacts related to operations and business relationships, including suppliers with significant impacts and activities in water-stressed areas. The information should be interpreted in light of the basin context.
For an AI buyer, this might mean requesting information about location and water management when the service is relevant enough. It doesn't mean every company needs to start by calculating milliliters per employee, especially when reliable data for that precision is lacking.
An estimate with many decimal places is no substitute for knowledge about the limits of measurement. It is more useful to know that a given figure only covers refrigeration and excludes electricity than to present a seemingly exact total without explaining what it represents.
Productivity is a benefit, but it is not environmental compensation.
Suppose a company saves thousands of hours with an AI application. This result can improve customer service, reduce repetitive work, and free up capacity for other activities. It's a significant operational benefit, but it doesn't, on its own, demonstrate a reduction in water, energy, or emissions.
It is also not necessary for every AI application to produce a direct environmental benefit for its use to be defensible. A tool may be useful for reasons of quality, accessibility, or productivity. In that case, the company should recognize that value without artificially converting it into a climate claim.
The reasoning changes when there is a verifiable physical alteration: less fuel used, less material wasted, or less energy needed to deliver the same service. Then there is an environmental outcome that can be measured and evaluated.
Result Which demonstrates What doesn't demonstrate on its own Hours of work saved Productivity gain Emissions reduction Lower operating costs Economic benefit Reduced use of physical resources Lower measured energy consumption Energy savings at the analyzed limit. Net benefit of the entire project Emissions avoided compared to an alternative. Difference between defined scenarios Automatic inventory offset
These measures can complement each other, but they are not interchangeable. Precise communication must preserve the difference.
Inventory reduction and avoided emissions are not the same thing.
When an application effectively reduces the energy consumption of its own operation, this change can be reflected in the corresponding inventory, according to the methodology used. It would not make sense to maintain the old emission as if the reduction had not occurred.
Another situation is estimating how much a solution has avoided compared to a hypothetical alternative. The GHG Protocol advises that these estimates of avoided emissions be presented separately from Scope 1, 2, and 3 inventories, and not deducted as an automatic credit.
This distinction is especially important for technology providers. A company might sell a system that helps customers reduce energy consumption. This doesn't translate the estimated savings for customers into authorization to offset the emissions from their own data centers.
It is also necessary to avoid counting the same gain twice: first as a decrease observed in the inventory and then as an additional deduction. Recognizing environmental benefits requires method, not just a narrative favorable to the project.
How can we verify if AI has truly reduced the impact?
A consistent assessment requires a baseline scenario: what would likely happen without the solution? The WBCSD guide on avoided emissions emphasizes precisely the definition of this alternative, the data used, and the validation of the solution's contribution.
The alternative should not be chosen simply because it makes AI seem better. In logistics optimization, for example, the appropriate comparison might be with the existing routing system, not with the complete absence of organization. In an industrial process, changes in production, maintenance, or equipment can explain part of the observed savings.
It is also important to preserve the equivalence of service. Reducing energy because a facility operated fewer hours does not, in itself, demonstrate improved efficiency. The result needs to consider how much was produced and under what conditions.
There are applications with significant potential. The IEA describes possibilities in buildings, power grids, industry, and transportation, but also highlights implementation barriers and indirect effects. Potential for reduction is not the same as reduction already achieved in a specific company.
What should we ask suppliers?
Contracting can include environmental information without turning every purchase into an extensive audit. The level of stipulations should consider the relevance of the service, the available data, and the type of claim the company intends to make.
A useful set of questions involves the scope of the data, the reference period, the location of the processing, and the attribution methodology. Does the indicator belong to the contracted service or the entire company? Does it include only active equipment or also supporting infrastructure? Has it been verified? How does the supplier distribute its consumption among clients and products?
Information requested Why does it help? Measurement limits Avoid comparing accounts with different components. Service data or allocation criteria It brings the analysis closer to what was agreed upon. Location and infrastructure conditions It allows for the assessment of energy and water differences. Methodology and verification It helps to judge reliability and uncertainty. Available technical alternatives Allows you to test for lower resource usage.
This proposal does not mean that all answers will be available. Lack of transparency may limit comparison, but it does not, in itself, prove that one supplier has a greater impact. Similarly, a detailed report improves assessment capabilities without guaranteeing good performance.
What to do when complete data is not available?
The GHG Protocol allows for different methods to estimate emissions from purchased goods and services, including supplier-specific data, hybrid methods, averages, and expenditure-based estimates. The guidance acknowledges that more specific data is not automatically more accurate: source quality and allocation rules also matter.
A reasonable approach is to start with the method compatible with the available data and acknowledge its limitations. When the source is relevant, it is worthwhile to deepen the data collection and improve the estimate. When it is not, it is possible to document the assessment without dedicating disproportionate resources to it.
Estimates based on expenses, for example, can be useful for screening, but a reduction in the contracted price should not be automatically confused with a physical reduction in emissions. Business changes can alter the invoice amount without changing the processing performed.
The important thing is to avoid two extremes: inventing precision or concluding that, because there is no precision, nothing can be evaluated.
How to adopt computational proportionality in the company
A usage policy can start with a simple question: what capacity is needed to deliver the result safely and with quality? The answer should be tested, and not defined solely by a preference for the most advanced model or the lowest cost.
For recurring tasks, the company can compare alternatives before expanding usage. Projects that execute many steps may have limits, interruption criteria, and referral for review. Model changes may also require reassessment because the system's behavior and its execution needs may change.
This doesn't require authorization for each individual message. The focus can be on processes that concentrate volume, cost, risks, or operational dependencies. A small pilot and a feature incorporated into all products don't need to receive exactly the same treatment.
Governance must also identify those responsible. Who monitors the results? Who decides to expand the project? Who verifies whether the projected savings have materialized? Without these answers, a project may continue to consume resources even when it has ceased to deliver the benefit that justified its adoption.
How to communicate without exaggerating or hiding limitations.
A statement like "our AI is sustainable" says little without explaining which service, which indicator, and which comparison. A more delimited formulation is more useful: "the application reduced the energy consumption of this process, in the analyzed period, compared to the described reference scenario.".
When figures are provided, communication should highlight the relevant limits and not conceal additional costs of the solution. The WBCSD recommends transparency in the assessment and disclosure of avoided emissions, including defining the baseline scenario and validating the contributions.
It's also not necessary to turn every improvement into a grand promise about the company. An advance in a particular process can be presented as such. Acknowledging that the data is still incomplete can be more responsible than filling the gap with adjectives.
On the other hand, a company doesn't need to apologize for every use of technology. It can explain its benefits and how it manages the impacts, as long as it doesn't confuse economic justification with environmental neutrality.
Benefits in one dimension do not solve all the others.
An application can reduce emissions and still present issues of privacy, discrimination, security, or accountability regarding decisions. NIST treats AI risk management as a process that involves individuals, organizations, and society, not just the technical performance of the system.
Therefore, energy efficiency should not function as a blanket authorization. A tool that uses few resources but causes significant errors or infringes on rights still requires correction.
The same applies within the environmental dimension. Lower carbon emissions do not automatically eliminate local water concerns, and reduced water consumption does not prove a lower impact on all other aspects. Choices may involve trade-offs between indicators, which need to be presented rather than hidden in a single score.
A responsible decision may include using, modifying, or not adopting.
A realistic analysis needs to consider more than one outcome. In some cases, the application delivers value with manageable impacts. In others, it works better with a different model, fewer steps, or usage limitations. There may also be situations where the complexity, risk, or total cost do not justify adoption.
There is no contradiction in an environmentally committed company reaching different decisions for different projects. The consistency lies in the criteria applied, not in approving or rejecting an entire technological category.
Similarly, a potential benefit should not be treated as unlimited justification. Adoption needs to respect applicable constraints and consider who receives the benefits and who bears the costs. A positive aggregate outcome does not automatically account for all local consequences.
Frequently Asked Questions
Can a B Corp company use artificial intelligence?
The use of AI alone does not allow one to conclude whether a company meets or fails to meet B Corp standards. The assessment depends on the organization's activities and applicable requirements, which encompass environmental, social, and governance dimensions.
Does productivity offset the environmental footprint of AI?
Not automatically. Time or revenue savings may economically justify an application, but an environmental claim requires demonstrating a change in resources or impacts. Economic and environmental benefits must be presented as distinct outcomes.
Does every company need to calculate consumption using prompts?
There is no reason to treat this indicator as a mandatory starting point. Depending on the business, an evaluation by service, project, or supplier may be more useful. The choice should consider materiality, data quality, and the purpose of the analysis.
Does a company need to prove that AI reduces emissions in order to use it?
Not as a general principle. An application can be adopted for other benefits. The important thing is to manage your footprint in a way that is compatible with the commitments made and not to present an operational gain as an environmental reduction without evidence.
Can a sustainable company use AI?
A sustainable company can use AI without needing to claim that the technology has no impact or that every application will produce a climate benefit. The most consistent commitment is to understand the relevant activities, compare alternatives, and distinguish between economic value, environmental footprint, and effectively demonstrated reductions.
This requires proportionality on both sides: neither rendering an important digital operation invisible nor devoting attention to it that is disconnected from its relevance to the rest of the business. The evaluation should help improve decisions, not just produce a favorable justification or a generic prohibition.
The intersection between AI and environmental responsibility lies in using the technology when there is a reason to do so, managing its costs, and communicating the results to the extent that the evidence allows. It's not the absence of impact; it's the ability to understand, prioritize, and improve.
Sources and limitations of the analysis
This text uses references from the GHG Protocol, GRI, B Lab, WBCSD, IEA, and NIST. The proposed supplier decision and evaluation guidelines are analytical recommendations, not universal requirements of these organizations. The accounting, regulatory, or certification framework depends on the organization's circumstances and should be verified according to the standards that are actually applicable.