AI Literacy in Sustainability: What the Evidence Says About Readiness, Risk and Responsibility

AI is quickly becoming part of everyday sustainability work, from reading disclosures to assessing climate risk. But the evidence from 2024 to 2026 shows a consistent pattern: adoption is moving faster than the skills, governance and trust needed to use AI well. This post brings together the latest data on where sustainability professionals stand, what can go wrong, which skills matter most and how responsibility for AI literacy is being shared across employers, educators, technology providers and regulators.

• 43%of corporate sustainability teams use AI in their workflows[1]
• 21%of those teams report significant time savings[1]
• 66%of workers rely on AI output without checking its accuracy[5]
• 6%of sustainability teams estimate the emissions of their own AI use[1]

1. Where AI is already used in sustainability work

AI adoption in sustainability is no longer experimental, but it is uneven. Reported adoption rates differ widely because surveys ask different people different questions. Executives tend to report company-wide use, while practitioners report what their own team actually does.

• C-suite executives using AI for sustainability efforts (2025)81%
  Company-level use, as reported by executives[2]
• UK financial firms using AI (2024)75%
  Up from 58% in 2022[3]
• Canadian federally regulated financial institutions using AI (2023)50%
  Up from about 30% in 2019; 70% expected by 2026[4]
• Corporate sustainability teams using AI in their workflows (2026)43%
  Nearly 60% plan to adopt or expand within a year[1]

The most common use cases

Most use today is back-office and process-focused rather than strategic. Among sustainability teams, the most mature applications are:

• Data ingestion and quality checks: collecting ESG data from many sources, flagging anomalies and filling gaps.
• Reporting automation: drafting narrative sections and mapping internal data to disclosure frameworks.
• Questionnaire responses: drafting answers to ratings agencies, investors and customers. One global technology company now answers about 30% of incoming sustainability inquiries with AI-generated drafts.^[19]
• Analysis and benchmarking: summarizing survey results, peer reports and regulatory updates.

• Data ingestion, quality checks and anomaly detection29%
  Most mature AI use case for sustainability teams[1]
• Reporting automation and framework mapping24%
  Drafting narratives and mapping data to disclosure standards[1]

The motivation is clear. Corporate sustainability teams spend an average of more than eight months a year on measurement and reporting, and half of organizations run their core ESG workflows with teams of one to five people.^[1] AI promises to free that time for decarbonization, risk analysis and engagement.

Climate and AI skills are converging

In finance, the same professionals are increasingly expected to understand both climate risk and AI. A global practice analysis of financial risk managers found that 40% already have climate-risk responsibilities, and 42% see AI and machine learning as a key area for continuing education.^[20]

2. The readiness gap

The central finding across studies is that broad adoption has not yet translated into value. Only about one in five AI-using sustainability teams reports significant time savings. The main reason is that time saved in generating content is spent again on manual verification.^[1]

What holds teams back

When sustainability teams that have not yet adopted AI explain why, access to tools is rarely the issue. Trust, data and skills are:

• Accuracy and reliability of outputs52%
  Top concern by a wide margin[1]
• Data security and privacy43%
  Confidential, client and portfolio data[1]
• Internal skills or training gaps37%
  Teams don't feel equipped[1]
• Budget or unclear return on investment35%
  Hard to justify without proven results[1]

Even among teams already exploring AI, maturity is low. A poll of environment, health, safety and ESG professionals found that most are still at the beginning:^[21]

• Just beginning to explore AI use cases41%
  Early experimentation
• Scaling AI across processes3%
  Moving beyond pilots
• AI embedded and delivering measurable value2%
  Fully integrated

This was a small conference sample, so these figures are illustrative rather than representative.

Individual use is running ahead of organizational support

Professionals are not waiting for their organizations. In one survey of corporate social responsibility professionals, 80% already use ChatGPT or other large language models in their daily work, yet 62% say their organization has no dedicated AI budget.^[22] When individuals adopt tools before policies, training and approved platforms exist, risks around confidentiality and quality grow quietly.

Trust and verification across the workforce

The broader workforce data explains why accuracy concerns are justified. A global study covering 47 countries found:^[5]

• Employees who rely on AI output without evaluating its accuracy66%
  Global, 47 countries
• Employees who have made mistakes in their work because of AI56%
  Global, 47 countries
• Canadians who have received any AI training24%
  Canada ranks 44th of 47 on AI training and literacy[6]

Canada's National AI Strategy identifies low literacy and low trust as key constraints on AI adoption.^[6] In Montreal, only 6% of business professionals surveyed by a local training program said they feel "very comfortable" using AI at work.^[23]

Understanding third-party AI

Financial institutions face a specific version of this gap: many rely on AI they did not build. In the UK, only 34% of financial firms report a complete understanding of the AI they use. Most of the rest report only partial understanding, largely because of reliance on third-party models.^[3]

• UK financial firms with complete understanding of the AI they use34%
  Most rely on third-party models[3]
• UK financial firms with only partial understanding46%
  The larger group[3]

Data quality also separates confident users from the rest. Reporting professionals who are confident using AI were almost twice as likely to have high-quality data (39% versus 21%) and were more likely to have AI governance or security policies in place (37% versus 23%).^[18]

3. How AI gets sustainability wrong

General-purpose AI tools fail in ways that matter specifically for sustainability work. Three failure modes appear repeatedly in research and practice.

Hallucinated facts and sources. Language models can produce confident, specific and false statements. In 2025, a major consulting firm agreed to repay part of a government contract after an AI-assisted report was found to contain fabricated citations and an invented court quote.^[7] In sustainability analysis, research on AI tools for reading sustainability reports documented a model attributing climate-oversight responsibilities to a board committee that the source report did not support.^[8]

Optimistic reading of corporate self-reporting. Sustainability reports are written by companies about themselves. Research found that general-purpose chatbots, without specialized tuning, tend to read these reports optimistically and are less critical of vague commitments and potential greenwashing.^[8]

Turning plans into achievements. A 2025 study showed that a language model asked to rewrite climate disclosures can easily turn planned actions into achieved ones.^[9] For anyone using AI to draft or assess transition plans, this is a direct greenwashing risk.

Domain-specific systems grounded in source documents perform much better. Academic studies report high accuracy for tools designed to extract ESG metrics directly from reports.^[24] The difference is not whether AI is used, but how it is designed and checked.

Why reliability matters beyond chatbots

Reliability is not only a chatbot problem. A benchmarking study of physical climate-risk data providers found that their assessments can diverge materially for the same assets.^[25] Sustainability professionals increasingly need to ask the same questions of any AI-driven score: what data it uses, how it was built and whether its results can be audited.

4. The five core AI literacy skills

Regulators have started defining AI literacy. The EU AI Act describes it as the skills, knowledge and understanding needed to deploy AI in an informed way and to be aware of its opportunities, risks and possible harms.^[12] A Canadian financial-sector forum defines it as understanding AI's capabilities, limitations, risks and responsible use.^[14]

For sustainability professionals, these definitions translate into five practical skills:

1. 1
   Critically evaluate AI outputs. Treat every AI-generated figure, citation or claim as a draft until it is traced back to its source.
2. 2
   Detect hallucinations and bias. Recognize the typical failure modes: invented references, optimistic readings of self-reported data, and planned commitments presented as achieved.
3. 3
   Protect confidential information. Know which tools are approved, what happens to data entered into them, and where client or portfolio data must never go.
4. 4
   Assess the reliability of AI-generated sustainability insights. Understand how models and third-party scores are built, what data they rely on and whether they can be audited.
5. 5
   Balance efficiency with responsible, low-footprint AI use. Choose right-sized tools, avoid unnecessary generation and account for AI's own environmental impact.

5. The environmental footprint of AI

For sustainability professionals, AI literacy includes understanding AI's own environmental impact. AI for sustainability must also be sustainable AI.

• ~415 TWhData-centre electricity use in 2024, about 1.5% of global demand[10]
• ~945 TWhProjected data-centre use by 2030 in the IEA base case[10]
• >90%Share of a GPU's lifetime electricity used for inference (everyday use)[1]

The International Energy Agency projects that data-centre electricity use will more than double by 2030, slightly exceeding Japan's current electricity consumption, with AI as the most significant driver. The agency stresses that these projections are highly uncertain.^[10]

What one prompt costs, and why scale matters

One provider's own measurement puts a median text prompt at about 0.24 Wh of electricity, 0.03 g of CO₂e and 0.26 mL of water.^[11] These are self-reported figures for text prompts only. They exclude image and video generation and model training, so they should not be treated as an industry benchmark.

The lesson is that a single query is small, but impact grows with scale and with everyday choices: which model is used, how often and for what purpose. Because most of a GPU's electricity goes into everyday use rather than training, users have real influence.^[1]

Footprint is rarely measured

• Sustainability teams that estimate or report their AI emissions6%
  Very few measure today[1]
• Sustainability teams that plan to start32%
  Intent is growing[1]
• US sustainability leaders who call AI’s environmental impact a major factor13%
  Still a low priority[26]
• CSR professionals who say companies should disclose AI's footprint78%
  Strong support for transparency[22]

The gap between belief and practice is striking: most professionals support disclosure, but very few measure. Part of the reason is the lack of standard units and limited transparency from AI providers.^[1]

6. Who is responsible for building AI literacy?

Responsibility is spread across four groups of actors, each with a distinct role.

Employers and financial institutions

Employers are where AI risk materializes, and regulation increasingly points to them.

• European Union: Since February 2025, organizations that provide or deploy AI systems have had to take AI-literacy measures for their staff. The European Commission has said this applies even to off-the-shelf chatbots and that relying on a vendor's instructions is generally not enough. National supervision began in August 2026, and the obligation was amended in July 2026 to require organizations to support staff AI literacy.^[12]
• Canada: OSFI's Guideline E-23, effective May 1, 2027, brings AI and machine-learning models, including third-party and vendor models, under enterprise-wide model risk management at federally regulated financial institutions.^[13]

Employers also report demand for more support:

• Investment employers who want industry-wide AI standards85%
  Strong demand for shared guidance[17]
• Employers who say their workforce seems anxious about AI60%
  Readiness is emotional as well as technical[17]
• Employers who report resistance to AI adoption48%
  Training builds confidence, not just skill[17]

Universities and professional training organizations

Education providers are adding AI content, but mostly in separate programs:

We found no Canadian credential that yet combines AI literacy and sustainable-finance practice in one program.

AI developers and technology providers

Technology providers control the features that make AI literacy usable in practice: traceable citations, audit trails, published accuracy metrics, data-handling guarantees and energy reporting. Per-query footprint disclosure has begun to emerge but remains voluntary and non-standardized.^[11] Providers can support literacy, but European guidance is clear that organizations cannot outsource it to vendor documentation.^[12]

Industry associations and regulators

In Canada, a financial-sector forum led by OSFI and the Global Risk Institute introduced the AGILE framework in March 2026: Awareness, Guardrails, Innovation, Learning and Ecosystem Resiliency. Its Learning pillar calls for building AI skills at every organizational level through continuous training and collaborative initiatives.^[14] Canada's National AI Strategy also commits to a national AI literacy initiative offering free entry-level training.^[6]

A timeline of key milestones

7. A layered model for AI literacy

The evidence suggests that AI literacy works best as a layered responsibility, with a clear owner for each layer rather than a vague shared duty.

Key takeaways

1. 1
   Adoption is broad, but value is thin. Most sustainability teams use AI for back-office tasks, and only a minority see significant time savings.
2. 2
   The main barrier is trust, not access. Accuracy, data security and skills are the top concerns.
3. 3
   Sustainability work has specific AI risks. Hallucinated facts, optimistic readings of self-reporting and planned actions presented as achieved results can all amplify greenwashing.
4. 4
   AI literacy for sustainability includes AI's own footprint. Very few teams measure it today.
5. 5
   Responsibility is layered. Employers own daily practice, regulators set the floor, educators set the standard, technology providers make outputs verifiable, and associations convene.

AI can make sustainability work faster, deeper and more accessible. Realizing that potential depends on people who know when to trust AI and when to challenge it.

Continue reading on Lucid Insights

AI literacy is one part of a bigger question: how AI can serve sustainability without creating new risks of its own. These earlier Lucid Insights pieces explore that question from other angles:

• AI for investors has to be sustainable: why the AI behind investment decisions must account for its own environmental and social costs.
• Who is accountable for AI governance in sustainable finance?: how responsibility for AI is shared across boards, teams and technology providers.

Explore all articles at lucidaxon.com/en/insights.