Loading...
September 22, 2026 · By Brandon Toews, Stephanie Bertels, and Wayne Moodaley
Share this post on:LinkedIn
Effective AI implementation requires a strong understanding of potential risks and impacts, strategic choices about the appropriate scale and scope of adoption, clear processes that support transparency, and appropriate governance and oversight.
In the first blog post in this series, we discussed the evolution of AI technology and explored how sustainability professionals are beginning to incorporate AI into their work. In our second blog post, we provided a summary of the impacts of AI technology on nature and society. This final blog post focuses on anticipating and mitigating the potential risks of AI and meeting evolving assurance expectations when using AI to support corporate sustainability disclosures.
AI should be used in a conscientious and targeted manner, with users and leaders understanding its limitations and impacts.
Fundamentally, AI is only as effective as the data that underpins its outputs. Data that is fragmented, incomplete, inconsistent, or otherwise varying in quality can make AI analysis difficult and its outputs subject to material errors.
AI tools continue to struggle with traceability, transparency, and explainability. AI algorithms and models are often deployed as “black boxes,” making it difficult to trace their decision-making processes, even for the AI researchers and engineers that design them. This creates challenges in understanding why certain outputs – especially erroneous ones – are produced. Even when model outputs appear to be correct, this lack of transparency introduces challenges when defending data and assessments from a regulatory and audit perspective.
Start by exploring your understanding and assumptions of what AI tools can and cannot do to support your sustainability efforts and identify and test potential high-value use cases before scaling. The best use cases are those where the workflow is bounded, repeatable, auditable, and subject to human approval. Workflow steps should be clearly defined, data sources should be controlled, success criteria should be clarified, and rigorous human review checks should be implemented at specific relevant points.
Professionals and organisations are responsible for any AI outputs that are acted upon. AI has tremendous potential to make work more efficient for sustainability teams, but this comes with a greater need for scrutiny.
It is essential that the process of evaluating, contextualising, and communicating content remains in human hands, guided by human judgement.
AI tools may provide inaccurate statements that are polished and sound plausible, making it far more difficult for reviewers to identify errors. These could include outputs containing specific information – such as numbers, values, or methodologies – that may be flawed, due to incorrect sourcing or simply due to confabulation (often referred to as “hallucinations”).
It is therefore important to treat outputs from AI tools as flags rather than as evidence. Key pieces of information should always be subject to human review.
AI tools are also designed to be sycophantic – inclined to tell users what they want to hear, a tendency that users often reinforce by rewarding or choosing responses that affirm them. The practical consequence is that AI is unlikely, on its own, to question the premise of a query or the intent behind it. Framing is also important – how you pose a question shapes both the answer you receive and what the system infers about your values. And, because the AI model draws on what it has learned from its training data, it may also simply have limited (or no) exposure to certain worldviews.
AI should not be used to replace experience or expertise – rather, it should be used to scale it and to enable talent to apply their judgement more broadly and uniformly.
Helping companies to operate in socially, environmentally, and economically responsible ways requires processing a broad range of information. Sustainability practitioners take in a wide variety of quantitative and qualitative data and apply a series of complementary lenses to assess the context – lived and experiential, theoretical and academic – to draw conclusions. When we use AI to do this work for us, we lose the most important part of the sense-making process: thinking.
A real and growing risk of AI use is the “cognitive debt” that comes with outsourcing to machines the work of critical thinking, of interpreting reality, making judgements on the world around us, and forming decisions accordingly. When we use AI to do this work for us, we risk degrading our skills and our ability to identify errors and apply our own judgment. Experience and expertise should be augmented, not replaced.
Sustainability practitioners take in a wide variety of information, data, and experiences to draw conclusions, whereas AI is a pattern-recognition engine bound by its training data and built-in assumptions. AI is not a substitute for human judgement, intuition, and instincts.
AI optimises against the value model it is trained on. An AI agent may pursue that objective at the expense of other factors that might seem obvious to a human decision-maker but that have not been explicitly identified as significant. It may overlook factors such as impacts on ecological and social systems, or the value of local and traditional knowledge, and instead may focus simply on maximising short-term financial value. It is therefore important to be explicit and include relevant environmental and social considerations when framing your query.
A key element to consider when designing and implementing your AI strategy is the development and deployment of a data classification framework for sustainability related data. For many organisations, sustainability data classification is less mature than it is for other types of operational data.
A good data classification framework does several things. It distinguishes between the data that employees can feed into personal, public, and enterprise AI tools, such as sensitive and legally restricted data; it defines permitted source documents and requires source-referenced outputs; it codifies version controls; it builds in human review checkpoints, calculation validation, and retention of prompts and outputs; and it assigns ownership to specific persons of published (both internally and externally) content that has received AI assistance.
Include your organisation’s use of AI in your impact materiality assessment process. For instance, AI use threatens to significantly increase the impact of knowledge-intensive work (and industries), such as administration, customer care, finance, law, education, and consulting by replacing lower-emission human cognitive labour inputs with higher-emission AI inputs. AI use should be included in your Scope 3 emissions profile, and – where it is deemed material – organisations may need to consider limits on non-essential AI use.
As models continue to evolve, a key part of proactively managing impacts stemming from AI use will be right-sizing the right tools. Identify the needs you are trying to meet and prioritise task-specific AI tools over those with higher relative water and emissions footprints. Reserve the use of more powerful and resource-heavy agentic systems for more complex, multi-step workflows. Simple, lower-impact LLMs can be used for predictable, stateless tasks.
Auditors are already thinking about AI, and there is growing scrutiny of AI in financial and sustainability reporting. Auditors will soon be asking which disclosure sections were AI-assisted and what the human review trail looks like.
You will need a well-organised evidence file with a clear review trail and a documented process showing how AI-assisted outputs were checked, approved, and retained. This can include explaining the tasks that were performed with AI, how the outputs were reviewed (and by whom), and assuring that AI did not make compliance conclusions, materiality judgements, or assurance decisions.
Be proactive in understanding and articulating your position on the use of AI in disclosures, as well as the language of how such use is described. If you will be introducing AI into your workflow, it is essential that you define this process early on and discuss it proactively with your auditors.
We will continue to track how sustainability teams are integrating AI into their workflows and how organisations are accounting for and working to mitigate the negative impacts of AI. Please reach out to share your experiences with us.