Loading...
July 24, 2026 · By Brandon Toews, Stephanie Bertels, and Wayne Moodaley
Share this post on:LinkedIn
Sustainability teams are exploring the potential of Artificial intelligence (AI) to support their work. Implemented well, AI has the potential to enhance decision-making, increase productivity, accelerate sustainability-oriented innovation in your operations and value chain, and reduce costs. Companies have been using AI to cut emissions and improve energy efficiency; better understand and prioritise impacts throughout their operations and across their value chain; engage in more sophisticated trend analysis and forecasting (for instance, to identify ways to mitigate the harm of extreme events); support setting more credible targets; provide real-time monitoring; and strengthen sustainability disclosures.
However, AI also carries risks that need to be understood and managed. This includes the risk of replacing jobs, amplifying the scale, scope, and consequences of bad judgement, or creating and intensifying negative impacts on communities and the environment.
In this blog, the first in a three-part series, we set the stage by exploring recent advancements in AI, key trends that are shaping the narrative around AI implementation, and how organisations and sustainability professionals are starting to use AI to support their work. In our second blog, we explore the impacts of AI on nature, community resilience, and the rights and wellbeing of workers. Our third blog offers precautions and recommendations to help your organisation limit its exposure to AI-related risks and to prepare for evolving assurance expectations.
AI is rapidly advancing, and it can be challenging for anyone – including experts – to keep pace with emergent AI technology. Here is a summary to help you catch up.
Over the past few decades, we have seen AI progress from simple rules-based systems to increasingly advanced machine learning and deep learning systems to the iterative advancement of foundation models.
More recent developments span several distinct capabilities. Generative AI moves beyond analysing existing data to creating new content (text, images, audio, and code) and includes conversational AI such as chatbots and virtual assistants. Predictive AI combines statistical analysis with machine learning to identify patterns and make forecasts. Extractive AI retrieves and structures information from documents, charts, tables, and other media. Agentic AI is different: rather than a way of processing data, it is a system capable of autonomously performing multi-step tasks on behalf of a user or system, often by building on the capabilities above.
One example of a foundation model is a Large Language Model (LLM), which is an AI system trained on a very large corpus of textual data so that it can interpret and generate human-like responses. An LLM learns statistical patterns in language and uses them to produce plausible outputs. It helps to think of an LLM less as a librarian looking up facts, and more as a fluent improviser who has read extraordinarily widely. The LLM generates the most likely continuation of your prompt rather than retrieving a stored answer. This explains why an LLM can ‘hallucinate’ and return incorrect information with a confident tone. LLMs are also unable to interact with other tools or learn new information after their training is complete. This is because the knowledge of an LLM is frozen at its training cutoff – in other words, LLMs do not learn, evolve, or update their understanding based on new interactions with users.
Current LLMs also lack persistent memory. They are fundamentally stateless functions; each query is self-contained, with LLMs receiving the input of a query and context and producing outputs based solely on that input and their training. Features that appear to "remember" you are generally handled by the surrounding application, not the model itself.
The implications of agentic AI for businesses and sustainability professionals are significant.
The distinction between LLMs and agentic AI rests primarily with the ability to engage external systems and tools to perform more complex, multi-step tasks. An LLM generates outputs in response to specific prompts, and its output relies entirely on its training data and the inputs provided. An agent wraps a model in additional architecture, such as memory, tools, orchestration, and safety systems, so that it can retain context across interactions, adapt to a user's preferences, and pursue a goal over several steps.
Another key difference between LLMs and agentic AI is autonomy. LLMs can generate summaries and code and respond to queries but lack independence. An agent can plan, execute, and manage a sequence of tasks, extending the capacity of the people and teams it supports.
Put simply, LLMs are the underlying models that generate text; agents are systems that use LLMs in addition to an architecture that includes memory, tools, orchestration, and safety systems; and interfaces are the UX layers that allow users to interact with them. Agents elevate the language capabilities of LLMs into broader, more open-ended use cases, albeit at greater, and growing, cost.
Forecasting AI advancements remains a challenge, but one thing is clear: the pace of change is accelerating.
In a few short years we have seen the technology evolve from answering simple and narrow questions and generating individual pieces of creative content to conducting comprehensive investigations, providing sophisticated analyses, and managing complex and aligned content strategies and coordinating complex workflows.
At present, and looking forward, we can expect further progress in multi-agent systems, industry specific models, increasingly autonomous systems that operate with limited human oversight, and further advancements in AI in robotics.
In spite of – and partly due to – the lofty and often utopian promises of AI technology, and alongside its real and expanding capabilities, there is growing evidence of skepticism and mistrust of AI by the general public.
Recent survey data indicates that despite widespread integration of AI tools, public sentiment – especially among Gen Z – is shifting in a negative direction. Since 2025, Gen Z’s anxiety has held steady at 42%, but anger related to AI has increased from 22% to 31%. Meanwhile, excitement has declined from 36% to 22%, and feelings of hope have declined from 27% to 18%. People also consistently say that they distrust AI-generated writing, which is contributing to growing efforts to conceal the use of AI in research and other publications.
These emotions and impressions largely center around the world of work. 48% of employed Gen Z respondents believe the risks of AI in the workforce outweigh its benefits, and the more workers use AI, the less they trust it. Despite various advancements in the technology, confidence in AI is collapsing – there has been a 35% decrease in confidence among baby boomers and a 25% drop among Gen X workers since 2025.
Meanwhile, we are seeing a growing number of companies push for the implementation of AI, including adding AI agents into their org charts, which is further exacerbating and amplifying feelings of anxiety. Whereas leaders fear missing out on the benefits of AI, staff increasingly fear becoming obsolete or being made redundant. There has been a rise in reports of workers actively sabotaging AI tools and related outputs to preserve their own roles and responsibilities.
AI carries the risk of amplifying the scale, scope, and consequences of biases and poor judgement. We are seeing a steady stream of new research highlighting the scale and scope of poor judgement and “deskilling” caused by AI. There is growing evidence that the use of AI tools may cause a measurable decline in cognitive abilities among children and adults, and rapidly degrades the abilities of professionals, leading to widespread deskilling even after short periods of reliance on AI tools.
Organisations face a growing imperative to preserve the capacities they have invested in and developed, in part due to mounting liability risks related to AI-generated outputs, such as the answers they provide to customers or the contractual agreements they make. Organisations must ensure that staff remain competent and capable of credibly overseeing and reviewing the work produced by AI.
For the past few years, many companies have been encouraging – if not mandating – experimentation and integration of AI.
More recently, in the wake of soaring subscription and token costs, companies are beginning to place restrictions on AI use.
Further, the growing scope and scale of AI use and advancements in AI technology are rapidly increasing token usage. Token requirements vary by task, AI model, and implementation, and companies are increasingly using AI for agentic tasks that use many more queries in the process of answering complex questions – a process that can consume 1000 times as many tokens as standard queries.
Surveys suggest that the vast majority of sustainability professionals are already using AI in some capacity in their roles, such as for extracting and assessing data across filings and systems; ESG disclosures; measuring, monitoring, and forecasting impacts; carbon accounting; risk analysis and scenario planning; and for supply chain optimisation.
Due to increasing use and excitement around AI, as well as companies feeling pressure to “keep up,” a growing range of specialised AI tools purpose-built for ESG-specific tasks are entering the market. It is important to understand, however, that not every use-case is going to be appropriate, and that no particular system is likely to meet all of your organisation’s needs.
While AI can help with drafting disclosures, this may be the least of its capabilities and value to sustainability professionals.
Instead, AI provides an opportunity to look at the sustainability reporting process and identify where there have been issues or gaps in the past, as well as clarifying where teams are finding information, organising it, reconciling it, and verifying it.
It can be especially helpful for reviewing current sustainability reporting processes, identifying “pain points,” and supporting specific parts of this workflow, such as finding information, organising it, testing consistency, and identifying gaps before reaching the drafting phase. In this way, AI becomes something that augments human judgement and review rather than replacing it.
AI provides excellent opportunities for supporting ESG data quality checks and reconciliation. AI tools can be used to flag missing, unusual, or inconsistent ESG data, as well as any other misalignments that may exist between internal data and external disclosures. It is also helpful for providing comparisons, as this does not require AI to determine whether or not data is correct, but rather to quickly identify inconsistencies and anomalies that humans can then investigate further.
Another effective use of AI is framework gap analysis, which entails taking draft sustainability reports and using AI to compare them against applicable sustainability-related disclosure standards and/or frameworks to identify potential shortcomings.
This first gap analysis can help your team to move more quickly from a broad compliance checklist to a more structured review of sections that may be incomplete or missing information, or that may simply require additional review. From there, another important possible use is building a claim and supporting evidence register for review. AI can be used to extract quantitative claims, forward-looking statements, and material commitments from draft disclosures and to link them to supporting documents. This can help your team to identify statements that have been made and to flag ones that may require additional evidence or additional review.
Finally, AI can be used to create an internal knowledge assistant. This assistant can be trained to query approved internal documents, such as prior period reports, documents, and summaries; surface relevant disclosures, methodologies, and source documents; specify and cite supporting documents, quotes, and page or worksheet references; and help reviewers trace outputs back to underlying documents. The assistant thus functions as a centralised, interactive repository for key information spread across different teams, folders, and documents.
In our next blog (coming soon), we will explore the impacts of AI on nature, community resilience, and the rights and wellbeing of workers.