Using AI for ESG reporting and data-driven decision-making

Hi, everyone. Uh welcome, good morning. Uh thank you for joining our session. First thing Monday morning. Uh I'm Aditi Suresh, Sustainability Specialist with the Worldwide Specialist Organization at AWS. I provide sustainability expertise to develop and bring to market scalable solutions on AWS and help our customers achieve their sustainability objectives.

Uh this morning, uh I am joined by Brian Rowe who is the Director of Information Security and IT at the Pacific Company, who are industry leaders in waste, recycling and supply chain solutions. And today we'll be sharing more about how we can use AI for ESG reporting and data driven decision making with that. Let's get started.

So in today's session, we'll kick it off with an overview of the ESG trends and challenges followed by the role that AI plays in ESG reporting and insights. We'll then have a look at how AWS AI capabilities can be leveraged for ESG reporting. We'll then dive deep into a couple of uh use cases and we'll show some art of possible demos as well. I'll then hand it over to Brian uh to share more about the Red Pacific Company's ESG reporting journey with us.

So you've probably heard of the term ESG which is Environmental, Social Governance. I'm not going to delve deeper into it, but in a nutshell, these are non financial factors ranging from carbon emissions, workforce diversity. so social impacts and so on. And financial institutions are increasingly using ESG data as an indicator of risk and performance.

So if you look at the ESG trends, uh regulations are here, the landscape has evolved tremendously over the past couple of years, we have regulations such as the Corporate Sustainability Reporting Directive, also known as CSRD, which have been passed in Europe and companies are increasingly being mandated to report on the ESG data.

The second trend we are seeing is there is an increase in exposure of risk of litigation. So there are penalties for non compliance and there is a reputation and operational risk associated with EU reporting now. And finally, there is a focus on evaluating the supply chain. Further looking at transparency, looking at the supply chain impacts emissions across the value chain.

And what we found is, you know, it's no surprise, 95% of organizations agree that technology is going to be the key enabler for successfully implementing EU reporting processes.

Now ESG reporting can be complex and burdensome organizations face many challenges when it comes to collecting, analyzing and reporting on ESG data. And the fundamental challenge is data like there are times when some of this data isn't being collected at all. And when it is, it resides in silos across different systems and even between different teams. So it is extremely time and resource intensive to collect all this data and then consolidate and report on it. And what we found is that 86% of companies are still recording their data manually in spreadsheets.

The other challenge is there is lack of unified frameworks for ESG reporting. There are so many different frameworks that with each country, there is a specific regulation. However, the metrics remain the same. So it becomes challenging to figure out like which ESD data to report, where to report whom to disclose it to. And then we have inconsistency in data quality and reliability as well. This results from immature data management practices and that limits our ability to derive meaningful insights at scale.

Ok. Now let's take a look at how AI can be a game changer for managing these ESG efforts and play a key role in addressing those pain points. I think everyone in the audience is familiar with AI and what machine learning is, but I'm going to double click on a couple of subsets of machine learning. One of which is deep learning. This is a type of machine learning which uses a technique called deep neural networks. So these it replicates how the human brain functions. So we are able to address more complex use cases than was previously possible.

And finally, we have generative AI, these are powered by large language models that are pre trained on vast amounts of data and with generative AI. Now we are able to create new content. So this includes conversations, text, images, videos, and stories, and so on.

Now, taking an ESU reporting lens on this. So how can we use AI to help move the needle for EU reporting? So we've categorized the use cases by key elements of ESU reporting. So from the left to right, you can see that on the very left, we have automated data collection. So we can use AI to automate document processing for various EU data sources. So these could range from, for example, utility invoices, different like interpretation of sustainability documents, standards and regulations.

Then we can use AI to automate information gathering from different silos and unstructured data sources. Uh you know, uh use AI to classify a SC data and standardize it. And then once you've collected the data, we want to bridge the data gaps and build models on that. So, you know, we can leverage AI for, you know, filling data gaps automatically by using industry validated proxy data. We can use ML models to streamline the ESGKP modeling build, you know, use generative AI to rapidly build models based on evolving ESG methodologies.

Then third, we have uh check for compliance. So once you've collected the data calculated your metrics, uh we can use AI to query and summarize different sustainability disclosure requirements, uh map those ESG metrics to multiple frameworks and then help third parties audit ESG reports. So accelerating their audits by analyzing multiple documents and creating a gap analysis of the disclosure requirements.

And finally, we have generated reports and insights. So we can use AI to automatically build different ES reports based on the framework of our choice. We can use AI ML models for anomaly detection, trending, alerting and so on. Then we can also use AI ML for actionable intelligence. So you know, for example, benchmarking data for supplier selection and trying to identify which supplier has lower impact.

So with AI driven ESG reporting and insights, we are now able to unlock many benefits. So first is you can reduce time and effort. So it frees up our time and resources cost savings. So we are going to reduce compliance cost by streamlining our reporting efforts. Then we have a proof and transparent reporting. We are better able to manage our operation and reputational risk. And then finally, we're able to accelerate time to insight. So we are able to gain an increased visibility and accountability across our E initiatives.

Now, let's have a look at how AWS AI capabilities can be used to accelerate our ESU reporting efforts. So as you see in this slide, we have a suite of services across AI across our AIM stack at AWS. We are innovating on behalf of our customers. And we have the broadest and deepest set of machine learning capabilities for builders of all levels of expertise. So our goal is to remove the undifferentiated heavy lifting and help our customers move faster.

Now, I'm gonna show you how we can map key AWS services to key functions required for ESU reporting. So the first uh we have here is intelligent document processing and search. So we have Amazon Textract, which is our AI ML service that you can use to automatically extract data from utility invoices such as electricity, water and waste and you can use it for your EKP calculations. Then we have Amazon Kendra, which is our AI powered intelligence search for faster answers from siloed and unstructured data sources.

Then we have data transformation. So this is required for ESGKP modeling. So we can use Amazon SageMaker for streamlining our modeling effort. So you can, you know automate emission factor selection for KP modeling. We also have our new geospatial ML feature in SageMaker which can be used to analyze biodiversity indicators like deforestation risk.

Then we have Amazon CodeWhisperer which is our generative AI service that you can use to quickly build ESG calculations based on rapidly evolving regulations, then we have data validation. So in this case, if we can use, we can build a proprietary ML models on SageMaker to detect errors, anomalies and trends and alerts. We can use generative AI on Amazon CodeWhisperer to benchmark and fill data gaps with proxy data.

Then we have data analytics and insights so we can use Amazon Forecast to forecast our ESG metrics. We can use generative AI to identify hotspots, recommend actions to improve ESG performance. And then we've launched our generative BI capabilities within Amazon QuickSight that where you can use natural language pros to build interactive dashboards.

And finally, for ESG reporting, there is a need to create a lot of content so that there is a need for content creation, summarization and insights. So we can use Amazon CodeWhisperer here to automatically build ESG reports, query different sets of sustainability regulations, summarize them. Accelerate our third party audits with ESG reporting disclosure gap analysis.

Now we are going to dive deeper into a couple of use cases and see how AI can help enhance ESU reporting and insights. So for today, we are going to explore three use cases which are highlighted in yellow. You've seen this slide before and we are going to share some like a glimpse of the art of possible demos and see examples of how we can use AI how we can use AWS for ESG reporting. So let's get started for the first use case.

This focuses on how generative AI can be used to query and summarize sustainability documents for quick insights. So today, sustainability professionals and financial analysts manually retrieve and analyze dense regulations and documents which can be over hundreds of pages long. And this is extremely time consuming, practically infeasible and very costly exercise. So there is a need to rapidly and accurately analyze large volumes of text based information, summarize the text and draw insights. For example, look at what are the different ESG disclosure requirements, what data needs to be collected, how to calculate the data.

So we'll see an art of possible demo where we use generative AI to query a couple of sustainability documents and extract insights out of those. And this is going to help us accelerate our time to insights and save on costs and cut down on standards of interpretation time.

So we built a proof of concept here. And what the demo will feature is a chatbot where we are going to query and access insights from key sustainability standards such as the Greenhouse Protocol, which is a standard that provides guidance on a carbon footprint calculation methodologies. Then we have also preloaded the demo with sustainability regulations such as the CSRD which is a Corporate Sustainability Reporting Directive. And we build this proof of proof of concept using Amazon CodeWhisperer, which is our fully managed service that provides a way for you to build and scale generative AI applications using foundation models.

So for our proof of concept, we've used um an Anthropic Cloud we two foundation model and we also use a RAG approach which is retrieval augmented generation within the framework to augment the context with relevant sustainability data from documents.

So I'm going to quickly move on to the demo here. So in this example, we are going to ask questions on the Greenhouse Gas Protocol, which many of you may be familiar is over 200 pages long. So this is the standard that provides us with a methodology to calculate the carbon footprint.

So we're going to ask a question on how to calculate scope three emissions from purchase goods services. So these are for example, this scope three carbon footprint is the emissions from your supply chain. So you'll see that it's able to quickly scan the document and it retrieves relevant information and it's accurately summarizing the top four methods to calculate the scope three emissions. So we have for example, supplier method, hybrid method and so on.

So now let's ask a follow up question on this and ask the model to elaborate on one specific method with the help of an example. So we're going to take the example of a supplier method and ask the model and say that hey, if i were a t-shirt company and i purchased cotton from a supplier in china for manufacturing clothing, how do i calculate my carbon footprint? So you will see that it's able to coherently provide responses like a step by step responses based on the example that we requested. So you can see that there is a six step process and it also tells us that, you know, we need to check for data quality, what type of data we need to collect from the supplier.

Moving on. Now, let's switch to an example where we compare two different documents. So there is a need for sustainability professionals to review multiple frameworks at a time. So in this example, we are going to see how we can compare two regulations. We have taken the example of CSRD which is a Corporate Sustainability Reporting Directive and the new proposed Climate rules and regulations.

So we are going to ask the model to summarize the key differences between these two regulations and in particular, walk us through those key differences based on the scope, the timing, the methodology and so on. And as you can see the model is able to review those two last documents. These are over 700 pages long and it summarizes the text and shares insights with us.

So you can see it's looking at the methodology, scope, materiality for the scope, you know, it's accurately saying that C the CSRD scope covers a broader range. We're looking at, you know, different ESG topics whereas the SECS regulation is focused on climate risk

Now, uh we can also see if it's able to retrieve, you know, citations and references. Uh so I'm going to ask a question on say you know what is double materiality. So how is double materiality defined? And we're going to ask the model to pull up the uh cite citations and the page numbers and see if it's able to return it accurately.

No. So you'll find that it's able to cite the page numbers and the sections. So this is going to significantly cut down the time spent in standards interpretation for sustainability experts and auditors.

Moving on to another example. Uh here we are going to query corporate sustainability reports. So we have preloaded amazon's sustainability report and we're going to ask a question on what is amazon doing to reduce its transportation emissions?

Um and we're gonna see if it's able to summarize all the different initiatives for amazon and our amazon sustainability reporters are over 80 pages long and you can see it's quickly able to uh summarize and draw insights for us. And like for example, is talking about our initiative like our investment in ev vehicles and you can also see have a look at the metadata and figure out and check for the references.

Mm so moving on to the next use case, this use case looks at how we can calculate supply chain emissions faster using ml powered emission factor recommendations.

Now, as we saw in the previous uh demo example to calculate supply chain emissions, that is scope three emissions, we require supply chain purchase data. So we require data such as you know, transactions for example, the dollar amount spent in products procured or the mass of volume that's procured from suppliers. And then we need to translate this data into a carbon footprint using what we call as emission factor.

So today, sustainability experts are manually mapping these product names to an appropriate emission factor in order to calculate scope three emissions more accurately. So we have a solution for this. Uh we have an ml powered model where you can automatically map these uh product names and associate them with the appropriate emission factor.

So this is going to help us reduce the time and effort in modeling as well as you know, this is a scalable solution when we have millions of skews that we need to model. And it also saves on time and effort and helps improve the accuracy of the scope three assessment.

So let's have a look at the demo. Now there are several emission factor databases and they vary by geography and sector. So in this example, we've selected the ueio database. Uh so this is a freely available government database and it provides carbon emissions per dollar of the product purchased.

So let's start uh uh emission factor mapping job. Uh what I'm gonna do is we're gonna add uh three different product names. So the first one we're doing is a 15 inch laptop and in the description, we can enter, you know, 1000 laptops. The second product we i added is t-shirts 1000 t shirts. And then finally, we're gonna switch it up to another sector, look at polyethylene or pe t.

Now, the goal is to see um the accuracy of the emission factor mapping. Um so as this uh model uh runs in the background, what it's gonna do is is gonna look at the us cio database and provide us with the top five mappings or recommendation and we can still have a human in the loop to validate whether it's an appropriate emission factor or not.

So in this example, for the laptop, it's selected electronic computer manufacturing, which seems uh right then for the second one, for t shirts, we have a better matching. So we like we can choose apparel manufacturing. And then the third one for polyethylene, you look at it says synthetic uh uh fiber and filament manufacturing with a high level of confidence and you can see it provides the emission factor, the carbon emissions per dollar.

So when we run the impact assessment, now it's going to return the uh scope three emissions for us for each product. So in this way, as a sustainability uh expert, we are significantly able to improve the efficiency of the modeling and also improve our accuracy of calculating scope three emissions.

Now let's look at our final use case. So this is automatically extracting data from utility invoices to accelerate esg reporting. Now, one of the esg reporting requirements focus on utility consumption. So there's water, waste, electricity and so on and we also need to report on the associated emissions with those.

So as we saw earlier, the biggest challenge today is manual data collection. So today, a lot of operational managers are manually retrieving this data from utility invoices and entering those into spreadsheets. So we have a solution to automate this. Uh we can use amazon textract, which is our a iml service that can read and automate document processing to accurately extract data from documents.

So this is going to help us reduce the manual effort and achieve cost savings and improve accuracy. So in this demo, we're going to see how our aws partner flex zero is using amazon text track to automatically extract the utility data from electricity invoices in order to process what we call a scope two emissions. So this is a esg reporting requirement.

So the scope two emissions are emissions associated with purchase electricity. So what we're gonna see is within the flex zero portal, uh you can upload any utility invoice in any format. And uh the amazon text track service is gonna help extract the kilowatt hours of electricity consumed and then we'll apply an emission factor uh which is, which gives us a carbon intensity based on uh the electricity grid mix of a region and then we can calculate the scope to emissions.

So this is our flex zero data entry portal. So we are gonna upload a sample utility invoice here. Uh for now, for just for the sake of this uh demo, we are gonna upload a sample of ppgne utility invoice and we're going to run the uh analysis and we can also uh do this in bulk. So we can upload uh you know invoices in bulk and also for many different languages and it's going to run this and return the kilowatt hour of electricity that was consumed.

Now, amazon textract also has a service where uh you can extract this information using uh queries feature as well as as you can see, the model has returned the kilowatt hour of electricity consume. Now, amazon text, right, as i mentioned earlier has a service where you can also for the same utility bill, you can use a query to extract the data using our queries feature.

So in this case, we're gonna ask the same question like how many kilowatt hours were consumed and is going to return the response. So this can be uh used, you know, for example, there are times when you don't have uh the utility consumption data and you probably need to use cost to estimate your uh scope to emissions. So you can see it uh return the response here.

And now we're going to ask another question on the total amount that was charged. So we can also use spend data to assess the emissions associated with electricity. And you can see it's able to extract the uh total cost data successfully.

Now, that's a wrap up of our demos. Um I'll now hand it off to brian, who's gonna share more insights on how re pacific company is partnering with flex zero to accelerate esu reporting. Thanks, brian.