AI Agents as Force Multipliers for Sustainability Leaders.

Good morning, everyone. Good afternoon, everyone, wherever you are in the world. My name is Inderjeet. I'm the co-founder of Pulsora. Really excited to be presenting the exciting developments of Pulsora to you, specifically around artificial intelligence and the opportunity that it provides. This is one of the most transformative and meaningful technologies we've seen in a generation, and it applies to sustainability quite a bit, as we've seen from our customers and our employees, who have actually come up with so many use cases.

Pulsora at a glance

First, just to introduce Pulsora. We are an AI-first platform for enterprise sustainability and carbon management. Full end-to-end regulatory reporting, data collection, carbon emissions and reporting, and performance management for the sustainability value chain. We have been recognized very eminently by analysts like Everest Group, Verdantix, and IDC. As you can see, we are in the leaders or the leader in this space, by far recognized by multiple analyst firms.

This is a smattering of our customer base. From this you can see that it's a true platform, not specific to any vertical. We have private equity firms like Apollo and BlackRock. We have manufacturing firms such as Xinyi Glass, the fourth largest glass manufacturer in the world. Manufacturing firms such as Molson Coors. Asset managers like Partners Group and Franklin Templeton. Chip manufacturers like ASM. And of course, retailers and luxury retailers like Tiffany. The point being that this is a true platform that applies to multiple verticals, and it scales and morphs to your vertical needs, rather than the opposite of making the customer scale to the platform.

How customers use AI today

So first, I'd start with a simple assertion. How are customers using AI today? There are a few answers we've seen across our customer base.

One, they're directly prompting large language models. I'm sure at this point pretty much everybody's doing this, just inserting things into ChatGPT or Anthropic's Claude, or Gemini, or your LLM of choice.

The second might be data extraction and collection. You might insert your sustainability report or maybe an Excel file and say, "Could you extract this data for me, perhaps even do some calculations for me?"

Once you start getting a little deeper, you might actually ask it to give you decarbonization strategies, insights, and recommendations.

By far the most common answer is: we're still exploring, and I don't trust the output of it. We're still exploring the use of it within the sustainability domain. Very rarely do you find "my AI agent handles everything," although, you know, give it five years or maybe even sooner. I think this is where the future is. AI agents are going to be handling a significant portion of operations.

And the last one: sometimes we do still hear "what exactly is an AI agent?", although very few times recently.

This is the reality on the ground from customers actually dealing with sustainability day in and day out. I'm sure you are in one of these buckets, depending on your journey towards artificial intelligence.

Why general-purpose LLMs fall short for sustainability

So then the question becomes: are general-purpose large language models like ChatGPT, Claude, Grok, Llama, Gemini, or pick your favorite, are they enough for sustainability?

Our assertion, drawn from real-life experience, is that sustainability AI needs two things that are missing from just directly using a large language model. One is the full enterprise context, which we can define exactly. And the second is trust and governance.

Overwhelmingly, the number one question we get from our customers is: "Can I trust what's coming out of AI?" I don't think they express the enterprise context question directly in a technical fashion, but they know the trust and governance is actually related to the richness of the context provided to the large language model.

Trust and governance

So let's talk about trust and governance. What does trust mean? Can I rely on the data, and can I also make sure the results are not biased? There's fairness, and there's no manipulation of the data. Essentially, no hallucination. We'll cover that.

From a governance perspective: is it accountable? Is there jurisdictional fragmentation? Is there concentration of power? Is there access and equity in standards and auditing? That's what trust and governance actually means.

Enterprise context

Now, from an enterprise context perspective, this is what you give the large language model. This means metrics, the data, the periods, the workflows, the temporal data, the audit trails, who approved what, for example. The transformations, calendarization, emission factors. Everything is fair game, but it should be structured and governed, and the relationships among the data are extremely important.

This is where just working directly with the large language model often falls short, because the relationships between the data are inferred, and that's where most of the problems lie. Providing a full enterprise context that is structured and governed from a sustainability application like Pulsora can provide a tremendous advantage here.

Second thing: organization hierarchies matter. Depending on who you are and what vertical you're in, from an investor point of view, it might be a fund, or a private equity fund, or a public investment fund that invests in companies. Companies have divisions. Divisions have locations. Locations might even have meters. Aggregation, transformation, and organization boundaries are tremendously important to sustainability. Not very apparent if you're simply using a large language model underneath the covers without any structured enterprise context. This is where an application like Pulsora can add a lot of value.

And of course, the emission factor datasets must be up to date. This is where the data itself matters, and the right emission datasets being picked up for a specific location, situation, good, or service is extremely important.

Audit trail and transparency: how the data got to where it is, extremely important to feed into the context. And of course, data must be complete, current, and correct. Without an enterprise context, what you're getting is a bad response out of bad data. This is why enterprise context is an extremely important component of artificial intelligence for sustainability.

The context graph

So what exactly is the context graph? What you see on the right-hand side is a picture that describes Pulsora AI. This is the architecture from a functional perspective.

The middle circle is the large language model. Frankly, one of the key differentiators for us is that we don't mandate the usage of a model. You can select your own model. We pick different models for different use cases as well. There are some large language models that might be really good for image processing analytics, others good for reasoning, others much better trained on decarbonization recommendations. That's what Pulsora AI in the middle for the large language model stands for.

Around this is the context graph. This is the most differentiating aspect, the most important aspect. It could be disclosure metrics, regulatory metrics, estimations, transformations, organization hierarchy, audit history, calendarization, transformations, greenhouse gas emission transactions. Everything that's needed for the large language model to be able to create the applications and agents that you see on the outer purple ring.

These are a smattering of them. We will show you a few today. For example, a data ingestion agent. A physical risk agent. A benchmarking agent that ingests publicly available data to benchmark your performance against your competition. Supply chain data aggregation. Public data extraction. Or maybe something more mundane but very valuable, like a pre-audit agent. We have customers spending a lot of dollars on auditing, but what the auditor does can be automated to a large extent by providing a pre-audit agent, so you can fix most of the issues before the auditor shows up. Or something like a decarbonization recommendation agent.

And of course, the easiest use cases are insights and analysis, the reporting agents, which have been around for a while. We've had them for a while. But what you see here is the expansion of the AI model to extend to what has traditionally been advisory and consulting tasks. That's what's different here. Our AI is not necessarily aiming to replace a sustainability consultant. It's meant to aid and assist to a large extent of what you would get from those services.

Carbon management

So that's the context graph. Let's take a few examples of sustainability AI.

The first is carbon management. High volume, messy, painful to get the data in and out, map emission factors, keep them up to date.

The enterprise context in this specific use case: you see a couple of screenshots on the left. They should be explainable and auditable. Cascading workflows is essentially a feature in Pulsora where the workflow can cascade at any level of the organization. This is what enables us to do data collection across an entire value chain that can start from an investor and go all the way down to a leader. You might be a manufacturing firm. One of our customers has 42 plants in 17 countries. A cascading workflow across all their facilities, divisions, and different goods and services they procure becomes extremely valuable, rather than only operating at the apex.

Data collection may be the customer's responsibility. An emission factor library. And natively integrated with reporting, so the same data is being fed into the AI context. It's natively integrated with reporting as well, so you see the latest, timely, accurate data being fed into the AI context. That's the enterprise context for carbon.

Examples of agents that might be high-value on top of this enterprise context. First is very mundane but extremely valuable: bulk upload of invoices. If you've been in sustainability for a while, this is one of the most mundane tasks and very painful because optical character recognition has been problematic and requires human intervention. Using artificial intelligence with an LLM doing this: upload thousands of invoices, and we extract the data for you. We can tell you where the issues lie. We can attach evidence automatically. That's it. Very efficient. Tremendous time savings.

Simple. But it goes from there to something more advisory, like a decarbonization agent. A company wants a path to net zero. We can model the path using multiple scenarios. Each scenario is composed of projects you might be doing. You might already be doing some projects, you enter them into the system. Or the AI can recommend projects based on your company profile that we know, including your size, your geographies, your vertical, what your competition is doing, taking that entire context into play. From there, you can create these pathways and model the output, essentially a marginal abatement curve. But because it's part of the entire sustainability platform, it's tracked over time. As your metrics improve, the goals are met or not met, and based on that, we can display the progress to sustainability leaders for specific scenarios.

Sustainability reporting

Now, let's move on from carbon to sustainability reporting. Here, the enterprise context is a bit different. It's not just large volumes of carbon transactions. It's modeling of a value chain.

The value chain in a private equity firm is from the investor in a specific fund to the portfolio companies in the fund. The model of the value chain for a manufacturer is all the way from their divisions to their manufacturing facilities, to their shipping locations, and all the way down to the meters. The enterprise context for the value chain for a retailer might be specific locations. We have customers modeling complicated multi-stakeholder value chains. For example, they might have one user or manager in Japan managing stores in Japan, South Korea, and China. They're responsible for collecting the metrics, or the primary data for that specific part of the organization hierarchy.

That also includes the workflow approvals, the revisions, the comments, and the full audit trail of who changed what. And by the way, this is not just human changes. It also means system changes. When you ingest data directly from a backend system, that could be environment data, social data, or governance data coming in from ERP systems, supply chain systems, procurement systems, with a full audit trail. That's the integrations into the entire enterprise ecosystem that provides the enterprise context.

Now, this might be lower volume but more complex than what you see in carbon. The point is that the context needs to morph to the specific AI use case. This is why working with a platform like Pulsora can actually add quite a bit of value here, because we know what the best context is to feed into artificial intelligence.

Based on this context for sustainability reporting, these are some of the agents and applications that make sense.

Data extraction. Instead of sending a workflow to a user saying, "Please fill in this data manually," you can simply say, "Upload your files." Could be Excel files, PDF files, a sustainability report, internal documents, Word documents, survey results. Just upload into the application. We extract the relevant data and build what we call in Pulsora the sustainability catalog from it. We do historical analysis of the data and build the historical data from these files as well.

Then could be a competitive monitor agent for mergers and acquisitions. What is the competition doing, and what's the risk profile of a potential acquisition? Or a briefing reports agent that could produce reporting out of it. Private equity firms use it for reporting to their LPs. You might have an agent that takes a single source of truth and reports it up to a mandatory regulatory agency or to a voluntary agency.

A DMA agent that allows for creation or recommendation of IROs (impacts, risks, and opportunities), backing them with the guidance from ESRS, for example. These days, sometimes double materiality is not required. You can just use single materiality, which might be just financial materiality.

Water and biodiversity risk agent based on WRI and other datasets. ESG data ingestion: you have backend data, simply upload into an agent, and we can ingest the data into the application for you. Physical risk agent. Supply chain agent to go into your supply chain, request primary data from suppliers, request from third-party sources. These become very rich extremely fast with artificial intelligence.

In the past, most of these agents and applications were separate applications you would buy, then integrate together. The point here is an AI-first platform allows you to do this in a very cost-efficient way in an integrated fashion, so you don't have to create the context, we do.

Demo: data extraction and benchmarking

Let's take some examples.

Here's a data extraction agent. Data is buried in PDFs and separate spreadsheets, becomes essentially usable sustainability data by simply uploading the file. We analyze the context. We know what the metrics are in Pulsora. Sometimes even more detailed than that: child metrics, dependencies, creation of tables. Stuff that is very involved to do, sometimes takes months for customers, can be done very quickly with an agent like this.

That is a benchmarking application. Some of our customers pay a large amount of money for benchmarking, and even then, it's not a continuous monitor. The intention here is that we can monitor the competition. You can upload your own context here, or we can create one for you, where we can start to benchmark specific metrics against your competitors.

The interesting thing here, with artificial intelligence, is that it's become much easier to infer the quality of non-quantitative metrics, the qualitative metrics: paragraphs, disclosures, and policies, very easily as well. With the ability to see the provenance of exactly where we got the data from and how we inferred specific scores. Very transparent, very auditable, and adjustable. What is done here is a sliding scale that allows you to normalize units and normalize intensities, specific metrics across a competitive set of companies based on number of employees, revenue, or other normalization or intensity-based denominators very easily. A very valuable application that can be done in a continuous fashion.

The simplest use case, but also very valuable: conversational AI. We believe at Pulsora that static dashboards are already gone. You can ask it a question, it will respond in real time with the latest data. It even suggests follow-on questions, completely auditable, where did you get it from, all the sources are present. It can draw the graphs for you. It can produce outputs in PDF, PowerPoint, and Excel. Reporting is the easiest use case. We've had it for a long time now. But it's very valuable. It's grounded in your data and traceable to your sources. That's the differentiation. It operates in real time to the context that exists in the application already.

The state of governance among customers

And governance. This is where most companies are. We have a very esteemed advisory board at Pulsora. We keep asking them questions on this. Trust in the data is the number one issue. There are a variety of gradations on how companies are dealing with trust and governance: still figuring out the basics, hitting risk or compliance roadblocks, pushing for more transparency and control. Very rarely do we find we have it figured out and ready to scale sustainability AI.

About six months ago, we would see hitting risk or compliance roadblocks quite a bit: "Don't feed any data into a large language model." That has largely gone away. Everybody is starting to use this technology. It's really about transparency and control at this point. This is where most of the investment we're making is going.

Demo: Sustainability Intelligence (Aramco walkthrough)

One second, let me check if there are any questions. Otherwise, I'll show you a couple of very quick demos of what these agents look like. Any questions so far? No questions so far. Okay.

I'm going to log into our corporate demo environment. This is our demo system, just the lay of the land of the application. This is the dashboard. You have your tasks, your disclosures. I'm not going to go into too much detail of the application itself.

The manifestation of these agents and applications is in an app store. The app store is essentially where you start launching these high-value agents and applications. Let's take a couple of examples.

Sustainability Intelligence. This application is designed to monitor regulations, standards, and competitors for you around the world. The first thing that happens: in this case, I've modeled one of our customers, Aramco. Based on just the website entered, we create the context. The context, in this case, is six pages and fourteen documents extracted from their entire website that mention anything about sustainability. Could be annual reports, 10-Ks, sustainability reports. This can be enhanced with your own context as well, but the point is we can extract it automatically.

Based on this, the regulation page allows you to go pick any regulations anywhere in the world. If you go, for example, to Asia, and you say, "Let's say I'm in Azerbaijan," these are the specific regulations available in Azerbaijan. Or in India, these are the specific regulations available in India. We make recommendations on this as well. Once you pick and choose, now you have a view-regulation page.

I'm going to, for example, California SB 261. A refresh of this icon uses artificial intelligence and builds a latest-news feed, which is very valuable in terms of what is happening with SB 261. We go after the exact source of the regulation.

The most interesting part is the impact analysis. The impact analysis tries to take the existing context of the application, which we've extracted from their publicly available reports, compare it to what the specific regulation requires, and then create a very detailed insight: what the topics are, what the current reporting is, what the requirements under SB 261 are, any identified gaps and notes, and then a specific set of recommendations. It even provides a very detailed impact analysis summary with full ability to go after the sources where the data came from. So there's no risk of hallucination.

This also gets to an interesting point made by one of our customers: "I don't want to simply look at a regulation from just the regulation's point of view. I also want to know what my competitors are doing with respect to that regulation." So in this case, the same analysis is provided for Aramco versus their competition: ExxonMobil, BP, and Shell, and specifically what SB 261 requires. Gives you a very easy way to look at exactly what competition is doing, what the regulation requires, and what our recommendations are.

By the way, this creates a digest on a periodic basis, every week or every month, and then emails it to the relevant stakeholders in PowerPoint, PDF, or HTML format.

Demo: Standards alignment (ISSB)

The next one is standards. This is where you might have the same idea, but let's say for ISSB. An alignment analysis uses artificial intelligence to compare the alignment of the current sustainability posture of the company in this context versus the specific requirements for that standard. It might actually give you a very easy view of what is required to conform with that specific ISSB requirement.

The output is very simple. It says what ISSB requires, the status is partially met for Aramco, what it meets, and what the gaps to address are.

Interesting when you start thinking about competition. In this case, let's see British Petroleum. The competitive details tell you specifically what the criteria are, measuring Aramco versus BP. Another feature that was asked for: analyze the competitor's progress over time. This is where you start to see the progress of how they've come on their sustainability journey from 2024, 2025, and 2026. The last three years. It tells you whether it has improved and gives you a bit of analysis on exactly what's happened. A very valuable application from a sustainability regulation and standards monitoring space.

Demo: Benchmarking

Another example, having to do with benchmarking. The key metrics extracted from the source company, the company that's using it, versus all of what they identify in their peer group. You can do major metrics versus all metrics. You can normalize this across specific dimensions like the denominator. It could be number of employees, amount of revenue, to give you a nice framework.

The AI analysis is interesting. It tells you the performance overview from the learning from artificial intelligence. In this case, it deals with significant variability when benchmarked against peers for scope 2 emissions. It also tries to get at Aramco's strengths: in circular economy, leverage waste management effectively, water usage, for example, against PetroChina. Very detailed and useful analysis for sustainability practitioners.

We don't stop at quantitative comparison. We get to narrative comparison as well, if you start to use the paragraphs and the policies that exist within these applications, within the documents and the context, and give you a score associated with that. The AI analysis is even more detailed. It tries to give you an analysis across what you can learn from how the competition is dealing with non-quantitative metrics, the softer disclosures. Very easy for us to do, and actually very valuable for a client to see in a single place.

The context matters here, because the entire thing is fed by the metrics and disclosures that are present in Pulsora from the company's context already. So we know how to measure, we know the audit trail for that, and that becomes the context. Then comparing it against the competition becomes much simpler.

Now, there are other areas in the application: alignment for ESRS or CSRD alignment based on the competition. Disclosure insights and transition monitoring. For example, customers might have net zero targets that are different from what this particular company has, and you start to track them. We continue to make this richer and richer.

Demo: Double Materiality Assessment (DMA)

I'll also show the DMA application. Within a disclosures agent, within CSRD, for example, you can click on double materiality. This is a composite of four specific agents. It creates your context. You are then recommended IROs. Each IRO can have a specific impact: applicability, what the assets are, what the transmission to look through, and details. Then it gets to the assessment part, where you can start to assess the impact materiality and financial materiality. Eventually you get to the materiality aspect of it, then eventually produce the materiality matrix.

This is all done with the IROs being recommended from artificial intelligence and allowing you to create a very simple way to manage your DMA, or an SMA if you wish to. There's no need to have to be a double materiality matrix.

This application morphs to your specific vertical as well. If it's a manufacturing company, it could be entirely different than an investment company versus a private equity firm.

A very quick summary, a short display of the demos for some of these applications. Coming back to our presentation: we can go on and on, and if you're interested in more detail, please contact us. We'd be happy to do much more detailed demos specific to your scenarios.

Accuracy and hallucinations: four techniques

Then the question does come up all the time: what about accuracy and hallucinations? Of course, you're never going to architect a system that is 100% free of hallucinations. You can minimize them to a large extent. By large extent, I mean de minimis.

The four techniques we use to do that:

One, system prompt engineering. This is where we've spent a ton of time to make sure this is operating in a box.

Two, source citations every time. Where the data is from, where the calculations were done.

Three, grounding in the context graph, so there's no going into the LLM without controls. We are telling the LLM how the data is related to each other, and we know it's accurate and timely.

Four, we incorporate confidence scores in everything. It becomes much more confident when there's a human reviewer.

We are never saying there's not going to be a human in the loop here. The intention is to make sustainability practitioners far more efficient at their jobs. Eventually, this does require legal oversight and oversight from humans as well. There's always going to be some, but it's going to be far more minimal.

The future: customers building their own agents

Now, how do you interact with this? What you saw in the application is a smattering of applications and agents that make sense. But the key thing we're moving into is that the context graph is so valuable that you, as a customer, will create your own agents, your own applications on top of it. You can choose your own, build it in Claude, Cursor, Replit, any vibe coding tools, what have you. Extremely valuable and very ingrained into your normal day-to-day life. This is where the future is.

‍