🦄 Making great presentations more accessible.
This project aims to enhances multilingual accessibility and discoverability while maintaining the integrity of original content. Detailed transcriptions and keyframes preserve the nuances and technical insights that make each session compelling.
Overview
📖 AWS re:Invent 2025 - Accelerating sustainability compliance with AI-powered document review (AIM237)
In this video, Kenta Sato, Solutions Architect at AWS, presents RAPID (Review and Assessment Powered by Intelligent Documentation), an open-source AI solution that transforms sustainability regulation compliance from a business constraint into a capability. He explains how regulations create market fairness for low-carbon transitions but face administrative bottlenecks due to complex document reviews. RAPID addresses critical AI challenges like "lost in the middle" and accountability gaps through a two-phase approach: AI-powered checklist extraction with human validation, followed by item-by-item document review using Amazon Bedrock agents. The serverless AWS architecture enables scalability from pilot to production, achieving up to 75% reduction in review time. The solution is available on GitHub and deployable via AWS AI Solution Box.
; This article is entirely auto-generated while preserving the original presentation content as much as possible. Please note that there may be typos or inaccuracies.
Main Part
Why Sustainability Regulation Matters: From Market Failure to Administrative Bottleneck
Hello everyone. Who in here loves regulation? Anyone? Yeah, kind of expected, right? But here's the thing: regulation is really important for driving sustainability transformation. It's not the regulation itself, but the endless paperwork and bureaucratic process that's probably the part you don't like, right? The EU is in the right session. Hello, my name is Kenta Sato, Solutions Architect at AWS. Today, I want to show you how we can flip that script, transforming sustainability regulation from a business constraint to a capability using AI.
Imagine regulation actually accelerates sustainability without slowing down your business. By the end of this session, you will have everything you need to get started, as we have open sourced the solution that you can deploy today. But first, why does regulation matter? Let me show you the world without regulation and with regulation. Without regulation, we face something called market failure in some sustainability areas. Take low carbon materials like low carbon concrete or low carbon steel. These materials tend to cost more than traditional alternatives. From society's perspective, it makes perfect sense to invest in those materials to protect the environment in the long run. But from individual companies' perspective, paying higher costs while your competitors are using cheaper alternatives doesn't always make sense if your consumers are not willing to pay for the premium.
But with regulation, everything changes. Regulation creates a level playing field where everyone must comply. Now, companies can invest in sustainability without being undercut by competitors. Those early adopters actually gain competitive advantage as they have already established relationships with their low carbon suppliers. So sustainability regulations are essential to guide the low carbon transition across the entire economy. These regulations are implemented across every major industry, such as renewable energy facility permits, manufacturing compliance, ESG investment reviews, or fleet electrification, and the list goes on.
There are different regulations across different industries. But they all share these common characteristics that create a huge administrative bottleneck. You deal with a high volume of complex documents that need to be verified against a large checklist that requires expert knowledge. Let me show you a concrete example: building permits applications for sustainable and energy efficient buildings construction. The applicant needs to prepare, say, twelve different types of documents that contain complex diagrams, charts, and text. The reviewers have hundreds of checklists that reference laws and regulations, and this checklist needs to be verified against these complex diagrams.
What does this mean for both sides? For applicants, it means they have to organize data for internal verification because if the application is rejected, that means additional rework, more cost, and more delays. For reviewers, they have to go through huge administrative work for document reviews. There is an economic risk, right? If the building construction delays too long, it risks slowing down the economy. So this kind of pressure makes the regulators put ambitious, strong sustainability regulations into place as they have to balance economic and environmental impact.
So that is the challenging part. Can we solve this with AI? Here is a straightforward approach: you have review documents and a checklist. Why not put them to generative AI and get instant review results? Simple, right? But when we tested this approach, we discovered there were critical problems that make it unsuitable for regulatory compliance workloads. The first problem is lost in the middle. When you put large context into generative AI, it is known to degrade performance. When you feed, say, two hundred pages of documents with one hundred items of checklist, the fiftieth checklist does not get the same attention as the first or the last one.
The second problem is the accountability gap. When the AI makes a mistake, who is responsible and who is there to correct it?
Introducing RAPID: A Two-Phase AI Solution for Intelligent Document Review
Regulators cannot simply accept "because AI said so" as an excuse. So if throwing everything at AI doesn't work, how can we solve this? This is why we developed RAPID: Review and Assessment Powered by Intelligent Documentation. RAPID is an open source solution that you can deploy today to streamline a large volume of document review. Let me show you the demo first, and I will explain how it works later.
We want to check if the new equipment installation in this demo complies with the energy efficiency standard. Here are the sample documents I'm dealing with: an Energy Efficiency Compliance Checklist on the left and an equipment specification document on the right. First, I'm uploading the checklist. You can just drag and drop the PDF file and click create. In the backend, RAPID works on extracting the checklist. Before using this checklist, an expert can review whether each item has been extracted correctly from the original document. If the check item is too ambiguous, RAPID will prompt you to clarify with suggestions. In this example, the checklist requires compliance with energy efficiency standard, but it doesn't have the specific details on the standard. So you might want to add detailed criteria so the AI can make an informed assessment.
Let's understand what just happened in the demo. This is phase one of using RAPID: AI-powered checklist extraction. The AI takes the checklist PDF file and extracts it as structured data. Then a human can validate whether the checklist has been extracted correctly. That's phase one. Let's see phase two. Now I'm uploading the equipment specification file for review.
All I do is upload the document and map it with the checklist I just extracted. Then I click compare. In the backend, RAPID processes each checked item one by one against the document. After a few moments, you can review the results. For each result, you can see the cost and the status of pass or failed, and you can adjust the confidence score. High confidence means AI has sufficient evidence to make its judgment. When you open the detail, you'll see AI's detailed reasoning on the decision as well as citations from the source text for humans to verify the results. This checklist requires rated power and nameplate to match. And they do actually match in the motor specification. If you want a human to validate, you can check out the source documents here.
Here's the nameplate: model ID 75, 75 kilowatt, and the cooling system, 30 watt and 30 kilowatts. So they match. AI's decision was right this time. Let's look at another item. Here you can see an item with a low confidence score. As we've seen in the checklist extraction phase, the checklist requires compliance with the standard requirements for heating equipment, but it didn't have specific details on how to check whether it meets the compliance. So AI wasn't confident enough and returned a failed status. In this case, you could go back and improve the checklist, or if you're a human expert, you can validate and check manually and override the result.
Like this. This is the human-in-the-loop component. Not just throwing everything at AI. But there's another approach we've developed into RAPID: to prepare and configure detailed requirements as an external knowledge source and configure RAPID to make it accessible via tools. It's easier to see in the demo. Here's the same checklist item, but this time it passed with 95% confidence.
When evaluating the checklist, if the rapid assessment didn't have enough information, it went on to search external knowledge to find how to calculate whether it meets the standard, and it returned the calculation method for validation. Then it creates the program. Based on the information it extracted and validated, it's above the requirements. So that's another approach.
But let's see the whole picture here. Now you have the whole picture. In phase one, you use AI to extract the checklist, then human validate it. In phase two, we use those checklists for document review. Each checklist item is processed separately against the relevant documents, then human validated the results. So this two-phase approach solves both of the problems I mentioned earlier. Remember? The first problem was lost in the middle. By limiting each AI code to a single check item with focused document context, we reduce the degradation that happens with long, complex prompts. And the second problem was the accountability gap. Human experts verify the checklist structure in phase one, then validate AI review results in phase two. So there's clear human oversight at every critical decision point.
Building and Deploying RAPID on AWS: Architecture, Impact, and Getting Started
Now, how have we built this on AWS? Here's a rapid serverless AWS architecture. Let me walk you through each component. The user interface is delivered through Amazon CloudFront, a content delivery network service, and static web files are stored in Amazon S3 and AWS WAF with application firewall for security protection. The backend includes Amazon API Gateway for handling all the API requests, and Amazon Cognito for managing all the user authentication. We have AWS Lambda handling all the business logic behind the APIs.
When checklist documents are uploaded to Amazon S3, a workflow orchestrated by AWS Step Functions triggers AWS Lambda to extract and structure the checklist. We use the latest generative AI model through Amazon Bedrock Service. Then all structured checklist data is stored in Amazon Aurora Serverless v2, a MySQL coordinated relational database service that scales seamlessly based on your workload. Our review workflow is also managed by Step Functions, coordinating between Amazon S3 for document storage and Amazon Bedrock Agent runtime integration. This AI agent approach enables intelligent documentation review and analysis, and the AI agent can access external tools or external knowledge via Amazon Bedrock Knowledge Base or run computational analysis through Amazon Bedrock Agent Core Code Interpreter functionality.
The main benefit of running serverless is you only pay for what you use. So you can start with ten documents per week for the proof of concept phase, and you can scale to production with tens of thousands of documents per week. No infrastructure management is required. AWS handles all the patching, scaling, and maintenance behind the scenes. And let me put some business numbers on the business impact. We've seen up to 75% reduction in human hours for document review in some of our early pilot projects. What could this mean? It could mean faster delivery of sustainability-related projects. Now, sustainability regulations can accelerate sustainability without slowing down your business. So this is the kind of transformation we are talking about.
So how can you benefit from RAPID? It depends on what you do. If you are an applicant, you can deploy RAPID internally, measure the time savings, and advocate for regulatory adoption. If you're a reviewer, a government agency, or inspection authority, start with a pilot project on the biggest bottleneck, measure the savings, then scale to other categories. If you are a technology partner like a system integrator or consultant,
RAPID is open sourced. You can customize it to solve your specialized industry's biggest challenges. Let me also show you a roadmap example. By starting with the end goal and walking through how to get there step by step, our ultimate goal is operation for production. RAPID is actively used for production, continuously streamlining document review operations.
Now, how do we get there? The first step is always business validation. You need to identify the domain with the highest business impact and organize a checklist and review documents. Once you validate the business case, you move into experimentation. Experiment with RAPID in a small but high impact domain. This is where you validate business performance and demonstrate that RAPID delivers expected value. You can iterate through business validation until you prove the concept.
Once you prove the concept, you move into production development. This is when you gather the production requirements and develop a full-scale solution. At this phase, you may want to involve AWS partners or professional services to fill in the requirements. And this brings us back to our goal. RAPID is actively used in production, streamlining document review and continuously improving.
Finally, how can you get started on RAPID? You can scan the QR code on the left. We have a GitHub open source repository on the left. And if you want to deploy quickly, we also have a QR code on the right with the AI solution box where you can deploy RAPID in just one click. You have the RAPID environment in just ten minutes. Transform sustainability regulation from constraint to capability. As I stated in the beginning, you now know how you can flip the script. Let's make sustainability regulations accelerate sustainability without slowing down our business.
And if you want to see more RAPID in action, please stop by the sustainability showcase booth in the Expo at the Venetian. We have a bunch of demos, including RAPID, to solve sustainability related challenges. I'm looking forward to seeing you there and ready to show you RAPID live demos and ready to help you get started. That's all. Thank you.
; This article is entirely auto-generated using Amazon Bedrock.
Top comments (0)