AWS re:Invent 2025 - Scaling Accessibility w/ Agentic AI: Siteimprove + Amazon Bedrock & Nova-NTA202

Scaling Accessibility with Agentic AI: Siteimprove + Amazon Bedrock & Nova-NTA202

Agentic AI: Transforming Business Beyond Automation

• Agentic AI systems go beyond traditional automation, empowering businesses to innovate faster, respond in real-time, and stay ahead of changing market dynamics.
• Agentic AI systems are designed with intent, planning, reasoning, and taking autonomous initiatives to drive outcomes, rather than just executing predefined rules.
• Agentic AI adoption is accelerating rapidly, with Gartner predicting that 40% of enterprise applications will embed task-specific agentic AI within 1 year, and agentic AI-driven software expected to capture 35% of enterprise application revenue by 2035.

Prioritizing AI Investment for Maximum Impact

• Focus on business outcomes and metrics, not just technology trends, when prioritizing AI initiatives.
• Use a structured framework to evaluate opportunities based on effort, value potential, trust, speed, and adoption.
• Invest in the right foundation to unlock transformative growth, balancing quick wins and strategic bets.

Siteimprove's Journey from Generative to Agentic AI

• Siteimprove is a SaaS platform that helps organizations ensure digital accessibility, compliance, and performance.
• Their journey started with generative AI for specific tasks, then progressed to AI agents capable of making suggestions and taking actions.
• The ultimate goal is to achieve agentic AI, where multiple agents communicate and work together autonomously to solve problems.

Key Agentic AI Use Cases at Siteimprove

1. Async Batch Processing for Accessibility Checks:

   • Processes millions of web pages per month for accessibility rule checks, using an architecture with input/output queues, S3 storage, and Bedrock batch processing.
   • Enables cost-effective, scalable, and automated accessibility testing at a massive scale.

2. Conversational AI Remediation for Accessibility Issues:

   • Provides an interactive experience where users can report accessibility issues and have the AI agent suggest fixes and have a conversational dialogue to resolve them.
   • Leverages Bedrock agents and the Siteimprove AI Accelerator architecture.

3. Contextual Image Analysis:

   • Analyzes images on web pages to understand the context, such as alt text, captions, and headings, and provide recommendations.
   • Uses an async processing model with Bedrock for the image analysis.

Lessons Learned and Best Practices

• Use cross-region inference to improve latency, resilience, and cost optimization.
• Invest in effective prompt engineering and optimization for different models.
• Understand regional throughput and quota variations to avoid production issues.
• Mitigate failed responses by using different models or model families.
• Enforce strict output formatting to ensure consistent and usable AI-generated content.
• Handle contextual information effectively to avoid issues like model hallucination.

Amazon Bedrock and Agent Core: Powering Agentic AI at Scale

• Amazon Bedrock provides access to leading AI models and enables the development of agentic AI systems.
• Amazon Agent Core is a fully managed service that handles the complexities of running production-grade agentic AI applications, including orchestration, memory management, security, and observability.
• Agent Core integrates with other AWS services and supports open standards, allowing enterprises to build and scale their agentic AI solutions.

Key Takeaways

1. Prioritize AI investments based on business outcomes, not just technology trends.
2. Siteimprove's journey demonstrates the progression from generative to agentic AI, unlocking new capabilities and business value.
3. Agentic AI use cases at Siteimprove showcase the power of autonomous, multi-agent systems to solve complex problems at scale.
4. Lessons learned highlight the importance of technical best practices and the role of AWS services like Bedrock and Agent Core in enabling enterprise-grade agentic AI.