Scaling Agentic AI for Digital Accessibility and Content Intelligence

siteimprove · Tech · 2025 · source Agentic: yes · promotional bias: med · source-backed · extracted with prompt 0.5.6 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Digital accessibility and content intelligence at enterprise scale

Siteimprove needed to scale AI and agentic workflows for accessibility, analytics, SEO, content remediation, and image analysis while preserving trust, speed, adoption, compliance, multi-region support, and cost efficiency across very large page volumes.

Constraints:

• Some customers have tens of thousands of webpages that must be crawled and checked.
• Accessibility rule checks can reach up to 100 million pages per month.
• Enterprise customers require security, governance, compliance, resilience, and multi-region support.
• The platform needs both batch and interactive workloads, plus access to multiple model families.
• The adoption path must move through trust-building stages before full autonomy.

Architecture

Siteimprove built an AI accelerator architecture around AWS services. It includes a batch manager for asynchronous page-scale workloads, a business logic manager for prompts and domain parameters, AI service adapters for Bedrock, Converse API, and third-party models, and use-case-specific paths for batch accessibility review, conversational remediation, contextual image analysis, and AgentCore-based conversational analytics. Bedrock and Nova models provide batch/interactive serving, multi-region deployment, and cost optimization; AgentCore connects agents, tools, MCP servers, and short/long-term memory.

Diagram confidence: explicit (topology described by the source).

flowchart LR
  A["Accessibility and content requests"] --> B["Siteimprove AI accelerator"]
  B --> C["Batch manager"]
  B --> D["Business logic manager"]
  B --> E["AI service adapters"]
  C --> F["Bedrock batch processing"]
  D --> E
  E --> F
  E --> G["Conversational Bedrock agents"]
  E --> H["AgentCore runtime"]
  F --> A
  G --> A
  H --> I["AgentCore Gateway"]
  H --> J["Short-term memory"]
  H --> K["Long-term memory"]
  I --> H

Components

Data flow

For batch accessibility review, requests enter through an ALB, Lambda enqueues them, another Lambda groups them into S3 batches, a submitter sends them to Bedrock, and result processors clean and return outputs to a queue. For conversational remediation and analytics, interactive Lambdas call Bedrock Agents or AgentCore. AgentCore agents interpret queries, call local tools, call other agents or an MCP server through the gateway, and use short-term and long-term memory.

Why this architecture

Siteimprove chose Bedrock/Nova and a shared accelerator because the product needed multi-region enterprise deployment, batch and interactive support, multiple model families, deep AWS integration, and flexible pricing. The accelerator separates business logic from service adapters and reuseable batch/agent infrastructure so each use case can pick the right model and execution mode.

Models used: Amazon Bedrock · Amazon Nova · Nova Micro · Bedrock Agents · Bedrock AgentCore · Claude · Llama

Scale

• Some customers have 50,000 webpages to crawl and check.
  “because some of our customers have 50,000 webpages that needs to be crowded, needs to be checked”transcript, stated
• Accessibility rule checks can run up to 100 million pages per month.
  “we need to process millions of pages, up to 100 million pages per month for accessibility rule check.”transcript, stated
• The batch path supports up to 100 batches of 100,000 requests and 10 million requests per day.
  “you can submit up to 100 batches, each batch 100,000 requests, and in one day, you can achieve 10 million requests done.”transcript, stated
• At 99% correctness and 100 million monthly requests, 1 million AI requests can still error.
  “my system is working 99% correct and perfect," but then we are scaling to 100 million or more requests per month, that translates into 1 million requests submitted to AI getting error”transcript, stated
• Nova Micro produced a 75% cost reduction for some Siteimprove use cases.
  “Nova Micro at that time when we checked, we had 75 cost reduction compared to the leading models at that time, for some of the use cases that we have.”transcript, stated

Alternatives considered (roads not taken)

Big-bang autonomy

Rejected because: Siteimprove frames AI adoption as a trust curve that must prove value before moving from human-in-loop to autonomous operation.

Always choose the strongest model

Rejected because: Cost and task fit matter at scale; Nova Micro delivered large savings for some use cases.

Interactive-only processing

Rejected because: Accessibility review requires asynchronous batch architecture for millions of pages per day.

Patterns

batch-offline-processing (serving)

Accessibility checks are grouped into queues, S3 batches, Bedrock batch submissions, and post-processing queues for millions of page requests.

multi-agent-orchestration (orchestration)

AgentCore connects agents, local tools, other agents, an MCP server, and memory to answer cross-domain analytics questions.

tool-use-orchestration (orchestration)

AgentCore agents use local tools for query interpretation, formatting, JSON/UX formatting, MCP access, and Bedrock-backed subtasks.

human-in-the-loop-gate (governance)

Siteimprove describes a staged adoption path from human-in-loop to human-on-loop to human-out-of-loop autonomy.

model-routing (model-selection)

The AI accelerator can use multiple model families and choose lower-cost Nova models where the task does not need the strongest model.

agent-memory (memory)

AgentCore gives agents short-term memory for session management and long-term memory for actor preferences.

Tradeoffs

Processing mode

Chose asynchronous batch for page-scale accessibility checks over interactive request-by-request processing.

Rationale: Batching makes millions of page checks feasible with acceptable cost and a 24-hour turnaround.

Cloud/model platform

Chose Amazon Bedrock and Nova through Siteimprove's AI accelerator over single-model or non-integrated AI service choices.

Rationale: Bedrock gave multi-region support, batch and interactive modes, agent/generative support, deep integration, and flexible pricing.

Autonomy progression

Chose human-in-loop to human-on-loop to human-out-of-loop over immediate autonomous operation.

Rationale: Trust and value must be proven before reducing human oversight.

Gotchas (what bit them)

Accessibility remediation at web scale is an orchestration problem, not only an AI prompt problem.

Root cause: Customers can have tens of thousands of pages with timeline constraints.

Resolution: Use asynchronous batch processing with queues, S3 batches, Bedrock submission, and result post-processing.

Small error rates become huge at AI-scale volume.

Root cause: A 99% correct system still produces many failures at 100 million monthly requests.

Resolution: Scale planning must account for error volume, trust, and staged adoption before autonomy.

Adoption stalls if trust is skipped.

Root cause: Users must trust the system and see value before moving to less supervised operation.

Resolution: Progress through human-in-loop, human-on-loop, and human-out-of-loop stages.

Using the strongest model for tiny tasks wastes money.

Root cause: At scale, model choice directly drives cost, and some workloads can use smaller models.

Resolution: Use model-routing discipline and Bedrock/Nova options to select task-appropriate models.

Evaluation

Siteimprove evaluates production readiness through scale, cost, trust/adoption, model fit, and operational business requirements rather than a single offline metric. The talk emphasizes requests per day/month, cost reduction, multi-region compliance, and staged trust.

Metrics: Up to 100 million pages per month · Up to 10 million requests per day through 100 batches of 100,000 · 24-hour batch turnaround · 75% cost reduction for some Nova Micro workloads · 99% correctness still means 1 million errors at 100 million monthly requests

Outcomes

• Siteimprove implemented an AI accelerator architecture shared across batch, interactive, image-analysis, and agentic use cases.
• The batch architecture supports up to 10 million requests per day and 100 million pages per month.
• Nova Micro delivered 75% cost reduction for some workloads compared with leading models.
• AgentCore-based conversational analytics connects agents, tools, MCP, and memory.
• The company formalized a human-in-loop to autonomous trust curve for adoption.

The case reports 50,000-page customers, up to 100 million monthly page checks, up to 10 million daily batch requests, 24-hour batch turnaround, and 75% cost reduction for some Nova Micro workloads.