From hundreds of pages to a first draft

AWS has released a reference architecture for generating PowerPoint presentations from enterprise material with Amazon Bedrock AgentCore. Published on 14 October, the example is framed around a financial-services team that must repeatedly read quarterly reports, compliance documents and market analysis before preparing client decks. AWS says its internal tests produced a presentation in under 90 seconds, although organisations should benchmark the system with their own document volumes and quality requirements.

The design does more than ask a model for slide text. It retrieves approved source material, creates an outline, generates supporting visuals and assembles a PowerPoint file. That matters because a presentation is a structured business artefact: claims need evidence, slides need a coherent sequence, and the final file must remain editable and reviewable by the person who will present it.

A supervisor selects specialised tools

A supervisor agent runs in AgentCore Runtime using the Strands Agents framework. It exposes one tool for generating a PPTX file and another for summarising material from the knowledge base. The generation tool invokes a containerised Lambda function that retrieves content, develops the slide outline, creates images and assembles the document. Keeping those operations behind defined tools gives the orchestrator a narrower surface than an unrestricted agent.

Conversation state is stored in DynamoDB and requests arrive through API Gateway to a .NET 8 minimal API running on Lambda. Generated files are placed in Amazon S3 and returned through presigned download links. This separation helps teams inspect each boundary: the user request, the retrieved evidence, the generated plan and the file that is ultimately delivered.

Grounding and multimodal retrieval

Enterprise documents are uploaded to S3, parsed into chunks and embedded with Amazon Titan Text Embeddings v2. OpenSearch Serverless stores the vectors for semantic retrieval. The pipeline includes special handling for tables, charts and images, which is important because many presentation-worthy facts are not contained in ordinary paragraphs. Retrieved chunks become context for Claude Sonnet 4.6 on Bedrock.

Stability AI Stable Image Inpaint can create slide imagery, but generated visuals require the same scrutiny as generated prose. A plausible chart or illustrative image can imply evidence that does not exist. Teams should distinguish decorative material from factual graphics, preserve labels for generated assets and require numerical charts to be derived from verified data rather than model imagination.

Security is part of the sample

The architecture combines VPC isolation, KMS encryption and least-privilege IAM. AWS WAF applies managed rules and rate limits at the API, Cognito provides authentication, and Bedrock Guardrails applies content filtering. S3 blocks public access and keeps versioning and access logs, while DynamoDB uses point-in-time recovery, deletion protection and customer-managed encryption. X-Ray traces Lambda activity for diagnosis.

Tenant IDs and session IDs are validated, and the design uses per-tenant role assumption through AWS STS. Those controls are essential when a single service processes confidential reports for multiple business units or customers. They do not remove the need to classify source documents and verify that retrieval cannot cross a tenant boundary.

A presentation still needs an accountable editor

The under-90-second result is best viewed as a first-draft benchmark. A reviewer must check whether every important assertion is supported, whether compliance language survived summarisation and whether the deck omits contrary evidence. Layout quality also needs inspection: crowded slides, unreadable charts and repetitive visual rhythm can make an accurate deck ineffective.

Teams evaluating the reference system should measure accepted slides, citation accuracy, corrections per deck and time saved after review rather than raw generation speed. The architecture is valuable because it makes the workflow reproducible and auditable. Its business value will come from reducing routine assembly while keeping a knowledgeable employee responsible for the story and the final client-facing claims.

Deployment choices that affect reliability

The sample’s infrastructure requirements are substantial: .NET, Python, Node.js, container tooling, CDK, a VPC and access to several Bedrock models. That is appropriate for an enterprise reference architecture, but smaller teams should identify which components solve a real requirement. A simpler workflow may be safer if it has fewer credentials, stores and failure paths, even when it produces a deck more slowly.

API Gateway’s integration timeout must also be raised beyond its default because presentation generation can exceed ordinary request durations. A production client should avoid appearing frozen during that interval. Job identifiers, progress states and retry-safe requests can prevent a user from submitting the same deck several times. The system should make failure explicit and preserve enough state to resume or clean up partial artefacts.

Model and retrieval changes need versioned tests. A new embedding model may select different passages, while a generation-model update may alter structure or tone. Keeping a representative deck set with expected facts, slide counts and compliance language gives teams a regression suite. That discipline protects the workflow from becoming less reliable as its individual managed services evolve.