AI for presentations just got faster: Amazon Bedrock AgentCore can turn quarterly reports into finished PowerPoint decks in under 90 seconds. Here's how it works, and how you can try the setup today.
If you've ever burned a Friday afternoon copy-pasting numbers out of a quarterly report into slide 34 of a client deck, one number from AWS is going to sting a little. A team there took that entire job — reading the reports, pulling the data points that matter, outlining the slides, generating images, and assembling the finished file — and handed it to an AI agent system built on Amazon Bedrock AgentCore. In their own internal testing, the finished PowerPoint came out in under 90 seconds.
The write-up comes from an AWS team working with a financial services group. Those people were spending serious time every single week building client presentations out of quarterly reports, compliance documents, and market analyses. Nothing exotic about it — just the kind of grind that eats a week in slices.
The old way: hundreds of pages per deck
Every deck started the same way. Read hundreds of pages. Find the handful of data points that actually matter. Decide what belongs on which slide. Then hunt for a visual that isn't the same stock chart everyone else is using. Then do the whole thing again for the next client, and again next week.
That's the loop the team wanted gone. Not because the work is hard — because it's repetitive, and repetition is exactly what a well-built agent is good at.
What AWS actually built
The centerpiece is a single "supervisor" agent running on AgentCore runtime. Think of it as a project manager with two specialists on speed dial. The agent's code runs inside its own isolated container on that runtime, with session management, built-in identity, and observability already handled — so you're not hand-rolling your own logging and access control from scratch.
The supervisor exposes exactly two tools:
• generate_pptx — builds the presentation. It pulls content from the knowledge base, drafts the slide outline, creates AI images, and assembles the final document.
• generate_summary — digs through the knowledge base with retrieval-augmented generation and returns an answer grounded in your documents instead of a plausible-sounding guess.
When a request comes in, the agent decides whether it needs to generate a document, summarize something, or both, then calls the right tool. Want a new capability later? Define a new tool. You don't have to touch the ones that already work — and that's the real selling point of the design.
Three models share the workload
All three are available on Amazon Bedrock:
• Anthropic Claude Sonnet 4.6 writes the content.
• Amazon Titan Text Embeddings v2 powers the vector search that finds the right source material. (Embeddings are just a way of turning text into numbers so a computer can measure how similar two passages are.)
• Stability AI Stable Image Inpaint generates the slide images.
The plumbing that makes it safe enough for finance
The team didn't skip the boring parts, and that's why this is more than a demo. A RAG pipeline uses Amazon OpenSearch Serverless to store those vectors and run semantic search across mixed content — text and images together. A REST API built with .NET 8 minimal APIs lets existing enterprise applications call the whole thing over ordinary HTTP requests. AWS Lambda provides the serverless compute, including containerized functions that actually produce the PowerPoint file. And the entire setup sits behind Amazon VPC isolation, AWS KMS encryption, and IAM least-privilege access control.
Translation: it's built to live inside a company that has auditors.
What it means for you
You don't have to work in financial services for this to matter. Strip away the AWS names and the pattern is: pull the source material, draft the structure, generate the visuals, assemble the file. That's the shape of most boring knowledge work you do.
At home
Think about the last time you had to explain something complicated to your family — a health insurance plan, a renovation budget, a school choice. You gathered the facts, then struggled to lay them out in an order that made sense. An agent like this doesn't care whether the source is a compliance document or your own messy notes. Point it at the material and it hands back a structured deck you can edit.
At work
Status updates, quarterly reviews, onboarding guides, vendor comparisons — the "read hundreds of pages, produce ten slides" loop is everywhere. The under-90-second number isn't really about speed. It's about the fact that your first draft stops being a chore and becomes a starting point you can actually think about.
For small businesses and freelancers
If you run an agency, deck-building hours are hours you can't bill at a premium. Turning a client's raw reports into a solid first draft in about a minute and a half means you sell the thinking and the polish, not the formatting. The deliverables stay the same size; the margin doesn't.
For students
Literature reviews and group projects are the same problem in a different jacket: a pile of sources, a required structure, and a deadline. Grounding matters here too — because the system pulls from the documents you feed it rather than whatever the model half-remembers, you can trace every claim back to a page.
For creators
Turning a research dump into a slide carousel, a workshop deck, or a course module is the exact workflow this thing automates. If you'd rather get a feel for AI tools before you go anywhere near an AWS console, mykreatool.com keeps a set of free AI tools you can test today.
As a side income
Here's the less obvious play: sell the setup, not the slides. A bookkeeper, a marketing consultant, or an IT freelancer could stand this pipeline up for small firms that drown in recurring reports, then charge for the integration and the guardrails. The AWS post publishes the code assets specifically so people can do this.
How to try it right now
Step 1: Start with the free option
The architecture runs on the Strands Agents framework, an open source SDK for building AI agents — free to use. The AWS post also links to a public repository containing the code assets, so you can read and copy the whole thing at no cost. Start with the two tool definitions. They contain the entire idea in miniature, and you'll understand the pattern in ten minutes without spending a cent.
Step 2: Open an AWS account and enable Amazon Bedrock
You'll need access to Anthropic Claude Sonnet 4.6, Amazon Titan Text Embeddings v2, and Stability AI Stable Image Inpaint on Amazon Bedrock. Request model access in the console. Heads up: AWS bills by usage, and the source doesn't quote a price, so check current rates before you scale anything.
Step 3: Build the knowledge base
Create an Amazon Bedrock Knowledge Base and point it at the documents you actually work from — reports, PDFs, spreadsheets. That's what turns generic AI output into something grounded in your own material.
Step 4: Deploy the supervisor agent
Deploy the supervisor agent on Amazon Bedrock AgentCore runtime using Strands. This is the piece that decides which tool to call.
Step 5: Wire up the two tools
Implement generate_pptx (via the AWS Lambda function that produces the PowerPoint) and generate_summary. Add new tools later without rewriting what's there.
Step 6: Add the API layer and lock it down
Expose it through a REST API built with .NET 8 minimal APIs so your existing systems can call it over HTTP. Then apply Amazon VPC isolation, AWS KMS encryption, and IAM least-privilege permissions. Don't skip this step if other people's data is involved.
Step 7: Run it against a real deck
Feed it last week's source documents and time the result. AWS measured under 90 seconds internally — your numbers will depend on your documents, your images, and your setup, so measure rather than assume.
Upsides and what changes
• The first draft stops being the bottleneck. Hours of assembly collapse into roughly a minute and a half of machine time in AWS's internal testing.
• Content stays grounded. Because retrieval happens against your documents first, you're editing sourced material instead of fact-checking a model's memory.
• It's modular by design. Two clean tools, a supervisor that routes between them, and room to add more.
• The security work is done for you. Container isolation, identity, encryption, and least-privilege access aren't afterthoughts here.
• No single-vendor straitjacket on the agent layer. AgentCore runtime is built to work with any framework or model, so you can swap pieces as the field moves.
What changes day to day is subtler than "AI writes my slides." It's that the assembly work — the part nobody enjoys and nobody gets credit for — quietly disappears.
Limitations
Be clear-eyed about what this is: a technical blueprint with published code, not a product you can buy and switch on. Getting it running takes real AWS experience — a knowledge base, a runtime deployment, Lambda functions, a .NET API, and a permissions model. Every piece you use costs money, and AWS doesn't publish a fixed price for the stack in this post. The 90-second figure comes from the team's own internal testing, so treat it as a benchmark, not a promise. RAG is only as good as what you feed it: messy or outdated source documents produce confident-looking slides full of the wrong numbers. And generated images still need a human eye before anything goes in front of a client — review the output like you'd review a junior colleague's first pass, because that's basically what it is.
Conclusion and one action for today
The story here isn't that AI can make a PowerPoint. Plenty of tools do that. It's that a team took a genuinely painful weekly workflow — hundreds of pages, compliance documents, market analyses, client-ready slides — and rebuilt it as an agent with two tools, real retrieval, and enterprise-grade security, finishing in under 90 seconds in their own testing. Whether you ever touch Amazon Bedrock or not, that's the shape of the next few years of knowledge work.
Your one action for today: pick the document you rebuild most often, and list every source file that feeds into it. That list is your project brief — and if you want to warm up first, grab a free tool and try the workflow once before you build anything.
Comments 0