Extending public sector intelligence with Agentforce and AWS
Amazon Web Services Christian Ramirez
AWS and Salesforce are wiring agent chat to processed evidence in S3. It lets staff ask for insights on files without leaving the Salesforce console.
Based on reporting by Amazon Web Services, Christian Ramirez — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
AWS is pitching a neat fix for a very old problem: agencies collect mountains of raw evidence, but someone still has to read, sort, and explain it before it’s useful. The new pattern ties Amazon Bedrock Data Automation to Salesforce Agentforce through the Model Context Protocol, so a user can ask for a summary and get one back inside the Salesforce console.
The setup leans on storage first. Files and documents move into Amazon S3 through the External Storage of Files with Amazon S3 connector for Public Sector. From there, an S3 event kicks off an AWS Lambda function, which creates a document ID, stores metadata in DynamoDB, and starts a Bedrock Data Automation job. When processing finishes, EventBridge triggers another Lambda function to save the results into a dedicated output bucket.
That processing step is where the useful stuff appears. Documents get text extraction, key fields, and structured summaries. Images can be described and scanned for objects or text. Video and audio files get transcriptions and scene-level summaries. On the Salesforce side, an Agentforce chat invokes AWS through MCP, passes through Amazon Bedrock AgentCore Gateway, and pulls the relevant results back into the agent’s context for a natural-language reply.
The post is careful to say this is modular, not a one-off. The same pattern can be adapted for permits, benefits claims, or compliance reviews. AWS even points out that you can add custom logic in the Lambda MCP Serverless Runtime, store more metadata in DynamoDB, or connect different agent front ends through MCP without changing the underlying pipeline.
There’s also a healthy amount of practical friction here, which is refreshing. The GitHub sample is not production ready, the Salesforce org has to support registering external MCP servers, and the allowlist should be kept tight in real deployments. That’s the right tone for this kind of integration: useful, but not magic.
My take — AI-written commentary, not fact-checked reporting
This is the right kind of enterprise AI: boring plumbing that actually moves work along. The hype crowd wants a model to “understand” public sector evidence; the grown-up version is storage, event hooks, access control, and a chat window that doesn’t make people jump systems. Very unglamorous. Very likely to ship.
Read more about this at: Amazon Web Services