AWS shows how public agencies can turn body-camera footage into searchable AI evidence

AWS has published an architecture for helping Salesforce Agentforce search body-camera footage, surveillance video, photographs, audio and scanned documents through AI. The practical point is not another chatbot perched on top of a filing cabinet: the design turns messy evidence into structured results that an agent can retrieve and explain inside the Salesforce console.

AWS AI Watch analysis

What happened

AWS says the pattern uses Amazon Bedrock Data Automation to process media stored in Amazon S3, with Lambda, DynamoDB, EventBridge and Bedrock AgentCore Gateway connecting the processing pipeline to Salesforce Agentforce through the Model Context Protocol. Documents can produce extracted text, key fields and summaries; images can produce descriptions and detected objects or text; video and audio can produce transcriptions and scene-level summaries.

The proposed workflow processes files asynchronously after upload, stores the resulting insights, and lets authorised Agentforce users ask questions without switching between systems. AWS says the design can be adapted for evidence management, while warning that its accompanying CDK sample is not production-ready and should be reviewed and hardened before deployment.

Why it matters

This is a concrete example of AI moving into the less glamorous, more consequential work of public-sector information handling. A person investigating a case could search processed media through an existing console instead of manually reviewing every file or juggling several storage and analysis tools. That could save time, but it also puts a great deal of weight on extraction quality, access controls and the ability to distinguish a useful lead from a confident-looking mistake.

The architecture keeps the underlying material in AWS storage and exposes selected results through an authenticated MCP connection. That gives organisations a route to connect agents to evidence without making raw files the model’s permanent playground. It does not, by itself, prove that the system is accurate, fair or suitable for live investigations.

Our read

AWS is offering a pattern, not evidence that public agencies should hand investigative judgement to an agent. The useful development is the plumbing: multimodal processing, structured outputs and source-system access are being assembled into one workflow that people can actually query.

The warning label matters. AWS says the sample is not production-ready, and the supplied post does not establish error rates, independent evaluation, retention practice or how conflicting evidence is presented. Agencies considering this route should test it against difficult real-world material, require human review for consequential decisions and make the provenance of every generated insight visible. Otherwise the filing cabinet has merely learned to speak with confidence.

What to watch

• Accuracy and provenance:
  whether agencies publish testing on transcription, scene summaries and extracted fields.
• Access controls:
  how narrowly MCP tools and evidence results can be scoped to authorised users.
• Human review:
  whether generated insights remain leads for investigators rather than unexamined findings.
• Production readiness:
  whether AWS or Salesforce publishes stronger operational, security and retention guidance.

Discussion spark: Should public agencies use AI to search and summarise investigative evidence if every result remains subject to human review, or is the risk of a plausible but wrong lead too high for that workflow?

not affiliated with or endorsed by Amazon Web Services (AWS)