Field noteNov 17, 20266 min read1 source

Novo Nordisk modernizes manufacturing process documents with Amazon Bedrock AgentCore

In one minute

Novo Nordisk’s Manufacturing Technology team built a process-intelligence assistant with AWS for questions across a large manufacturing-document vault. Its design combines a curated knowledge graph, hybrid retrieval and clarification before search to produce cited answers grounded in applicable documents. For builders, the case shows how to make scope, document status and relationships explicit in a regulated workflow.

What changed: Novo Nordisk applies Bedrock AgentCore to build an assistant for manufacturing process documents, a distinct customer deployment beyond the existing document-processing tutorial.

What happened

AWS’s announcement describes a system developed by Novo Nordisk’s Manufacturing Technology team within the Tablet area, in collaboration with AWS. The team’s challenge was not simply locating a document: an answer might depend on the manufacturing site, product generation, equipment, document version and lifecycle state, as well as links between documents. The wider vault holds on the order of four million documents, while full-text contents are indexed incrementally for onboarded domains.

The assistant uses a GraphRAG architecture with two separate stores. Amazon Neptune holds the knowledge graph: an organization-level process hierarchy of roughly three levels, plus domain entities such as sites, product generations, equipment and process steps. Amazon OpenSearch Serverless indexes document content. Search combines dense vector similarity using Amazon Titan Text Embeddings with lexical BM25 matching, fused through reciprocal rank fusion. A reranking stage reorders candidate passages before the model uses them. Shared document numbers bridge search results to graph nodes, allowing the system to expand from a retrieved passage to related documents and relationships.

A query agent built with the Strands Agents SDK runs on Amazon Bedrock AgentCore and uses Anthropic Claude models on Amazon Bedrock. The ingestion pipeline uses Amazon Bedrock Data Automation for multimodal parsing, orchestrated with AWS Step Functions and AWS Lambda; the infrastructure is defined with AWS CDK. In guided mode, the agent checks whether the question’s scope is clear before retrieving. It asks about common attributes such as site or product generation when needed, then resolves deeper distinctions, such as equipment or batch size, against graph values. If multiple graph-backed candidates remain, it offers the attribute that distinguishes them rather than inventing choices.

AWS says answers default to approved, effective documents and the latest applicable version, unless the user explicitly requests drafts or historical revisions. A deterministic indexed-state flag distinguishes documents whose metadata is present from those whose contents are searchable. The assistant is designed to cite approved sources, acknowledge missing or ambiguous content, and flag cases requiring subject-matter expert, quality or regulatory review. An evaluation harness checks source relevance, document status, version applicability, scope compliance, citation accuracy and preservation of requirement strength—whether a source says must, should or may. AWS reports that some expert tasks that once took hours or days now take minutes.

What the archive adds

New Runtime’s earlier coverage of semantic layers for agentic AI on AWS (October 22, 2026) argued that shared definitions and permissions can govern agent access to enterprise analytics. This is a new event and a distinct manufacturing case: here, graph structure and document lifecycle state govern retrieval and answer scope. An earlier note on another AgentCore engineering-intelligence application (November 3, 2026) makes this a further example of AgentCore being applied to domain-specific information work.

Trend signal

This case supports agent security moves to runtime boundaries: the design relies on managed session isolation, IAM integration, observability and explicit document-state and scope checks, rather than trusting a prompt alone. It also reinforces company memory needs write loops, though less directly: domain experts curate the taxonomy and can promote new entities through a human-in-the-loop process. The announcement describes a built system, but does not establish independent performance or deployment-wide outcomes.

Why now

The team’s documentation burden is shaped by recurring regulatory changes, process improvements, equipment modifications and audit findings. AWS frames the assistant as a response to the difficulty of keeping thousands of active controlled documents current and finding the applicable answer across a vault of roughly four million documents. The timing is therefore tied to a persistent operational problem, not a newly announced general release.

Connections to other events

New Runtime’s earlier coverage of Amazon Bedrock’s single-shot and agentic retrieval modes (November 10, 2026) concerned retrieval choices in the same product ecosystem. The Novo Nordisk system is a separate application: it combines hybrid search with graph expansion and asks for clarification before retrieval, rather than describing a general retrieval-mode release.

Why it matters

In a regulated workflow, retrieving a plausible passage is not enough. A team needs to establish whether that document is current, approved and applicable to the site, product and equipment in the question. Novo Nordisk’s design makes those conditions part of the system: a curated taxonomy expresses domain structure, search finds candidate passages, graph traversal supplies relationships, and guided clarification narrows scope before the answer is composed.

The transferable lesson is architectural, not that every team should adopt the same AWS stack. Separate the searchable text from the structured model of applicability when each has different requirements; connect them with stable identifiers; and evaluate whether answers cite the right version and preserve the force of requirements. For teams working with high-consequence records, explicit “not indexed,” ambiguous and out-of-scope states are useful product behaviors, not merely retrieval failures to conceal.

Limits and uncertainty

The outcomes are reported by AWS in a vendor-authored customer case, not independently measured here. The announcement says some tasks fell from hours or days to minutes, but does not give a task-by-task baseline, sample size, time period or error rate. It does not establish how often the assistant’s answers are accepted without expert review, how often clarification is required, or how quality changes across document domains and languages. The account describes architecture and intended controls, but does not provide an independent audit of compliance impact or quantify reductions in outdated documentation or manufacturing delays.

Skill to build now

For a document assistant this week, choose one question whose correct answer changes by site, product version or equipment. Write down those scope attributes and the approved-document states that matter; then require the workflow to resolve missing scope before retrieval. If you have a graph or structured metadata, use its real values to generate clarification choices. Add a small test set that checks whether answers cite the latest applicable approved document and preserve words such as “must” versus “should.”

Reader takeaway

The useful pattern is not “put four million documents behind an agent.” It is to model where documents apply and how they relate, connect that structure to retrieval with stable identifiers, and clarify scope before answering. In regulated work, a cited answer that can say what it does not know is more valuable than a fluent answer that silently guesses.

Evidence / sources

  1. 1aws.amazon.com/blogs/machine-learning/novo-nordisk-modernizes-manufacturing-process-documents-with-amazon-bedrock-agentcore
All Field Notes

Next reads

Read next

Open the graph
  1. 01topicAnthropicExplore the anthropic topic hub.
  2. 02topicClaudeExplore the claude topic hub.
  3. 03archiveField NotesOpen the latest editorial analysis.
  4. 04sourcesSourcesSee the original sources this archive cites.
Who read this page?Machine requests, hidden until opened

Loading the privacy-safe route aggregate...

Open the JSON contract