Novo Nordisk Taught AI Agents to Answer from Manufacturing Procedures: Minutes Instead of Hours and Days
Novo Nordisk, a maker of diabetes and obesity drugs, has taught an AI assistant to answer questions from factory procedures. Asked "Which samples are taken at the granulator?", the assistant first clarifies the site, product generation and equipment, and only then searches. Expert tasks that once took hours or days now take minutes, with quotes from approved documents. In pharma and other regulated industries, asking a question is safer than confidently giving a wrong answer.
AI-processed from AWS Machine Learning Blog; edited by Hamidun News
Imagine the question: "Which samples are taken at the granulator?" At a tablet manufacturing plant, there is no single correct answer. It depends on the site, the product generation, the specific installation and the batch size. If the assistant pulls an outdated instruction or a document from another plant, in pharmaceuticals that turns into quality problems and audit issues. This is the task the Manufacturing Technology team from Novo Nordisk's tablet business set out to solve. They built an AI assistant on AWS, and according to the authors, work by experts that used to take hours or days has been cut to minutes.
Why ordinary search doesn't work here
Novo Nordisk makes drugs for diabetes and obesity. Every significant production step is described in controlled documents: standard operating procedures, work instructions, forms and training materials. Documents are constantly updated due to new regulatory requirements, equipment replacement, process improvements and auditor comments. What needs revision, what is awaiting approval and what urgently needs attention was largely tracked in spreadsheets and scattered systems. This is slow, error-prone and scales poorly: an outdated document can stay in circulation for a long time.
There is a second problem. Knowledge is spread across versions, sites and programs, and finding the right thing often required an experienced employee who knows where to look. Keyword search returns a list of files, not an answer. Standard vector RAG (semantic search followed by answer generation by a model) handles questions in natural language better. But it gets confused by exact identifiers such as document numbers and equipment codes, and it handles questions about relationships between documents poorly. Most importantly, it cannot honestly say that no suitable material exists, which is critical in a regulated industry.
How Novo Nordisk's GraphRAG works
At its core is GraphRAG: a knowledge graph is added to text search. The graph in Amazon Neptune stores structure and relationships: which process a document belongs to, what it references, and which site, product and equipment it is tied to. Search in Amazon OpenSearch Serverless combines semantic comparison (Amazon Titan embeddings) with exact word matching using the BM25 algorithm. The results of the two methods are merged with reciprocal rank fusion, that is, by a combined ranking. This way, documents matching the meaning of the question and documents matching an exact equipment code are both found.
The two stores are deliberately not merged into a single engine. They are linked by the document number: it is present both in the retrieved fragment and in the graph node. As a result, a found paragraph becomes an entry point into the graph without fuzzy matching. The graph then adds neighboring documents and references, and a reranker model (a cross-encoder in Amazon Bedrock Rerank) reads the question together with each candidate and reorders them by precision.
The agents run on Amazon Bedrock AgentCore, are written with the Strands Agents SDK and use Anthropic Claude models. Documents enter the system through a pipeline built on AWS Step Functions and Lambda, file parsing is handled by Bedrock Data Automation, and the infrastructure is described in AWS CDK.
The graph schema is fixed and was drawn up by experts. The top level is a process tree of roughly three levels, and each document is attached to the lowest suitable node. The bottom level consists of domain entities: sites, product generations, equipment, process steps. The authors explain the choice this way: when a model invents entities on its own, the result is noisy, hard to maintain and hard for auditors to verify. In pharma, the industry already has an expert ontology, and new entities are added with human involvement.
Clarify first, then search
The most notable idea is guided mode. Standard RAG searches first and then answers. Here the order is reversed: if a question is ambiguous, the agent asks a short clarifying question and only then searches.
Universal attributes, such as site and product generation, it asks about directly. Everything else, such as equipment or batch size, is taken from the graph at query time. The agent matches the user's term to real graph nodes, filters out what is already known and, if several options remain, proposes an attribute that actually distinguishes them.
Options are drawn only from existing values, which reduces the risk of invented choices. When experts add a new distinguishing attribute to the taxonomy, the assistant starts asking about it without any changes to prompts or code.
The answer is assembled from approved sources, with links to them, and the assistant states directly when the needed content is absent. Working versions by default are approved and effective ones, not drafts or superseded revisions.
What to take from this
There are caveats. This is AWS describing its own customer, and the claim of "hours and days reduced to minutes" is given without a measurement methodology or independent verification. The available text contains no detailed answer-quality metrics.
The practical lesson is still clear for anyone building enterprise RAG. If your document base includes versions, statuses and areas of applicability, a vector database alone is not enough: you need an explicit relationship schema, as Novo Nordisk has. It is better to ask the user one clarifying question than to deliver a confident answer about the wrong site. And an answer of "no data available" should count as a full result of the system's work.
- Nov 21, 2026A 360-line MCP server with no SDK: five decisions that stopped the agent from breaking links
- Nov 21, 2026Text-to-cad: AI Agent Draws Robot Parts From a Description, Project Passes 26,000 Stars
- Nov 21, 2026A viral GTA 6 Claude clip looks fake: half an hour of control burned through the 5-hour Pro limit
- Nov 21, 2026Claude Code Speeds Up Snowboard Kids 2 Decompilation: From 20% to 45% of Functions in Three Months
Key Questions Answered
Why ordinary search doesn't work here?
Novo Nordisk makes drugs for diabetes and obesity. Every significant production step is described in controlled documents: standard operating procedures, work instructions, forms and training materials. Documents are constantly updated due to new regulatory requirements, equipment replacement, process improvements and auditor comments. What needs revision, what is awaiting approval and what urgently needs attention was largely tracked in spreadsheets and scattered systems.
How Novo Nordisk's GraphRAG works?
At its core is GraphRAG: a knowledge graph is added to text search. The graph in Amazon Neptune stores structure and relationships: which process a document belongs to, what it references, and which site, product and equipment it is tied to. Search in Amazon OpenSearch Serverless combines semantic comparison (Amazon Titan embeddings) with exact word matching using the BM25 algorithm. The results of the two methods are merged with reciprocal rank fusion, that is, by a combined ranking.
What are the key details on Clarify first, then search?
The most notable idea is guided mode. Standard RAG searches first and then answers. Here the order is reversed: if a question is ambiguous, the agent asks a short clarifying question and only then searches. Universal attributes, such as site and product generation, it asks about directly. Everything else, such as equipment or batch size, is taken from the graph at query time.
Ship AI Faster with Founders, Engineers & Weekly Architecture Reviews
Inside Vibe Club (hamidun.com/club) — weekly live masterminds with Zhemal Khamidun, Tech Radar, full LMS access, and 370+ production Claude Code skills.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.