
Institutional Knowledge AI: What Natera's Lab Agent Teaches Operations Teams

By Erin Harris, AI Strategy & Business Transformation Consultant
The business case for an institutional knowledge AI system is not that it can answer questions. It is that it can help authorized staff find the right operational guidance, trace where that guidance came from, and record whether it worked.
AWS and Natera published a November 12 case study of SWARM, an agent system supporting lab instrument operations. The company reports average first response dropping from 17.4 hours to about eight minutes, with automated first-response coverage of 96.2 percent. The evaluation also distinguishes tickets with engineer-verified outcomes from the larger ticket set. Source: https://aws.amazon.com/blogs/machine-learning/how-natera-designed-an-agentic-ai-system-for-lab-ops-with-amazon-bedrock-agentcore/
Why this matters beyond laboratories
A small HVAC company, logistics operation, or service organization may have the same structural problem on a different scale: exceptions are solved by experienced staff, but the reasoning is rarely captured in reusable form. When those people are absent, the business relies on memory, calls, and guesswork.
An agent can surface documented guidance. It cannot manufacture missing institutional knowledge. Natera's design highlights the importance of equipment-specific documentation, structured skills, and engineer feedback. Its own report describes improving accuracy by adding missing vendor manuals rather than changing the model.
A practical knowledge-continuity workflow
First, select one recurring problem family with a measurable baseline: number of escalations, first-response time, and repeat incidents. Do not start with every department.
Second, assemble approved source material. Include current procedures, vendor instructions, resolved cases, and explicit exceptions. Assign an owner to each document and identify its review date.
Third, separate deterministic rules from judgment. A ticket number, equipment type, or required field can be validated by software. A proposed diagnosis needs supporting evidence and may require human review.
Fourth, design a response format that names the suspected issue, cites the underlying source, lists safe next steps, and states when to escalate. A confident paragraph without a traceable source is not an adequate operational record.
Fifth, compare the agent's suggestions with verified outcomes. Count correct, partially useful, and incorrect responses. Track the denominator, not only the attractive examples. Revisit the knowledge base when failures cluster around missing documentation.
Worked hypothetical
A regional equipment maintenance firm receives 40 service tickets each day. One senior specialist knows the unusual failures of an older machine line. Rather than deploying an unrestricted troubleshooting agent, the firm documents 20 resolved cases, adds the current service manual, and tests retrieval against 30 historical tickets. It limits the assistant to drafting a technician briefing. The technician signs off on any repair action. These numbers are an illustrative pilot design, not measured results.
Decision criteria
Proceed only if the source set is sufficiently complete, permissions are clear, staff can challenge the recommendations, and an owner will maintain the material. Pause if the agent is expected to issue safety-critical instructions without verification or if the documentation is obsolete.
The practical objective is continuity, not an impressive demo. EH Synergy's AI Opportunity & Transformation Assessment can help identify where expert knowledge is trapped, what needs to be documented, and whether an AI implementation is justified. Inquire through EH Synergy's existing contact channel.
Key Takeaways
- Begin with knowledge capture and source ownership.
- Evaluate verified outcomes separately from response coverage.
- Preserve human approval for consequential actions.
- Build a maintenance process for documentation and skills.

