Natera says its lab-support AI delivers first responses 130 times faster

Natera’s laboratory engineers now receive AI-generated equipment diagnostics within minutes, rather than waiting through a first-response process that averaged 17.4 hours. In a November 12 AWS case study, Natera says its SWARM system delivers about 130-times-faster first responses and 96.2% automated ticket coverage. The system prepares troubleshooting reports; engineers still investigate and fix equipment.

The deployment addresses a knowledge problem as much as a queueing problem. Natera’s service engineers maintain multiple families of precision instruments, each with different failure patterns and vendor documentation. When a senior technician left, years of equipment expertise went with them. Instrument downtime was delaying sample processing and patient reporting.

A diagnostic head start, not an automated repair

Before SWARM, the team faced more than 65 support tickets daily across three Jira queues. Natera built the system with AWS using Amazon Bedrock AgentCore, a managed service for running AI agents. The team wanted autonomous ticket processing, connections to Jira and Confluence, and built-in security and monitoring without maintaining its own agent infrastructure.

SWARM splits the work between two agents. The background swarm-agent selects skills using the equipment type and reported symptoms, searches the knowledge base, and generates a report. That report gives engineers a leading explanation, ranked troubleshooting steps and a confidence label, with a link added to the Jira ticket within minutes.

Engineers use the second agent, swarm-hive, for follow-up questions, image uploads and alternative troubleshooting paths. It retains conversational context for each engineer across sessions. The reported 96.2% figure describes automated ticket coverage, not diagnostic accuracy or a share of instruments repaired without people. The performance figures are company-reported results, not an independent evaluation.

Better manuals before a different model

Early prototypes produced inconsistent results when they searched documents without equipment-specific metadata. Natera responded with a custom document-processing pipeline for technical manuals. The knowledge base brings together vendor documentation, historical tickets and operational data, giving the agents equipment-specific material to consult rather than relying only on a model’s general knowledge.

That choice produced a measurable improvement for one equipment family. Natera lab engineering leader Ariel Jirau says adding missing vendor manuals raised diagnostic accuracy from 71% to 75%, without changing the model. The team now checks documentation coverage before tuning the model when accuracy stalls, because missing manuals have been its most common source of missed diagnoses.

The software is also divided into 33 versioned skills: 24 specific to instruments and nine shared across workflows. Python handles fixed tasks such as parsing ticket fields and formatting reports. Language-model skills interpret symptoms and generate diagnostic explanations. Each skill has a defined interface and can be tested independently, helping engineers isolate a faulty component.

Checking the diagnosis and separating patient data

Evaluation continues after deployment. Engineers record what they actually found and fixed, and an evaluator grades the agent’s response as Accurate, Partial or Wrong, grouped by instrument family. Logs and traces follow each run from ticket intake to report, recording time and cost per tool call so the team can locate slow or expensive skills.

The clinical setting also shapes what the agents can see. According to the case study, tickets contain equipment and workflow metadata only; patient-identifiable information reaches neither the searchable vector index nor the language model. SWARM runs on HIPAA-eligible AWS services under a Business Associate Agreement, and every agent invocation is audit-logged. Deployment credentials are temporary rather than long-lived keys stored in the pipeline.