>/BASH

Natera reports faster lab-support responses with SWARM AI system

Natera reports that SWARM cut average first-response time from 17.4 hours to about 8 minutes in a 90-day evaluation. Automated reports reached 96.2% of incoming tickets before engineer review, but the findings do not establish faster repairs or better patient outcomes.

Natera reports faster lab-support responses with SWARM AI system

Natera says its SWARM AI system substantially shortened the wait for an initial equipment-troubleshooting report during a 90-day production evaluation. An AWS Machine Learning blog published on November 12, 2026, reports an average first-response time of approximately 8 minutes, compared with a previous average of 17.4 hours—about 130-fold faster.

The system addresses a knowledge-sharing problem in laboratory operations supporting cell-free DNA testing. Engineers handle more than 65 tickets daily across three Jira queues, drawing on expertise that can be difficult to retain when experienced staff leave. SWARM retrieves vendor documentation, historical tickets and operational information to help engineers investigate equipment faults.

Built on Amazon Bedrock AgentCore, SWARM combines 33 skills: 24 specific to instruments and nine covering broader workflows. Python handles predictable tasks such as parsing ticket fields, while language models interpret symptoms and generate diagnostic narratives. One agent automatically prepares a likely cause, ranked troubleshooting steps and a confidence label; a second supports interactive follow-up with engineers. The reports offer equipment troubleshooting, not patient diagnoses.

Natera evaluated 524 support tickets, using Jira metadata to measure response time and coverage. It reports that 96.2% of incoming tickets received an automated diagnostic report before engineer review. That figure measures response coverage, not successful repairs or autonomous resolution. The source does not explicitly establish whether the coverage denominator was the same 524 tickets.

Accuracy assessment used only 269 tickets with both a confidence label and an engineer-verified resolution. Reported accuracy was 74% for High-confidence reports across 154 tickets, 65% for Moderate across 78, and 65% for Low across 37. Moderate and Low therefore showed no separation in reported accuracy, and the Low group was small. The other 255 tickets lacked an eligible label or verified resolution, potentially biasing the accuracy estimates.

Engineers recorded findings and fixes, and an evaluator graded responses as Accurate, Partial or Wrong. The evaluator’s identity, blinding, grading criteria and treatment of Partial responses in the accuracy figures were not specified. No uncertainty intervals were reported, and independence between tickets involving shared instruments, incidents or engineers was not established.

The team also reports that adding missing vendor manuals raised accuracy for its lowest-scoring instrument family from 71% to 75%, a four-percentage-point improvement without changing the model. Sample size and evaluation details for that comparison were not supplied.

These are company-reported findings in a vendor blog, with no evidence of peer review or independent replication in the supplied material. There was no reported randomized or concurrent control, and the historical comparison and ticket mix were insufficiently described to establish causation. Evaluations were rerun after updates, but an independent held-out test set and safeguards against overlap with retrieved historical tickets were not described. Faster initial reports therefore do not establish equivalent diagnostic quality, shorter instrument downtime, faster repairs or improved patient reporting.

Cover illustration generated with AI; conceptual, not research evidence.

Original sources

View this story on Instagram