Cutting deviation cycle time with guided AI.
How a global pharmaceutical manufacturer built a suite of AI assistants to accelerate deviation investigations — collecting evidence, guiding writing, and standardizing quality documents.
Three assistants, one governed documentation workflow.
Thousands of deviations, no standardization.
Potential deviations from established procedures are captured, investigated, and documented as part of the quality process.
To meet resolution targets at scale, the following bottlenecks needed to be addressed: manual fact-gathering and inconsistent technical writing across a large, distributed pool of quality management investigators.
AI mapped to every step of the workflow.
We delivered a phased suite of AI assistants aligned to the deviation management workflow — Initiate, Collect Facts, and Populate — using retrieval-augmented generation over historical events plus structured SOP and attachment data.
Everything runs on the client's self-hosted LLM Mesh, so privileged deviation data never leaves their environment — no anonymization or redaction required.
A phased suite built on a shared RAG foundation.
Three assistants shipped in sequence, each one drawing on the same retrieval layer.
Investigator Assistant
RAG search across historical deviations surfaces relevant precedents, SOPs, and information gaps during fact collection.
Initiation Chat
An interactive chat guides operators through initiating a report and capturing observations consistently.
Guided Writing
Suggestions help populate each report section, guiding rather than generating, to stay compliant.
Vector + Relational
Milvus recalls semantically similar events; a relational store handles structured SOP lookups.
Local & Governed
Runs entirely inside the client environment; no external integrations, no data leaving the boundary.
Unified Workflow
A single guided three-step interface tracks state across Initiate → Collect Evidence → Populate Report.
Feedback & Logging
Thumbs-up/down capture plus token-usage and session-time logging quantify cycle-time gains.
Model-Agnostic
Configurable across GPT and Claude models, tuned for the best balance of accuracy, cost, and latency.
Faster, more consistent investigations, by design.
Complete at intake
Interactive chat intake guides operators at submission, ensuring critical information is captured up front.
Faster fact collection
Automated retrieval of precedents, SOPs, and attachments replaces slow, manual searching.
Consistent writing
Field-level guidance standardizes report quality across distributed investigation teams.
Compliance-safe
Guides rather than auto-writes, and keeps all data on the client's LLM Mesh.
Measurable
Token and session logging quantify time savings against the five-day cycle target.
The full stack behind the suite.
App & UI
AI & Retrieval
Data
Methodology
Trustworthy AI,
inside your walls.
Jahnel Group builds trustworthy AI systems for regulated, quality-critical industries.