AI agent·In production at Bayer · expert review before scores are used
ARS (Assessment of Regulatory Success) Automation
A production AI agent at Bayer. It reads a safety data sheet and drafts the toxicology sections of country ARS workbooks.
| Before | Now | |
|---|---|---|
| Tox / ecotox | Typed from the safety data sheet | Drafted for review |
| Share of the job | The slow part of every ingredient | About 70% pre-filled |
| Who decides | Expert starts from a blank sheet | Expert checks before scores are used |
Problem
Country ARS (Assessment of Regulatory Success) portfolios live in multi-sheet Excel workbooks. Experts read an MSDS, map the hazard lines, then apply that country's scoring rules. It is slow, it is the same job for every ingredient, and it holds up decisions on which active ingredients are worth registering.
Value
Toxicology and ecotoxicology look like about 70% of the work per active ingredient. If those cells are pre-filled, the expert reviews instead of typing, and regulatory managers get a clearer view for resourcing. Scoring stays in the workbook's own formulas.
How I built it
Pull the classifications out of the MSDS, apply the schema sheet that already lives in the country file, and write back into the existing template without breaking the live score formulas. Same idea as the use-summary table: a draft, then a person.
Deployment
Interactive pilot plus a local Python path, calling Bayer's AI gateway for model work. Secrets stay on the server. Same Pre-Flight guardrails as the use-summary tool: kill switch, rate limit, daily cost cap, metadata audit trail, injection-phrase flags, Agent Card and drift check. An expert still reviews before scores are used.
Stack
- Python
- PDF / MSDS extraction
- Excel write-back (openpyxl)
- Country schema sheets
- Bayer AI Gateway
- AI Pre-Flight guardrails