Classifying 10K+ emails a month with RPA + ML.

§ 01 — The problem

Enterprise clients received high volumes of email across 200+ configured sources — support, operations, finance — that all needed reading, categorizing, and routing to the right team or system. Done by hand, it was slow, inconsistent, and a poor use of people who should have been resolving issues, not sorting them.

§ 02 — The classification model

Working with an AI scientist, I helped train and deploy an email-classification system on UiPath's Communications Mining API. It learned the clients' real categories from historical mail and reached 85–90% accuracy — high enough to route confidently, with low-confidence cases falling back to a human.

§ 03 — Bridging ML to enterprise systems

A model that classifies isn't worth much until its output drives real work. I integrated the classification results into SAP and ServiceNow — turning "this email is category X" into a structured action: the right ticket, the right queue, the right downstream process. That gap, from a prediction to a business outcome, is where most of the value (and most of the engineering) lived.

§ 04 — Hardening the workflow

I redesigned the exception handling, routing logic, and validation across the SAP/ServiceNow integrations — catching the malformed, ambiguous, and edge-case inputs that break naive automations. Tightening those paths improved overall workflow throughput by 70%.

§ 05 — Computer Vision for legacy interfaces

Some of the systems involved had no API — only old desktop UIs over remote sessions. I used UiPath Computer Vision to automate extraction and form-filling across those legacy interfaces, so the pipeline could reach systems that otherwise couldn't be touched programmatically.

§ 06 — Outcome

The system classified and routed 10K+ emails a month across 200+ sources at 85–90% accuracy, and the workflow rework lifted throughput by 70%. It also gave me the instinct I carry into LLM work now: a model is the easy part; the hard, valuable part is wiring its output safely into the systems and processes that depend on it.

§ 07 — What I took from it

This is where I learned to treat model output as a signal to be validated and routed, not trusted blindly — confidence thresholds, human fallback, exception paths. It's the same posture I apply to LLMs today: schema-validate, retry, and keep a human gate where a wrong answer is expensive.