At Tata Teleservices, the security team received more than 1,000 abuse emails every day: phishing reports, spam complaints, DDoS alerts, suspicious attachments. Someone had to open each one and decide whether it mattered. The reports behind them were long and unstructured, so even the important ones took a while to act on.
My job was to automate that. By the end, manual analysis work had dropped by more than 95%, and the team was saving over 10 hours a week. The numbers are nice, but what I actually learned came from the gap between the problem I thought I was solving and the one the team had.
Sorting was only half of it
I started by thinking of this as a classification problem: read an email, put it in the right bucket. That part worked well with NLP on the text plus the email's metadata. But sorting alone didn't save the team much time. Knowing an email was phishing still meant reading a long, messy report to figure out what to do about it.
So the second half of the project was summarization: turning each unstructured security report into a short summary the team could act on. That's where most of the saved time came from.
The best tool is the one they already use
It would have been easy to build a separate dashboard and ask the team to check it. Instead I integrated everything into the Azure-based workflow they were already using, so the triage and the summaries showed up where the work already happened.
What I took away
The model is rarely the whole product. What made this useful was everything around it: understanding what the team did with an email after it was sorted, writing the summary for the person who has to act on it, and fitting into their day instead of adding to it. That's the part of building with AI I care most about now.