Building an Escalation Root-Cause Agent with Gemini and ADK

#ai #googlecloud #machinelearning #genai

Gen AI Academy APAC — Track 1 (AI Agents with Gemini, ADK, and Cloud Run)

Why I built this

I lead a customer service team of 25 agents at Amazon, handling both
buyer-side and marketplace seller support. A big part of my job is
reviewing escalated cases — calls or chats where a customer asked for
a supervisor — and figuring out why they escalated in the first
place. Was it a policy gap? A training issue? A system limitation
nobody flagged?

Right now, that review is manual. Every escalation gets read, tagged,
and turned into a coaching note by a human — usually me, or one of my
leads. It works, but it doesn't scale well, and patterns across dozens
of cases are easy to miss when you're reviewing them one at a time
between everything else on your plate.

So for Track 1 of the Gen AI Academy APAC program, I built an agent
that does the first pass of this analysis automatically: read an
escalation summary, classify the root cause against a standard
taxonomy, flag whether it looks like a repeat pattern, and draft a
coaching note — the same way I would, just faster and more
consistently.

What it does

The agent takes a case summary like this:

Customer requested a refund for a damaged item outside the standard
return window. Agent denied it citing policy; customer says a rep
last month approved a similar exception for someone else.

And returns a structured analysis:

{
  "root_cause_category": "policy_misapplication",
  "severity": "medium",
  "is_likely_repeat_pattern": true,
  "pattern_reasoning": "Inconsistent policy application across agents
  suggests a training or documentation gap rather than an isolated
  error.",
  "coaching_note": "..."
}

It's built on Google's Agent Development Kit (ADK) with Gemini
as the underlying model, and deployed as a live service on Cloud
Run. The agent has one tool — a lookup function for the standard
root-cause taxonomy — which keeps the categories consistent and easy
to update without touching the core prompt.

For batch review, I also built a runner script that processes a whole
CSV of cases at once and outputs a summary report — closer to how I'd
actually use this day to day, reviewing a batch of the week's
escalations rather than one at a time.

What I learned building it

A few things stood out going through this:

Getting the taxonomy right mattered more than the prompt engineering. Once I had six clear, non-overlapping root-cause categories with real definitions, the model's classifications got noticeably more consistent.
Structured output is where this becomes actually useful. Free-text analysis is fine to read once; a fixed JSON shape is what lets you aggregate patterns across dozens of cases and actually spot something you'd have missed manually.
Deployment had more friction than the coding did. Getting a fresh Google Cloud project's IAM permissions set up correctly for Cloud Build and Artifact Registry took longer than writing the agent itself — worth knowing going in if you're new to GCP.

Try it / see the code

Live agent (Cloud Run): https://escalation-root-cause-agent-507816071353.us-central1.run.app
Source code: https://github.com/aadilmajeedlone/escalation-root-cause-agent

The repo includes a README.md with full setup instructions if you
want to run it yourself, plus a sample CSV of synthetic escalation
cases to test against (no real customer data used anywhere in this
project).

What's next

The natural extension is Track 2 — instead of feeding it a CSV
manually, connect it directly to a real ticketing system via MCP so
it can pull escalations and analyze them without a manual export
step. That's the version I'd actually want running against my team's
queue.

Built for the Google Cloud Gen AI Academy APAC program, in
partnership with Hack2Skill.

DEV Community