Case studyDecathlon

Multi-Agent System for IT Support

Proof of concept: 80% of recurring 'price mismatch' tickets resolved without a human

Automatic resolution

80%

Of the targeted ticket category

Specialised agents

3

Analysis, Research, Resolution

Orchestration

Airflow + LangGraph

Scheduled, stateful workflows

Stage

Proof of concept

Validated on a real ticket flow

Overview

IT support was saturated by a recurring ticket category — price mismatches — where every ticket triggered the same manual investigation across the same internal sources. I built a proof of concept: a set of specialised agents that qualify, investigate and correct these tickets autonomously, and escalate the rest with the investigation already attached.

The Challenge

A single category of recurring tickets was consuming a disproportionate share of support capacity. Each one required the same cross-checks across several internal systems, and the quality of the resolution depended on which engineer picked it up.

Key pain points

  • A recurring ticket category absorbing support capacity through near-identical investigations
  • Diagnosis requiring cross-checks across several internal data sources
  • Support time drained away from genuinely critical infrastructure incidents
  • Resolution quality varying with whoever handled the ticket

The Solution

An orchestrated multi-agent architecture with narrow, specialised roles rather than one generalist agent — chosen so that behaviour stays predictable and debuggable.

1

Orchestrated multi-agent architecture

Airflow schedules and picks up the ticket flow; LangGraph handles the state machine between agents, with explicit transitions rather than free-form delegation.

2

Specialised agents rather than one generalist

Analysis qualifies the ticket, Research gathers evidence across sources, Resolution applies the correction. Each has a narrow scope, which keeps the system predictable and debuggable.

3

Autonomous correction of price discrepancies

Where the discrepancy is identified with sufficient confidence, the system applies the correction itself rather than proposing it.

4

Auto-closure or documented escalation

Resolved tickets are closed automatically. Anything outside the confident path is escalated by email with the investigation attached, so a human starts from evidence rather than from scratch.

Results & Impact

80% automatic resolution

On the targeted ticket category

Measured on the recurring 'price mismatch' flow the proof of concept was scoped to.

Support capacity freed

Redirected to critical infrastructure

Engineers stopped repeating the same investigation and went back to incidents that actually needed judgement.

Escalations arrive pre-investigated

Evidence attached

When the system escalates, the human receives the analysis already done rather than a raw ticket.

Deliberately scoped as a POC

Not a production deployment

The goal was to validate the multi-agent architecture on a real ticket flow before committing to production.

Technologies used

PythonVertex AILangGraphAirflowMulti-Agent Systems

Ready to move your AI project forward?

Let's talk about what a similar engagement would look like in your context.

Get in touch