All work
162021MoHRE

Citizen Service Request Auto-Router

Routing bilingual citizen requests to the right team — cutting misrouting from 28% to 6%.

Before

Bilingual citizen requests were misrouted about 28% of the time, causing SLA breaches across labour-service teams.

After

An intent + topic classifier routes each request to one of 22 specialist teams with a confidence-gated escalation path. Misrouting dropped from 28% to 6%, first-touch resolution improved 19 points, and labour-services SLA breaches fell 41%.

Challenge

Bilingual citizen requests were misrouted ~28% of the time, causing SLA breaches across labour service teams.

Approach

Intent + topic classifier on top of the request text routes to one of 22 specialist teams, with a confidence-gated escalation path.

How it was built

  1. 1

    Intent + topic classification

    Phase 1

    Built an Arabic-capable classifier over the request text to predict intent and topic across 22 specialist teams.

  2. 2

    Confidence-gated routing

    Phase 2

    Auto-routed confident predictions and escalated uncertain ones, so the tail of ambiguous requests didn't degrade into misroutes.

  3. 3

    Measure against SLAs

    Phase 3

    Tracked misrouting, first-touch resolution, and SLA breaches — the metrics leadership actually feels — to prove the impact.

Key architecture decisions

Confidence-gated escalation

Why · Forcing a guess on low-confidence requests is what caused misroutes. Escalating the uncertain ones is what pushed misrouting down to 6%.

Bilingual classification from the start

Why · Citizen requests arrive in Arabic and English; handling both natively was table stakes for accurate routing.

Impact

  • Misrouting dropped from 28% to 6%
  • First-touch resolution improved 19 percentage points
  • Reduced SLA-breach rate on labour services by 41%
-22pp
misroute
+19pp
ftr
-41%
sla

What I'd tell someone building this

  • 01 · Let the classifier abstain — a confidence gate beats a forced wrong guess.
  • 02 · Tie the model's metrics to the ones leadership feels (SLA breaches), not just accuracy.
  • 03 · Routing quality compounds: fewer misroutes lifts first-touch resolution too.

Tech stack

BERT (Arabic)ML.NETAzure MLWeb API

Ask anything about Citizen Service Request Auto-Router

AI scoped to this project · GPT-OSS 120B