evidoria

← Back to browse

Good practice Imported

AI-Powered Semantic Triage for IT Support Tickets — Madeira's Regional Informatics Directorate

Portugal · Funchal · See the Portugal profile

Evidence: Observational / pre–post Top 50% 60/100 · Ask Evidence Copilot about this practice

Madeira's IT directorate (DRI) deployed a BERT + Random Forest triage system on 12,843 support tickets (2022-2024), reaching 83% accuracy, cutting repeat complaint categories 27% and raising classification efficiency 32%, per a peer-reviewed case study funded by Portugal's PRR.

83%
Classification accuracy (Jan 2022-Dec 2024)
0.80
F1-score
-27%
Repeat complaint categories change (6 months post-deployment)
+32%
Classification efficiency increase (6 months post-deployment)
0.21 seconds
Per-ticket processing time
0.47
Clustering silhouette score

Details

Maturity
Established
Promoter
Direção Regional de Informática (DRI), Governo Regional da Madeira — with ISTAR-IUL and VIPA (Funchal)
Period
2022-2025
Region (NUTS)
PT30
Keywords
public administration, IT service management, NLP, complaint triage

Context

Portugal's regional public services see growing volume and complexity of IT support requests (locally called PATs) submitted through the PatOnline and PATS systems used by the Autonomous Region of Madeira's government, with ambiguous language, inconsistent terminology and fragmented reporting across departments making conventional keyword-based triage unreliable. The Direcao Regional de Informatica (DRI), working with researchers from ISTAR-IUL (Instituto Universitario de Lisboa) and the VIPA laboratory in Funchal, built an AI pipeline combining multilingual BERT sentence embeddings, K-Means clustering into four complaint categories, a Random Forest classifier for automated categorisation, and Isolation Forest for anomaly detection, with SHAP values for interpretability. The project was funded under Portugal's Recovery and Resilience Plan (PRR).

Results

Evaluated on 12,843 complaint/ticket records drawn from 7,615 user interaction logs between January 2022 and December 2024, the system reached 83% classification accuracy (F1-score 0.80) and processed each ticket in 0.21 seconds end-to-end. In the six months after deployment, repeated complaint categories fell by 27% and classification efficiency rose by 32% (clustering silhouette score 0.47). The results come from a single regional agency and a specific ticketing context; the authors note that applying the approach elsewhere would likely require domain-specific fine-tuning or adaptive retraining, and no independent replication outside Madeira has yet been published.

Implementation

Indicative cost
Low (< €50k) — No budget figure is published; classified as low cost given the narrow single-department scope delivered through an academic research partnership under PRR co-funding.
Time to results
Medium (1–3 years) — Data and evaluation span January 2022 to December 2024 (about three years), including six months of post-deployment measurement.
Staffing & skills
Direcao Regional de Informatica (DRI), Governo Regional da Madeira, ISTAR-IUL (Instituto Universitario de Lisboa) researchers, VIPA laboratory, Funchal

Conditions for success

  • Combining multiple ML techniques (BERT embeddings, K-Means clustering, Random Forest, Isolation Forest) with SHAP interpretability
  • Sustained government-university partnership
  • PRR funding to support the initial build

Common failure modes

  • Authors note the model was evaluated in a single agency/context and would likely need domain-specific fine-tuning or retraining to transfer elsewhere
  • No independent replication outside Madeira has yet been published

Commonly funded by

Recovery and Resilience Facility (national plans) National / regional programmes

Indicative funding routes for practices of this type — always check each programme's current calls and eligibility rules.

Do you run this practice? Claim it — verified implementers get a public contact pathway and can propose corrections.

Data sources

Where this practice's information was retrieved from, and when.

Similar practices you may find useful