Software ERP supportato dall’IA per gestione rifiuti e riciclo

Risultati della tesi di master sull’identificazione, valutazione e prototipazione di casi d’uso di IA in ERP di settore.

2025-2026 Master Thesis AI & ERP Recycling 16 Use Cases 10 Expert Interviews

Le aziende di rifiuti e riciclo gestiscono flussi di materiali eterogenei, ordini urgenti e una regolamentazione rigorosa, mentre processi ERP come presa ordini, disposizione, ricevimento merci, routing e accesso alla conoscenza restano molto manuali. La tesi ha identificato 16 casi d’uso di IA, ne ha analizzati sei in profondità, li ha validati con dieci interviste a esperti e ha implementato il caso meglio classificato come prototipo NLP.

Vedi risultati chiave Sfoglia casi d’uso

Lo studio in sintesi

Research question: Which AI use cases are practically relevant for industry-specific ERP systems in waste management and recycling, how should they be prioritised, and what can a first prototype demonstrate?

16

AI use cases

Identified across core ERP-supported industry processes

6

Deep-dive cases

Evaluated in detail with structured scoring

10

Expert interviews

Validation with practitioners and domain experts

1

Working prototype

NLP-based order intake as top-ranked implementation

Risultati chiave

Data quality is the bottleneck

AI value is capped by incomplete, inconsistent or siloed ERP and operational data.

Non-technical barriers dominate

Organisation, process fit, trust and change management outweigh pure algorithm choice.

Assistance over full automation

Experts prefer AI that prepares decisions rather than replacing accountable staff actions.

Modular ERP integration

Use cases succeed when added as modular services with clear human review points.

No universal best use case

Ranking depends on company maturity, data readiness and regulatory context.

Skills shortage raises assistive AI value

Where specialists are scarce, assistive AI can protect throughput and knowledge access.

Classifica dei casi d’uso

Prioritisation of the six deep-dive use cases by weighted evaluation score.
Rank Use Case Score Assessment
1 UC-01 NLP order intake 3.78 Best entry point
2 UC-06 RAG compliance knowledge 3.68 Strong knowledge leverage
3 UC-03 AI disposition support 3.30 High process impact
4 UC-02 Visual classification 3.25 Promising with sensors/cameras
5 UC-04 IoT fill-level prediction 2.82 Depends on sensor coverage
6 UC-05 Dynamic routing 2.73 Complex constraints, later phase

Casi d’uso prioritari

UC-01 – NLP order intake 3.78

Extract structured orders from emails and messages.

Problem: Orders arrive as unstructured email/text and are typed into ERP manually.
Solution: NLP extracts customer, material, quantity, time window and site suggestions.
ERP fit: Creates draft sales/order objects for clerk confirmation.
Interview: Seen as fastest visible win with controllable risk.
Maturity: Prototype implemented in the thesis.
UC-02 – Visual material classification 3.25

Computer vision support for material identification.

Problem: Visual inspection is experience-heavy and hard to scale.
Solution: Camera-based classification suggestions with confidence scores.
ERP fit: Enrich goods-receipt / quality records.
Interview: Valuable where lighting and labelling variance are manageable.
Maturity: Conceptually strong; hardware-dependent.
UC-03 – AI disposition support 3.30

Assist planners with capacity and sequencing suggestions.

Problem: Disposition reacts late to short-notice changes.
Solution: Predictive/prescriptive hints for slots, resources and priorities.
ERP fit: Decision support inside planning boards.
Interview: High impact if trust and override UX are solid.
Maturity: Requires clean historical planning data.
UC-04 – IoT fill-level prediction 2.82

Predict container/fill levels to trigger logistics.

Problem: Pickups are often calendar-based rather than need-based.
Solution: Sensor + forecast driven collection triggers.
ERP fit: Generate service orders from predicted thresholds.
Interview: Attractive but capped by sensor roll-out cost.
Maturity: Feasible in niches with existing IoT.
UC-05 – Dynamic routing 2.73

Optimise routes under industry constraints.

Problem: Routing ignores live constraints and service windows.
Solution: Constrained vehicle-routing optimisation with dispatcher override.
ERP fit: Sync tours and orders with logistics planning.
Interview: Valuable later once order and geo-data quality improve.
Maturity: Complex; not the best first project.
UC-06 – RAG compliance knowledge 3.68

Retrieval-augmented answers over policies and SOPs.

Problem: Compliance and process knowledge is scattered across documents.
Solution: RAG assistants grounded in approved internal sources.
ERP fit: Contextual help beside transactions and master data.
Interview: Strong support where audits and regulation load are high.
Maturity: High leverage if governance of sources is clear.

Ulteriori casi d’uso

UC-07 – Predictive maintenance

Forecast equipment risk from usage and sensor history.

UC-08 – Sustainability reporting assist

Draft ESG/recycling KPIs from operational records.

UC-09 – Digital product passport support

Assemble traceability artefacts for materials and lots.

UC-10 – Material-flow anomaly detection

Flag unusual quantity/quality patterns early.

UC-11 – Returns / reverse logistics assist

Support classification and routing of return flows.

UC-12 – Inventory optimisation hints

Suggest stock and buffer adjustments by material class.

UC-13 – Demand forecasting

Predict inbound/outbound volumes for planning horizons.

UC-14 – Pricing / margin assistance

Provide context for commercial decisions under constraints.

UC-15 – Plant utilisation coaching

Highlight bottlenecks across yards and processing lines.

UC-16 – Assisted quality inspection

Combine checklist guidance with vision/NLP evidence capture.

Risultati delle interviste

Data often exists but is hard to use

Experts report data availability without trustworthy readiness for AI.

Process fit beats model novelty

Use cases fail when they ignore real ERP transaction paths.

Acceptance needs transparency

Users adopt AI when suggestions are explainable and overridable.

Technology is rarely the main blocker

Governance, ownership and change capacity dominate delays.

Domain adaptation is mandatory

Generic AI tools underperform without recycling-specific vocabulary and rules.

Start with low-regret entry points

NLP intake and knowledge assistants were preferred first steps.

Prototipo

  1. Step 1 – Capture inbound message

    Ingest customer email/text order requests into the prototype pipeline.

  2. Step 2 – NLP extraction

    Identify entities such as customer, material, quantity, date and location.

  3. Step 3 – Confidence & validation

    Score fields and highlight uncertain extractions for review.

  4. Step 4 – ERP draft mapping

    Map extracted fields to draft order structures for clerk confirmation.

  5. Step 5 – Human confirmation

    Staff accepts, edits or rejects before any operational posting.

Raccomandazioni

  1. Phase 1 – Data readiness

    Clean master data, define ownership and close critical media breaks.

  2. Phase 2 – Pick an entry use case

    Start with UC-01 or UC-06 where risk is controllable and value is visible.

  3. Phase 3 – Human-in-the-loop UX

    Design review, override and audit trails before automation claims.

  4. Phase 4 – Modular integration

    Expose AI as services beside ERP transactions, not as a monolith rewrite.

  5. Phase 5 – Pilot with KPIs

    Measure handling time, correction rate, adoption and exception volume.

  6. Phase 6 – Governance

    Set model, prompt and document-source governance for compliance.

  7. Phase 7 – Scale by process

    Extend to disposition, vision or IoT only after data and trust foundations hold.

  8. Phase 8 – Capability building

    Train key users and establish MLOps/DevOps ownership for sustainment.

Prospettive future

Process mining

Discover real ERP variants before automating the wrong path.

Digital twin of yard/plant flows

Simulate capacity and material-flow interventions.

Agentic AI assistants

Multi-step agents for preparation tasks under strict approval gates.

Multimodal models

Combine text, images and tabular signals in one decision context.

Tabular foundation models

Improve predictions on ERP table-heavy workloads.

Time-series foundation models

Stronger forecasts for volumes, fill levels and demand.

Digital product passports

Regulatory traceability will pull AI into documentation workflows.

Limitazioni

Not statistically representative

Ten interviews cannot generalise to the whole industry.

Company-dependent ranking

Scores shift with maturity, systems and regulation exposure.

No production field trial

Business benefit was not measured in live operations.

Prototype channel limits

NLP focus on message-style intake does not cover all order channels.

Data sensitivity constraints

Privacy and commercial confidentiality limited dataset breadth.

Feasibility ≠ value

Technical success does not automatically imply positive ROI.

Stack tecnologico

Backend

PHP
Laravel
API layer

Frontend

Livewire
Blade
Interactive review UI

AI / Analytics

NLP
ML
RAG concepts

Infrastructure

Docker
API deployment
Experiment tooling

Linea di sviluppo

  1. 2022 – Practical project

    DevOps cycle and customer portal foundation.

  2. 2023 – Bachelor thesis

    ERP-near enterprise software and digital intake.

  3. 2025/2026 – Master thesis

    AI use-case portfolio, expert validation and NLP prototype.