AI/ML 
Software Development 
Businesses write off an average of 8% of revenue as bad debt every year – not because debtors are unreachable, but because collections workflows treat every overdue account identically (Hilton Baird / Chaser 2026). 92% of businesses report invoices are typically paid after the due date; 55% of B2B invoiced sales in the US are overdue (Atradius 2023 / Chaser 2026). Late payments cost the UK economy £11 billion per year, forcing 50,000 SMEs to close annually (DBT 2025). VM.PL builds custom ML systems for accounts receivable and debt recovery: payment probability scoring, risk classification with SHAP explainability, channel-optimized collections automation, and LLM-powered debtor self-service portals. GDPR- and CCPA-ready, EU-based legal entity. First system in production in 6-12 weeks.
The AI for debt collection market was $3.34 billion in 2024 and is projected to reach $15.9 billion by 2034 at a 16.9% CAGR (Kaplan Group 2026). AI increases recovery rates by 25-35% while cutting collection costs by 20-50% (Kaplan Group / Experian 2024) – and conversational AI handles up to 80% of routine debtor queries, freeing collections staff for complex cases. 91% of mid-sized firms with fully automated AR systems report improved cash flow and growth; among firms that have automated only a few AR tasks, that figure drops to 44%.
What rule-based collections cost – and where ML outperforms
Rule-based collections apply fixed intervals and undifferentiated escalation steps regardless of debtor risk. A customer who missed a single payment for the first time gets the same collection letter as a chronically late-paying account. Valuable payers receive unnecessary friction; steerable cases lose the early intervention window.
| Criterion | Rule-based approach | AI-powered solution by VM.PL |
| Risk assessment | Fixed debtor categories: good / medium / bad · updated periodically | Continuous ML scoring in real time · updated daily or on each transaction · SHAP explanation per decision |
| Collection timing | Fixed schedule (+14 days, +30 days regardless of risk) | ML-optimal timing per debtor: earlier for high-risk, delayed for low-risk with high probability of self-cure |
| Channel selection | Letter or email by default across all accounts | Channel optimization: email, SMS, self-service portal, voice · selected per debtor engagement profile |
| Escalation | Time-based: all accounts escalate through same stages on the same schedule | Score-based: faster escalation for high-risk, gentler path for low-risk · late-stage handoff to collections agencies triggered by ML, not calendar |
| Explainability | Internal rule logic · not auditable for individual decisions | SHAP values: every scoring decision explained · GDPR Art. 22 / CCPA audit trail · challengeable by the debtor |
AI collections increase recovery rates by 25-35% (Kaplan Group / Experian 2024) and reduce operational costs by 30-50%. On a $10M annual receivables portfolio with a current 8% bad debt rate ($800K written off), a 25% improvement in recovery rate recovers $200K per year. Timing matters critically: Tesorio analysis of more than $100 billion in receivables processed found that once an invoice crosses the 120-day mark, collection probability drops to approximately 20-30%. Every day an account sits in an undifferentiated queue is a day closer to that cliff. In a documented $2M portfolio of accounts 60+ days past due, AI collection recovered $936K versus $252K from a traditional agency and $320K from in-house collections – a 3.7× improvement over the agency model (AgentCollect 2026, vendor benchmark on their own client portfolio).
How VM.PL calculates ROI for AI-powered AR – savings, break-even, and TCO
Accounts receivable AI recovers investment through higher collection rates, lower processing cost per account, and earlier intervention on high-risk receivables – all three compounding from the first production month. VM.PL models all three in the Discovery Phase using the client’s actual receivables data – not industry benchmarks.
| Cost block / lever | Description | Range |
| Recovery rate improvement | ML-optimized timing and channel selection increases payment collection in the early-stage window; fewer accounts reach expensive late-stage escalation | 2-8 percentage point improvement in collection rate, depending on receivables volume, current DSO, and debtor profile mix |
| Processing cost reduction | Automation of collection runs, channel routing, and communication classification reduces manual handling per account | 30-60% fewer manual processing steps on Tier-1 accounts · collections staff redeployed to complex exception cases |
| Collections agency and legal cost avoidance | Score-based escalation transfers fewer accounts to external collections agencies or attorneys; early payment arrangements reduce late-stage costs | Collections agency fees typically 15-25% of receivable face value · each prevented transfer saves that fee directly |
| Discovery + Definition | Data audit, architecture planning · delivers Concept Document and Architecture Blueprint | Lowest cost block · Go/No-Go decision before main investment |
| Delivery + Direction | Model training, ERP integration, monitoring · scales with complexity and receivables volume | Main investment · typical payback 6-18 months depending on receivables volume and current write-off rate |
Break-even calculation: VM.PL quantifies the break-even point from the client’s actual receivables volume, average invoice value, current DSO, and write-off rate – as a concrete part of the Concept Document, before any commitment is made. McKinsey (January 2025) found that optimizing AR procedures can improve receivables-related working capital by 30% or more within weeks. For organizations with $100M in annual revenue, each 10-day reduction in DSO returns approximately $548,000 to working capital (Lydonia AI). The Billtrust 2026 AR Benchmark Report shows that automated AR clients averaged 39-day DSO in 2025, against a global average of 50-54 days – a 15-day advantage that compounds across the fiscal year.
How VM.PL builds payment probability and risk classification models
Payment probability scores are generated for every open receivable at 30, 60, and 90-day horizons – and updated whenever new data arrives, not on a monthly review cycle. Risk classification happens automatically and dynamically: payment behavior, contact patterns, and receivable structure feed the model continuously.
| ML method | Application | Data requirements |
| XGBoost / LightGBM | Payment probability (binary: pay/default · or multi-class probability bands) · risk score calculation · feature importance for SHAP explanation · fastest to train and most interpretable for business stakeholders | Historical payment data, receivable structure, account attributes · minimum 12 months · works well from 10,000+ labeled examples |
| LSTM (time-series models) | Temporal payment patterns: seasonal default cycles, behavioral change detection, multi-period trend modeling · captures sequential dependencies XGBoost misses | Sequential payment history · minimum 18-24 months · particularly valuable for subscription billing and recurring payment scenarios |
| Isolation Forest (anomaly detection) | Sudden behavioral changes, fraud indicators, account takeover signals · flags accounts that deviate significantly from their own historical pattern without requiring labeled fraud examples | Live transaction data · no fraud labels required · runs continuously as a second layer over the primary scoring model |
| SHAP (explainability layer) | Generates a human-readable explanation for every individual scoring decision · applied on top of all models · essential for GDPR Art. 22 compliance and debtor challenge rights · enables audit trails | Applied to all primary models · no additional data required · adds negligible inference time |
A debtor receives Risk Band C. The SHAP output reads: “Payment history weight +35%; last contact channel response −15%; receivable age +20%; account tenure +10%.” The collections agent sees the reason, not just the score. The debtor can challenge the classification. The compliance team has a documented audit trail. All three requirements – GDPR Art. 22, CCPA, and internal governance – are satisfied by the same output.
Data preparation for ML models: audit, ETL, and feature engineering
50-70% of a debt recovery ML project is data preparation, not model development. Accounts receivable data from ERP and billing systems is rarely ML-ready: inconsistent debtor IDs, missing payment timestamps, booking errors, and historical system migrations create gaps that would systematically bias any model trained on them. VM.PL starts every project with a structured data availability audit.
- Data audit (Discovery Phase). VM.PL analyzes payment history, debtor IDs, receivable structure, and contact history data for completeness, consistency, and GDPR/CCPA relevance. Typical findings: gaps in payment data across historical system migrations; inconsistent debtor classification; missing contact channel records. The audit result determines whether the project receives a Go.
- ETL pipeline development (Definition Phase). VM.PL builds robust pipelines that extract, clean, and transform raw data from SAP FI-AR, Microsoft Dynamics 365 Finance, Oracle Financials, NetSuite, Sage Intacct, or custom billing systems into ML-ready features. Stack: Python 3.11+ / FastAPI · pgvector/PostgreSQL 16 as feature store · Azure Data Factory for batch extraction from enterprise ERP systems. Schema validation, null handling, and type consistency run automatically on every pipeline execution.
- Feature engineering. Raw data becomes forecast-relevant signals: Recency/Frequency/Monetary metrics from payment history, contact channel response rates, receivable age vectors, seasonal indices, and account tenure signals. VM.PL defines a feature catalog in the Definition Phase documenting why each feature enters the model – GDPR Art. 22-compliant and audit-ready.
12 months of payment history for XGBoost/LightGBM models; 18-24 months for LSTM time-series models. The Discovery Phase data audit assesses this before any training commitment.
Calibrating scoring models: precision-recall trade-off and cut-off optimization
A payment probability model outputs a score between 0 and 1 for each receivable. Where that threshold sits is a business decision, not a modelling choice.
| Parameter | Description | Business consequence |
| Precision | Share of accounts classified as high-risk that actually default | High precision = fewer unnecessary escalations · valuable payers are not over-communicated · customer retention protected |
| Recall | Share of accounts that actually default that are correctly flagged as high-risk | High recall = fewer missed defaults · more early interventions on genuinely at-risk accounts · lower write-off rate |
| F1 Score | Harmonic mean of precision and recall · used when both matter equally | Balanced optimization for general-purpose risk scoring |
| Precision-optimized cut-off | Higher threshold: fewer accounts flagged as high-risk | Use when customer relationship protection is the primary objective: telecoms, insurance, premium B2B |
| Recall-optimized cut-off | Lower threshold: more accounts flagged · higher escalation volume | Use when maximizing recovery rate is the primary objective: consumer credit, BNPL, high-volume low-value receivables |
| AUC-ROC | Area Under the ROC Curve · model-level performance metric · measures discrimination ability independently of cut-off | Typical target: AUC-ROC > 0.80 for production deployment · VM.PL validates on held-out data before go-live |
VM.PL simulates the financial impact of different cut-off settings on the client’s actual receivables data. The output is a precision-recall curve with dollar values attached: “a cut-off of 0.65 recovers $X more per year than 0.75, but increases escalation volume by Y accounts.” The client selects the cut-off based on that tradeoff, not an abstract metric. Finance teams that measure AR effectively and implement ML-driven cut-off optimization reduce DSO by 20-25% and improve collection rates by 30-35% (ChatFin 2026 AR benchmarks).
LLM-powered debtor communication and self-service portals
Collections communication is a high-volume, high-variance problem. A large receivables operation sends thousands of communications per week – each requiring the right tone, the right offer, and the right timing for the specific debtor profile. Rule-based templates work on straightforward reminders. The moment a debtor’s situation has any nuance – a partial payment dispute, a business relationship that matters, a hardship case – the template produces the wrong tone. LLM-based generation handles that variance.
- Channel-optimized outreach generation. The ML model selects the optimal channel per debtor (email, SMS, in-app message, voice) based on past response rates. The LLM generates the communication – tone and offer calibrated to the risk score. A debtor scoring 0.8 (high payment probability) receives a gentle reminder with a self-service payment link. A debtor scoring 0.2 receives a more direct communication with an installment plan offer.
- Debtor self-service portal. An LLM-powered chat interface allows debtors to query balance, request payment plans, submit documentation, and arrange extensions – 24/7, without a collections agent. Conversational AI handles up to 80% of routine debtor queries (industry benchmark). The portal is powered by LightRAG (lightrag-hku 1.4.9.11) on pgvector/PostgreSQL 16, with Auth.js v5 for role-based access. Rule-based outreach is also a compliance liability: the CFPB received 207,800 debt collection complaints in 2024, an 89% year-over-year increase; 45% were about debts consumers say they don’t owe, pointing to human error in manual processes (CFPB FDCPA Annual Report 2025). Consistent, documented AI interactions reverse that pattern.
- Early-stage, mid-stage, late-stage routing. Early stage (0-30 days overdue): digital self-service, automated payment link, installment plan offer. Mid-stage (30-90 days): score-based channel selection, payment arrangement with explicit timeline. Late stage (>90 days): ML decides on transfer to external collections agency with automated handoff package – complete case documentation, communication history, and risk score – ready for agency intake. Regulatory context for US deployments: FDCPA and CFPB Regulation F cap outbound contact attempts at 7 per 7-day period per debt (federal floor); the NYC SHIELD Rule (effective September 2026) tightens this to 3 attempts per 7 days for New York City. The FCC’s February 2024 ruling classifies AI-generated voices as an “artificial or prerecorded voice” under TCPA, requiring prior express consent for AI voice calls. California SB 1286 (effective July 2025) extends the Rosenthal Fair Debt Collection Practices Act to commercial debts up to $500,000 with statutory damages of $100-$1,000 per willful violation. VM.PL maps applicable regulations in the Definition Phase and builds them into contact frequency logic, consent management, and audit trail.
Industry-specific requirements for ML-driven debt recovery
A consumer credit portfolio and a B2B manufacturing receivables book share almost nothing – different default drivers, different contact sensitivities, different regulatory frameworks. VM.PL maps the specific requirements for each in the Definition Phase before building the model.
| Industry | Typical receivable structure and default patterns | Specific requirements and VM.PL adaptations |
| Insurance (P&C, life) | Premium receivables with known due dates · seasonal default cycles · policyholder data history goes back years | NAIC Model Bulletin compliance (US) · FCA/PRA (UK) · GDPR Art. 22 for profiling in policyholder decisions · SHAP explainability required for complaint management · Human-in-the-Loop for policy-affecting decisions |
| Energy / Utilities | High share of residential customers · seasonal peaks (annual bill reconciliation) · high-volume low-value receivables | Collections cost economics: AI automation particularly valuable on small-amount high-volume receivables where manual cost > receivable value · smart meter data as additional prediction feature · state utility commission rules on collections timing (US) |
| E-commerce / BNPL | High volatility, short payment windows, broad debtor spectrum · Buy-Now-Pay-Later structurally increases default exposure | Real-time scoring at point of purchase or checkout · integration into shop backend or payment gateway API · Isolation Forest for fraud signal detection · CFPB (US) / FCA (UK) consumer credit compliance |
| Fintech / Factoring | Sale of overdue receivables portfolios · portfolio-level valuation, not individual case management | Portfolio scoring model · Monte Carlo simulation for default scenario ranges · SHAP-explained portfolio value for investor due diligence · SEC / FCA disclosure requirements |
| B2B trade credit | Commercial receivables · longer terms (net 30-90) · lower volume, higher individual invoice value · relationship sensitivity | Precision-optimized cut-off (protect client relationships) · multi-contact hierarchical communication · SOX Section 13 record-keeping for public companies · preference for early payment arrangement over escalation · note: FDCPA applies to third-party debt collectors, not first-party creditors collecting own B2B debt – many enterprises adopt FDCPA standards regardless as brand integrity best practice |
| Residential / HOA | Rent arrears and HOA dues · social sensitivity component · lower case volume, high per-case importance | Gentle communication approach required by law in many jurisdictions · payment plan and hardship program integration · GDPR / CCPA for sensitive household data · US: Fair Debt Collection Practices Act (FDCPA) compliance |
Integrating AI debt recovery into ERP and billing systems
Existing ERP and billing systems stay untouched. VM.PL builds a decoupled Python 3.11+ / FastAPI service layer that communicates with SAP FI-AR, Microsoft Dynamics 365 Finance, Oracle Financials, NetSuite, Sage Intacct, and custom billing systems via REST API or direct database interface. The AI layer reads receivables data, writes risk scores and collection actions back to the existing system, and is fully rollback-capable.
| System type | Compatible systems | Integration method |
| ERP / Finance | SAP S/4HANA (FI-AR), Microsoft Dynamics 365 Finance, Oracle Financials, NetSuite, Sage Intacct | REST API · SAP OData / IDOC · Azure Data Factory for batch extraction · real-time event trigger on invoice overdue status change |
| Billing / Subscription | Stripe, Zuora, Chargebee, Recurly, custom billing platforms | REST API or webhook · event-driven: payment failed event triggers immediate ML scoring update |
| CRM | Salesforce, Microsoft Dynamics CRM, HubSpot | REST API · risk score surfaced in CRM account record · collections action log written back to CRM |
| Communication | Email (SendGrid, Mailgun) · SMS (Twilio) · voice (Twilio, Amazon Connect) · in-app notifications | Channel-selection API · LLM-generated message body · delivery and open tracking feeds back into scoring model |
| Collections agencies | Standard electronic file transfer for late-stage handoff · agency-specific format requirements | Automated handoff package: case documentation, payment history, communication log, risk score · format per agency contract |
Stack: Python 3.11+ · FastAPI · LightRAG (lightrag-hku 1.4.9.11) · OpenAI SDK via Azure OpenAI Service · pgvector/PostgreSQL 16 · Drizzle ORM 0.45 · asyncpg · Azure Container Apps · Azure Blob Storage · Next.js 16 + shadcn/ui for collections dashboard · Auth.js v5 (role-based access: COLLECTIONS_AGENT / MANAGER / COMPLIANCE) · Vitest 4 for Next.js frontend.
The 4D methodology for AI debt recovery projects
No ML model without a validated data foundation. VM.PL starts every AR/collections project with a data availability audit – payment history, data quality, GDPR/CCPA compliance – before any architecture is proposed.
| Phase | Activities | Deliverable | Duration | Go/No-Go gate |
| Discovery | Receivables data audit (payment history, debtor IDs, receivable structure, contact logs) · GDPR / CCPA / FDCPA assessment · use case validation · risk class design · ROI modelling on actual write-off rate and DSO · ERP integration feasibility | Concept Document (Go/No-Go) | 1-2 weeks | Go if: min. 12 months payment history available; sufficient labeled default examples; ERP API accessible; compliance path identified. No model training before this gate. |
| Definition | ML architecture (XGBoost / LSTM / Isolation Forest) · feature catalog · cut-off design and precision-recall trade-off · channel integration plan · SHAP explainability design · ERP/billing API integration design · GDPR Art. 22 / CCPA / FDCPA compliance architecture | Architecture Blueprint | 1-2 weeks | Approved by AR director and compliance lead; SHAP and Human-in-the-Loop design signed off; ERP integration confirmed |
| Delivery | ETL pipeline · feature engineering · model training and validation on held-out data · cut-off calibration · ERP/billing API integration · LLM communication pipeline · self-service portal (Next.js 16 + shadcn/ui) · Azure Pipelines CI/CD · Vitest 4 for Next.js frontend · formal AUC-ROC and precision-recall accuracy gate | Production ML system | 6-12 weeks | AUC-ROC target met on held-out data · ERP integration stable · SHAP output validated · portal live · rollback tested |
| Direction | MLOps monitoring (Azure Application Insights) · score drift detection (PSI > 0.20 triggers review) · automated re-training on new payment data · cut-off reviews (quarterly or on regulatory change) · FDCPA / CCPA / GDPR compliance updates | Ongoing roadmap | Ongoing | Monthly accuracy report · score drift alert triggers review within 48h · annual compliance documentation refresh |
Timeline and cost. First production payment probability model with ERP integration: 8-10 weeks, $35K-85K build, $7K-20K/yr operating. Full platform (scoring + LLM communications + self-service portal + MLOps): 12-18 weeks, $70K-$160K. Break-even: typically 6-18 months depending on receivables volume and current write-off rate. The Discovery Phase produces the ROI model from your actual data.
Why US and UK companies work with VM.PL on AR and debt recovery
VM.PL Software House (founded 2004, Wrocław, Poland) has shipped 475+ projects for US, UK, and Western European clients. 150+ engineers, NPS 71. English-first communication, 6-9 hours ahead of US East Coast, 1-4 hours of daily synchronous overlap.
Five differences from off-the-shelf AR automation platforms:
- Custom ML on your payment data. SaaS collections platforms train on industry-average default patterns. VM.PL trains on the client’s own payment history – the specific debtor behavior, seasonal patterns, and risk signals that appear in their portfolio. A model trained on 3 years of your own receivables consistently outperforms a generic platform model on your accounts.
- SHAP explainability as standard. Every risk classification is explainable to the debtor, the compliance team, and the regulator. GDPR Art. 22, CCPA automated decision-making requirements, and FDCPA obligations are addressed by the same SHAP output layer. This is built into the architecture, not an optional add-on.
- No ERP replacement. SAP FI-AR, Microsoft Dynamics 365 Finance, Oracle Financials, NetSuite – the ML layer integrates via REST API or native ERP API without data migration or system replacement. Rollback is a configuration change.
- Privacy by architecture. Debtor data is pseudonymized before ML processing. On-premise deployment is available as default for clients where receivables data cannot leave the enterprise infrastructure. No data transfer to third-party services. DPA (Data Processing Agreement) with VM.PL is a standard contract component.
- Full ownership. Source code, model weights, ETL pipelines, and dashboards belong to you. No ongoing VM.PL licensing fees. Transfer to an internal ML team or a different vendor without rebuilding.Reference: https://vmsoftwarehouse.com/operational-efficiency-for-debt-collection-companies/