Payment-behaviour prediction & collections prioritisation (AR, dunning, debt recovery)

Most companies still chase receivables by age bucket: everything 30 days overdue gets dunning letter 2, everything over 60 days gets a call. That wastes collector time on customers who would have paid anyway and reaches the real risks too late. Payment-behaviour models predict, for each invoice or account, when it will be paid, how likely it is to slip, and which action actually changes the outcome. Collectors get a ranked worklist, low-risk customers get a light digital reminder or nothing, and finance gets an invoice-level inflow forecast. In consumer debt recovery the same models must also detect customers in financial difficulty and offer forbearance.

Problem & value

• Who has the problem: B2B accounts-receivable teams in the Mittelstand and in enterprises (shared-service centres), and first-party collections teams at banks, BNPL providers, telcos, utilities and public creditors. Debt-collection agencies apply the same logic to purchased or assigned portfolios.
• Today: static dunning runs in the ERP (e.g. SAP dunning procedure, DATEV OPOS lists), age-bucket worklists and collector intuition. All customers get the same letter at the same time, disputes are found late and expensive channels (calls, agencies) are spent on the wrong accounts.
• Why AI helps: gradient-boosted models on payment history predict payment date and lateness much better than averages. Uplift or treatment models estimate which action changes behaviour rather than which customer is merely risky. Capacity-aware prioritisation focuses scarce collector hours on the cases with the highest expected value.
• Reported results (vendor-reported, no independent audit): HighRadius lists a 20-day DSO reduction at distributor DXP [1]. Upflow states a 30% average DSO reduction within 3 months [3]. collect.AI (Aareal Bank Group) claims ≥30% lower receivables-management cost and >90% of receivables settled before due date across >EUR 20 bn processed [4]. InDebted claims 35% lower operating cost and 15% more cash collected for its first-party platform [2]. Treat these as upper bounds and verify them in a controlled pilot.

Variants

• B2B AR (core variant): invoice-level payment-date prediction plus worklist prioritisation in ERP-centric order-to-cash. Typical in manufacturing, wholesale, professional services and SaaS.
• SMB / freelancer: the accounting tool or a lightweight AR app (Upflow, Chaser, Kolleno-type tools) adds "likely late" flags and automated reminder sequences. Rule-based dunning plus a simple score is often enough.
• Consumer collections & recovery (banks, BNPL, utilities, telcos): propensity to pay, self-cure probability (will the customer pay without contact?) and roll rates between arrears buckets drive treatment paths. These paths include digital self-service, payment plan, agent call, forbearance or sale or assignment to an agency. This variant must include vulnerability detection and forbearance under CCD2 (Directive (EU) 2023/2225, Art. 35) and, for servicers of bank NPLs, the Credit Servicers Directive (EU) 2021/2167, transposed in Germany by the Kreditzweitmarktgesetz under BaFin supervision [6].
• Debt-collection agencies (Inkasso): portfolio valuation and case-level recovery scoring for bought or assigned debt. Agencies such as EOS, Intrum, Lowell, PAIR Finance and Creditreform run this in-house.
• Public sector: municipal treasury and tax or fee offices (Kasse, Vollstreckung) prioritising reminders and enforcement. Keep a human decision on enforcement and see the regulation section.
• Private use: not a target audience. Consumers are on the receiving end. Debt-advice organisations (Schuldnerberatung) are a separate, non-AI-scoring setting.

How it works

1. ERP/AR open items + payment history + contact log + external credit signals
2. feature store (customer-level lateness stats, trend, terms, disputes, seasonality, credit-agency data)
3. model 1: payment-date / lateness prediction (GBM regression or classification; survival/AFT model)
4. model 2: treatment response / uplift per action (no action, email, SMS, letter, call, plan, agency)
5. prioritisation: expected value = amount × P(late) × uplift − contact cost, under collector capacity
6. policy layer: compliance rules (contact frequency, quiet hours, vulnerability, disputes, forbearance)
7. worklist + scheduled digital actions (dunning engine / workflow) + LLM-drafted messages
8. outcomes logged
9. retraining and A/B or bandit evaluation of treatments

• Prediction: LightGBM, XGBoost or CatBoost on invoice- and customer-level features is the workhorse. HighRadius, for example, names Random Forest and gradient-boosted trees for its "Predict Payment Date" feature [1]. Survival models handle censoring because open invoices have no payment date yet. lifelines (Cox, AFT, cure models) is a common Python toolkit for this [8].
• Treatment choice: uplift models (two-model or causal forests) or contextual bandits learn which action works for which segment. Rule-based champion/challenger testing is the conservative start.
• LLM layer: drafts tone-adjusted reminder emails, summarises account history for the collector, classifies replies (dispute, promise to pay, hardship). That layer is AGENT-023's territory; this use case supplies the scores.
• Human in the loop: collectors can override the worklist. Payment plans, forbearance, credit blocks, agency hand-over, credit-bureau reporting and enforcement remain human decisions.

Data preparation

Payment prediction needs the open-item and payment history from the ERP or AR ledger (SAP FI-AR, Dynamics, DATEV, NetSuite), customer master data, the dunning and dispute log, and contact notes. In an SMB this typically lives in the ERP plus Excel lists kept by individual clerks, and promises to pay sit in e-mails or phone notes. Typical gaps are duplicate customer accounts, missing dispute reasons, payments booked without clear invoice matching, and two or three years of history with changed payment terms.

Preparation steps: export ledger data via standard interfaces (PREP-005), merge duplicates and define one customer key, agree on a clear label such as "paid more than 14 days late" (PREP-019), harvest it from booked payments (PREP-020), and pseudonymise before any external scoring (PREP-022). Collectors should record disputes and promises in structured fields instead of free text (PREP-021). Check readiness with GUIDE-003.

Rough share of effort: about 30 to 40 percent of the project, mostly on matching, cleaning and label definition. Buying a product shifts part of this to the vendor connector, but master data clean-up stays with you.

Implementation paths

Which path for whom

• SMB: Path A. Mature AR and dunning products with DATEV, SAP and Dynamics connectors exist at SMB prices and dunning is a standard process.
• Enterprise: Path B. Receivables are core to cash flow and volumes are high, so scoring on your own ledger data via cloud APIs or an internal platform avoids seat costs and lock-in.
• Developer: Path C. A gradient-boosted model on payment history runs on HW1 with open-source libraries and gives full control over features and explainability.
• Public: Path C. Citizen debt data is sensitive and procurement rules favour sovereign, self-hosted or German-cloud operation.

Path A — Buy (off-the-shelf)

• ERP-native: SAP S/4HANA Collections Management and SAP "behavioral insights for contract accounting" (FI-CA payment-risk prediction for utilities, telco and public sector) [5]. Microsoft Dynamics 365 Finance "Finance insights" with customer payment predictions.
• AR / order-to-cash suites: HighRadius, Sidetrade (FR/EU, Aimie AI), Esker (FR/EU), Serrala (DE), Billtrust, Quadient AR, Emagia, Versapay. EU-focused mid-market: Upflow (FR), collect.AI (DE, Aareal Bank Group) [4], Kolleno (UK), Chaser (UK).
• Consumer collections platforms: receeve (Berlin-founded, now InDebted's first-party platform "Receive") [2], Collenda (DE), Experian PowerCurve Collections, FICO Platform for collections.
• Choose A when: you have a standard ERP and up to tens of thousands of open items, and want results within one quarter. Check EU hosting, the ERP connector (SAP, DATEV, Dynamics), whether scores are explainable, and whether you can export outcome data.

Path B — Build on cloud APIs

• AutoML / tabular services in EU regions: Azure Machine Learning AutoML (West Europe / Germany West Central), Google Vertex AI Tabular (europe-west3), AWS SageMaker Autopilot (eu-central-1), Databricks AutoML. Train on an extract of AR history, serve scores via batch jobs back into the ERP.
• LLM APIs (Mistral, Aleph Alpha / STACKIT-hosted models, Azure OpenAI EU Data Zone, Claude on AWS Bedrock Frankfurt) only for message drafting and reply classification. Keep personal data minimal.
• Orchestration: n8n or Power Automate triggers dunning actions. SAP BTP integration for S/4HANA.

Path C — Self-host (open source)

• Stack: Python with LightGBM/XGBoost/CatBoost, lifelines or scikit-survival for time-to-payment, EconML or CausalML for uplift, Vowpal Wabbit or simple Thompson sampling for channel/timing bandits, OR-Tools for capacity-constrained worklists. MLflow for model registry, Feast for features.
• Tabular foundation models: TabPFN 3.5 handles up to 1M rows [7] and is strong on small AR datasets. Note that its newer weights are non-commercial unless licensed from Prior Labs.
• Hardware: HW1, a CPU server is sufficient. A small LLM (e.g. Mistral Small or Qwen3 8B on HW2) is needed only if letters are drafted locally.
• Ops effort: moderate. Monthly retraining, drift monitoring and a clean ERP data extract are the real work.

Path D — Custom training

Worth it for banks, BNPL providers, utilities and large shared-service centres with hundreds of thousands of accounts. Here, proprietary treatment-outcome data (who was contacted, how, and what happened) is the main asset. You need 12–24 months of contact and payment history, logged randomisation or holdouts for unbiased uplift estimates, and model-risk governance (validation, calibration, fairness). Reinforcement learning for multi-step treatment policies is still rare in production and should be gated by offline policy evaluation.

State of the art (2026)

• Predictive AR is mainstream in large suites. The shift in 2025–2026 is from scores to agentic order-to-cash: vendors bundle prediction with agents that send the right reminder, read replies and log promises to pay. HighRadius advertises 180+ AI agents [1], and Upflow ships collections and insights agents plus an MCP server for querying AR from Claude and ChatGPT [3].
• Techniques: gradient boosting remains best on tabular payment data. Survival and cure models handle censored open items. Uplift and bandit methods distinguish "risky" from "persuadable". Tabular foundation models (TabPFN line) reduce data needs for SMB-size ledgers [7].
• Consumer side: self-cure models that suppress unnecessary contact, digital self-service payment plans, and vulnerability detection. LLM classifiers flag hardship mentions in messages for human follow-up.
• Limits and failure modes: models learn from past treatments, so outcomes are confounded unless holdouts exist. Disputes and invoice errors masquerade as late payment and should be fixed upstream. Large one-off customers dominate amounts but have few observations. Macro shocks (rate rises, energy-price spikes) break calibration quickly. Scores trained on dunning history can encode socio-economic bias in consumer portfolios.

Shared resources & combinations

• Same tabular-ML stack (tabular-ml, HW1): PRED-003 churn, PRED-004 CLV, PRED-009 credit scoring (origination vs. this account-management stage), PRED-005 lead scoring. One feature store and MLOps pipeline can serve all of them, and the customer payment-behaviour features are reused in PRED-009 and PRED-003.
• Downstream: PRED-016 cash-flow & liquidity forecasting consumes invoice-level expected-inflow curves. DATA-002 report narratives cover DSO dashboards. OPT-003 / PRED-011 plan collector and call-centre staffing from forecast workload.
• Execution channels: AGENT-023 order-to-cash agent (cash application, dunning, disputes), CONV-008 outbound voice agent for reminders, REC-006 send-time and channel optimisation, LANG-002 letter and email drafting, AGENT-001 inbox triage for customer replies, SPEECH-002 call analytics for collector QA and hardship detection.
• Upstream data quality: AGENT-006 reconciliation and LANG-033 bank-transaction categorisation give clean payment records. REC-003 next-best-action shares the uplift and bandit machinery.

Effort, cost & KPIs

• Path A: 4–12 weeks incl. ERP connector. Licences are usually per user, per open item or per customer account. Enterprise O2C suites are a six-figure annual commitment, while SMB AR tools cost a few hundred euros per month (check current vendor price lists).
• Path B/C: 2–4 months for a first model plus ERP write-back by one data scientist and one ERP/integration engineer. Compute cost is negligible because data preparation dominates.
• Path D: 6–12 months with model-risk governance in regulated lenders.
• KPIs: DSO and best-possible DSO, overdue % by bucket (CEI, collection effectiveness index), bad-debt write-offs, cash collected per collector hour, roll and cure rates, self-cure rate of uncontacted accounts, payment-date MAE (days), calibration of P(late), uplift vs. holdout, complaint rate, share of vulnerable customers correctly routed to specialist teams.

People & change

Start with management and the head of finance, who own the goal (lower DSO) and the decision on how customers are treated. Involve IT early for the ERP export, and the data owners in accounting for master data. The data protection officer must clear the legal basis and the AVV with any vendor, because payment behaviour of individuals is personal data and automated dunning decisions touch Art. 22 GDPR; a DPIA is likely for consumer debtors at scale.

Works council: depends. Scoring customers alone does not trigger co-determination, but worklists that log collector activity or measure their performance can fall under §87(1) no. 6 BetrVG, and §90(1) no. 3 requires timely information about AI-supported work processes. Involve the council before the pilot; in public bodies the Personalrat applies.

Pilot with a small collections team and one customer segment, and run the model in shadow mode next to the existing process. Typical concerns are "the tool will judge my customers" and "it replaces me". Address them by showing the reasons behind each score, keeping dunning-level and payment-plan decisions with humans, and letting collectors override and give feedback. Support AI literacy with a short training for the team (AI Act Art. 4). See GUIDE-004, GUIDE-005 and GUIDE-008.

Risks, regulation & compliance

• EU AI Act: depends. B2B receivables prioritisation on companies is minimal risk (Art. 4 AI literacy applies). Annex III point 5(b) makes AI that evaluates the creditworthiness of natural persons or establishes their credit score high-risk [9]. A consumer collections model is ambiguous. If it only orders contacts, many lawyers see it outside 5(b). If its output feeds decisions on credit limits, account termination, forbearance eligibility or credit-bureau reporting, treat it as high-risk (risk management, data governance, logging, human oversight, FRIA for public bodies and banks). Annex III 5(a) applies when public authorities use AI to reclaim public-assistance benefits [9]. No Art. 5 prohibition applies in normal use, but scoring that draws on unrelated social behaviour would approach the social-scoring ban.
• GDPR: debtors who are natural persons (consumers, sole traders, contact persons) are data subjects. Use legitimate interest or contract as the legal basis, inform them about profiling (Art. 13/14) and handle objections. Under the CJEU SCHUFA judgment (C-634/21, 2023), a score is an automated decision under Art. 22 when it plays a decisive role in a decision with legal or similar effect. Agency hand-over or bureau reporting based purely on a score is therefore risky. Keep humans in the loop and provide reasons. Germany: § 31 BDSG for scoring, and the RDG registration and information duties for Inkasso.
• Consumer credit & servicing: CCD2 Art. 35 requires creditors to exercise reasonable forbearance before enforcement and to limit charges on default. The Credit Servicers Directive / KrZwMG regulates servicer conduct and fair treatment of borrowers [6]. EBA guidelines on arrears and forbearance apply to banks. The UK FCA Consumer Duty and vulnerable-customer guidance are a benchmark for EU vulnerability policies.
• Commercial law: the Late Payment Directive 2011/7/EU (under revision by the Commission [10]) and default rules (e.g. § 286 BGB) define when dunning fees and interest are allowed.
• Risks & mitigations: aggressive contact patterns against vulnerable people need contact caps, hardship detection and a specialist team. Bias needs outcome monitoring by region and age, with proxies removed. Disputes treated as debts need a dispute flag that blocks dunning. Feedback-loop bias needs permanent random holdouts.

Getting started

1. Extract 12–24 months of AR history (open and cleared items, dunning runs, disputes, contacts) and compute baseline DSO, CEI and the share of invoices paid late per customer segment.
2. Start with your ERP's native prediction or one AR tool (Path A) on a test segment. Alternatively, build a LightGBM payment-date model and compare MAE against "average days late per customer".
3. Run a controlled pilot: half the accounts on the AI worklist, half on the old age-bucket process, with a small no-contact holdout to measure self-cure.
4. Define policy guardrails: dispute blocks, contact limits, vulnerability routing, human approval for agency hand-over and bureau reporting. Document them for GDPR and, for consumers, AI Act purposes.
5. Feed the invoice-level inflow predictions into the cash-flow forecast (PRED-016) and retrain monthly.

Sources

1. HighRadius — Collections software (AI features, Predict Payment Date, customer results) — https://www.highradius.com/lp/autonomous-receivables/collections-management-software/
2. InDebted — Receive (formerly receeve) first-party collections platform — https://www.indebted.co/solution/receeve/
3. Upflow — AR & collections platform, agents, MCP server, customer figures — https://upflow.io/
4. collect.AI (Aareal Bank Group) — AI dunning & payment probability assistant — https://www.collect.ai/
5. SAP — Predict the risk of late payment for invoices (use case page) — https://www.sap.com/uk/use-cases/predict-risk-of-late-payment-for-invoices
6. BaFin — Kreditzweitmarkt (Kreditzweitmarktgesetz, credit servicers) — https://www.bafin.de/DE/Aufsicht/Kreditzweitmarkt/kreditzweitmarkt_node.html
7. Prior Labs — TabPFN repository (versions, limits, licences) — https://github.com/PriorLabs/TabPFN
8. lifelines — survival analysis in Python — https://lifelines.readthedocs.io/en/latest/
9. EU AI Act, Annex III (high-risk use cases, point 5) — https://artificialintelligenceact.eu/annex/3/
10. European Commission — Late payment (revision of Directive 2011/7/EU) — https://single-market-economy.ec.europa.eu/smes/sme-strategy/late-payment_en
11. CJEU press release on C-634/21 SCHUFA (GDPR and scoring) — https://curia.europa.eu/jcms/upload/docs/application/pdf/2023-12/cp230186en.pdf
12. Directive (EU) 2023/2225 (Consumer Credit Directive, CCD2) — https://eur-lex.europa.eu/eli/dir/2023/2225/oj
13. Directive (EU) 2021/2167 (Credit Servicers Directive) — https://eur-lex.europa.eu/eli/dir/2021/2167/oj