Data Department helps finance companies, insurers, payment companies and fintech lenders use machine learning (ML) and AI (artificial intelligence) for decisions about cash, credit and customers: how much will be collected, which policyholders to call first before renewal and how much credit to extend. Our work includes a collections forecast whose error fell from 32% to 9.5% and an insurance renewal programme with about 6% revenue gain.
Where does AI pay off in financial services and insurance?
Financial decisions are frequent, measurable and backed by good historical data. The value comes from turning risk estimates into an action: a cash plan, a call list, a credit limit.
| Decision | What ML or AI does | Typical data |
|---|---|---|
| How much cash will be collected, and when | Collections forecasting built on customer-level risk scores | Receivables, payment history, customer risk scores |
| Which customers to contact first before renewal | Churn and propensity-to-buy models feeding a priority-optimisation scheme | Policy, renewal, product and contact history |
| Whether a borrower is likely to default | Credit-risk scoring, for example probability of default with logistic regression | Accounting, banking and commerce data |
| What repayment plan a financing option implies | Modules that generate repayment plans from the option's parameters | Financing terms, revenue history |
| Whether to approve a merchant's credit-limit increase | A model that supports the limit decision with a risk-based recommendation | Merchant transaction and repayment history |
| Which transactions or claims look unusual | Anomaly and fraud detection that ranks cases for review | Transactions, claims, customer and device data |
| Where an answer sits in policies and procedures | A RAG (retrieval-augmented generation) assistant that cites its sources | Policy wordings, procedures, internal guidelines |
Relevant case studies
- Collections forecasting: a finance company's treasury. Using customer-level risk scores, forecast error measured as MAPE (mean absolute percentage error) fell from 32% to 9.5%.
- Insurance renewal and retention: an insurance business. A churn predictor plus a priority-optimisation scheme, built on churn and propensity-to-buy models, decided who to call first; the result was about 6% revenue gain.
- Revenue-based financing risk models: a European fintech lending marketplace where SaaS (software-as-a-service) companies finance growth by selling a share of future recurring revenue instead of equity. Embedded with the team for 3–4 months in 2023, we analysed the financing approach, built modules that generate repayment plans from a financing option's parameters, and built credit-risk scoring (probability of default, logistic regression) on accounting, banking and commerce data pulled from the borrower's systems through a financial-data API (application programming interface).
- Dealer credit risk scoring: a payment services company. Models for automatic limit-increase decisions, built around predicting dealer default and scoring dealer risk from limit-request, transaction and chargeback history.
What are typical first projects in finance and insurance?
- Cash and collections forecasting for treasury, measured on forecast error against the current method.
- Renewal and retention prioritisation: a ranked call list for the sales or service team, tested against a control group. See our churn prediction and customer analytics service.
- Credit-risk scoring for a defined product or customer segment, with interpretable models.
- A document assistant for policy, procedure and compliance questions. See our generative AI assistants service.
Each starts with a two-week Discovery: ranked use cases, a data check and a go/no-go decision. The free 30-minute AI opportunity review is a quick way to test fit.
Data and integration notes
- Core systems such as policy administration, loan management and accounting hold most of the history these models need.
- CRM (customer relationship management) and contact-centre data are needed to turn retention models into call lists and to measure the result.
- External financial data can be pulled from a borrower's systems through APIs, with the borrower's consent.
- Infrastructure: we deliver on AWS (Amazon Web Services), Google Cloud, Azure or on-premises, so sensitive data can stay in your environment.
Frequently asked questions
Can the models be explained to a risk committee?
Yes. Where explanation matters, we prefer interpretable methods such as logistic regression for credit-risk scoring, and we document which inputs drive each score.
How do you measure a retention or renewal programme?
On a business measure such as revenue or retained customers, not on model accuracy alone. Where possible, the effect is tested against a control group. In our insurance renewal work, the reported result is about 6% revenue gain.
Does customer data have to leave our systems?
No. We can work on your own infrastructure, including on-premises servers.
What does a collections forecast need?
Receivables and payment history by customer, plus customer-level risk scores. Our treasury project used such scores and reduced MAPE from 32% to 9.5%.