Automated Bank ImportAI rule recognition for financial reconciliation
Engineered AI rule pattern recognition and predictive late-payment risk classification — 8+ features across 6 releases, contributing $2M+ in product revenue growth.
The problem
Every bank, every format, every client meant another set of manual mapping rules. Finance teams spent days per cycle on statement ingestion and exception handling, and each new client onboarding carried a rule-building tax that throttled growth.
Discovery
Onboarding time was the hidden bottleneck in the sales funnel. Analysis of implementation logs showed rule configuration was the single longest step, and exceptions clustered in a small number of recurring patterns.
Customer research
- Interviews with finance controllers, implementation consultants and support
- Analysed thousands of historical statements to quantify format variance
- Pain point: exception queues with no prioritisation — everything looked equally urgent
- Insight: users trusted automation far more when it explained why it matched a rule
Product strategy
Learn the rules instead of writing them. Use pattern recognition on historical mappings to propose rules at onboarding, and layer predictive risk classification so analysts work the riskiest exceptions first.
Prioritisation & tradeoffs
- RICE-scored against onboarding time saved and deals unblocked
- Chose rule suggestion (high confidence, reversible) before full auto-posting (high blast radius)
- Explicitly deprioritised exotic bank formats until the long tail justified it
Solution
- AI rule pattern identification: automatically recognises and reconciles bank payments within customer statements
- AI-powered customer risk assessment predicting late payments in accounts receivable
- Intelligent batch processing using ML to optimise CPU utilisation at peak financial workloads
- One-click accept/override with learning from every correction
- Reconciliation dashboard showing automation rate and exception ageing
Architecture
- Statement ingestion → normalisation → feature extraction → classification model
- Rule proposal service with confidence gating; below threshold routes to human
- Feedback loop: every override becomes labelled training signal
- Immutable audit log for every automated posting decision
Execution
- Partnered with design-partner clients for shadow-mode validation before go-live
- Ran automation in parallel with the manual process for a full close cycle
- Compliance and audit sign-off built into the launch checklist
- Supported global presales across Europe, North and South America
Lessons learned
- Shadow mode is the cheapest trust-building mechanism in enterprise AI.
- Ranking exceptions was worth more to users than eliminating a few more of them.
- Next time I'd instrument override reasons from day one — the taxonomy of 'why the AI was wrong' is gold.