LaSpend: read-only AI money assistant over UK Open Banking

Consumers lose money to forgotten subscriptions and recurring charges they never review. The product owner wanted an AI assistant that finds them, scores the waste and helps the user cancel, without ever becoming a payments product. The board owned the risk appetite: useful, but never in a position to move money.

FCA-regulated UK Open Banking data, PII minimisation before anything reaches a model, and a hard rule that the assistant never moves money. Models fail or drift, so the user-facing behaviour had to stay predictable when they do.

A read-only pipeline: Open Banking transactions are normalised and PII-minimised, recurring-spend detection runs deterministically first, and the LLM layer explains, scores waste and drafts the cancellation steps. Every user-triggered action requires explicit human approval and produces a verified action receipt. Deterministic fallback replaces the model output when confidence or availability drops. I specified the flow and the acceptance tests; AI coding agents generated the code under a three-gate review and evaluation process that I ran.

• UK Open Banking API (read-only scopes)
• Stripe subscriptions
• Cloudflare Pages, Functions and D1
• Multi-provider LLM routing with deterministic fallback

Read-only by policy, enforced at the API scope and at the action runtime. Human approval gate for every user-triggered action, verified action receipts stored for audit, PII-minimised prompts, and a deterministic fallback path when models fail. Production incidents escalate to me.

Live in production. Measured by detection precision on recurring spend (checked against labelled statements) and by cancellation completion rate once a user approves an action. Adoption and precision figures are not published here; I share them on a call.