Sentient banking stalled by messy transaction data

Banks hold the largest consolidated record of consumer financial behaviour, covering card payments, transfers, salary flows and recurring obligations. Product teams say apps remain generic because transaction feeds lack the structure machines need to infer behaviour reliably.

The term sentient design was coined in 2024 by designer Josh Clark. Clark and Veronika Kindred published the book Sentient Design: Crafting Intelligent Interfaces with AI in June 2026. Design strategist Adam Dragus of Ergomania has discussed the idea with multiple European bank product teams; the concept reframes users as curators who set limits and preferences while systems act on context.

Most banks today run intent-based features: customers request breakdowns, forecasts or chatbot replies and the system returns an answer. Sentient inference operates earlier. It monitors patterns across a customer ledger, detects deviations from a personal baseline and draws conclusions without an explicit prompt. Practitioners emphasise that inference depends on a stable baseline built from high-quality transaction enrichment.

Teams point to consumer apps outside banking as reference points. Apple Shortcuts learns routines from behaviour and Waze reroutes commutes before a user opens the app. Designers recommend starting with small, low-stakes actions that do not move money, to build user trust before expanding automation. Adam Dragus urged a gradual approach: “teach the young lion — one thing at a time — earn trust, then extend.”

Regulatory and user-trust constraints shape any deployment. Practitioners list three requirements that must be met before automation takes consequential actions: informed consent so customers approve actions that affect them; explainability so customers can see the data and reasons behind a suggestion; and graduated control so customers set how much authority the system has, from narrow supervision to broader delegation.

Design teams describe an “escalator” pattern: automation that always has a manual fallback. One implementation is an opt-in, sandboxed pocket of funds the customer grants to the AI under explicit rules. The pocket is reversible and visible in the app so customers can withdraw authority or review actions.

Engineers and product leads say those behaviours rely on an enrichment layer. That layer must provide clean merchant identity so a single retailer does not appear as many vendors, accurate categorisation so changes in spending register as meaningful signals, deduplication across acquirer IDs and points of sale, and reliable detection of recurring payments so fixed commitments are not mistaken for discretionary spending. Without those elements, teams report that models produce false signals about spending patterns and risk.

Ergomania recommends an AI-readiness assessment that maps whether an institution has working AI features, who owns them internally, and what an achievable starting point is given regulatory and product maturity. The agency also maps “sentient flows” to convert the concept into implementation roadmaps. Banks contacted by Ergomania vary in approach: some plan modest pilots while others seek broader automation after data cleanup.

Practitioners advise a sequence of work: choose the first behaviour for the system to learn, set consent and explainability rules, design manual fallbacks, and prepare the transaction feed so machines can read it. Banks and vendors say progress on sentient features will depend on improvements to merchant identity, categorisation, deduplication and recurring-payment detection before wider automation is enabled.