Most companies are bolting AI onto a business built for a world without it. That’s why the spend never shows up in the results.

Installing a full detection pipeline end-to-end: graph construction over the transaction population, community detection to find natural groupings of accounts, and a sequence model trained to recognise the behavioural signatures of twelve criminal typologies — structuring, layered chains, mule networks, trade-based laundering, professional money laundering and others. Hard Concrete gates to control what the model attends to. Low-rank adaptation heads so the model can be steered to new typologies without full retraining. A predicate router that maps model output to investigator-ready case narratives. Everything built against the bank's own data, not a benchmark.

On handover a detection capability the financial crime team operates directly, with explainable outputs an investigator can read and a regulator can examine. Period one covered nearly a hundred thousand accounts, surfaced over six thousand reportable entities, and produced a separation ratio that makes the prioritisation meaningful rather than nominal. The team owns the pipeline, the weights, and the architecture. No licence, no dependency, no ongoing seat.

Installing a focused four-week assessment across the operational and control teams — risk, compliance and customer operations — reading the current processes and constraints, running a working session with the people who own them, and testing each opportunity for value, feasibility and risk.

On handover a prioritised set of practical AI opportunities, with a clear recommendation on sequencing, governance and next steps — built to lift consistency and free capacity, not just cut cost, and grounded in the bank's own priorities.

Installing a five-week pilot that takes one operations use case — intelligent document processing and reconciliation — end-to-end against the bank's real workflows and volumes, built to run on live cases, not demo on slides.

On handover a working capability the operations team runs and owns, with outcomes measured against the value hypothesis and a clear, evidence-based read on whether to scale into a longer-term engagement.

Before the everyday-banking AI question sat as scattered point ideas, with no shared basis for deciding what to build, in what order, or how it fit an ai-native operating model.

Now a prioritised portfolio of ai-native propositions — each defined, pressure-tested against strategy, regulation, and the competitive field, with a build case and sequencing — owned by the bank and actionable without us.