Banks' AI race is being won in the back office

JPMorgan says AI has more than doubled the volume of transactions its screening operation handles while cutting manual checks in half. The bank's most recent disclosure does not get the same press as a chatbot launch, but it is the closest thing to a hard number on where bank AI is actually paying off.

A 24 August 2026 LiveMint column by Manish Agrawal uses that result to draw a three-layer picture of banking AI. Layer one is what the customer sees: chatbots, in-app assistants, product recommendations. Layer two is the decisioning layer: fraud detection, credit underwriting, transaction screening. Layer three is the back office: liquidity forecasting, deposit modelling, cash-flow analysis, compliance. Public attention drops off sharply as you move inward. The Agrawal column argues that the importance of each layer does not drop off with the attention. The most durable AI advantage in banking is accumulating in the layers the annual report never photographs.

Fraud detection is the layer where defensive AI and offensive AI are the same AI. Phishing kits, synthetic identity builders, and deepfake voice tooling all run on commodity models, and the lead a bank's detector earns is matched, sometimes inside a quarter, by an attacker's gain on the other side. The LiveMint framing calls it an arms race, which is the right word. Catch-up is the default state, and any quarter of detector advantage is rented, not owned.

Credit decisioning sits in the same layer with a different failure mode. Models trained on cash-flow, income, and spending signals can underwrite borrowers the static bureau score would reject, a real expansion of access for thin-file customers. They can also exclude borrowers more aggressively, with rationales the underwriter cannot reconstruct in a dispute call. The durable AI lead in credit goes to the bank that can show its model is both wider and fairer, not just wider.

The back-office layer is the one the Agrawal column argues will decide who wins, and it is also the one with the least public reporting. AI applied to deposit modelling, liquidity forecasting, and cash-flow analysis does not need to run the balance sheet to matter; small improvements in the forecast shape the funding decisions that sit behind a bank's wholesale book. For now, the layer-three thesis rests on directional argument and one columnist's reading of the incentive. The BCG 2026 Treasury Benchmarking report, expected in this window, is the most likely place to find the first cross-bank numbers.

Chatbot launches, image-generation features, and in-app assistants are marketing surface. The numbers worth asking for are screening throughput, alert-review hours saved, default-decision turnaround, and forecast error on liquidity and deposit models. The banks that publish those are the ones taking the race seriously. The ones that do not are showing you the cheapest layer to build.