White paper
When agents become the customer
Payment leaders have moved past basic digital infrastructure and on to autonomous payments and real-time transaction intelligence, running on structured ISO 20022 messaging.
Time-series and transaction foundation models (TFMs) could do for payments what large language models did for text: stronger prediction, paired with real demands around explainability, latency and compliance with frameworks such as the EU AI Act and DORA.
This paper shows what it takes to move TFMs from research to production, grounded in deployments at RBC and Revolut.
We have arrived at a critical juncture where we now have profitable AI because we have useful AI. Compute is revenue.
What's inside the paper
- Deconstructing NVIDIA TFM: from static rules to neural payment flows.
- The strategic case: promise versus reality.
- Institutional AI pioneers: deep dives into RBC and Revolut.
- Value creation: data monetization and business outcomes.
- How Thoughtworks can help: from AI hype to production alpha.
- Sibos action plan: a four-step roadmap.
No form fill required.
The proof points
32ms
inference latency at 10,000 transactions per second in our reference TFM pipeline.
65%
more caught fraud cases for Revolut's PRAGMA model, with 17% higher precision.
42%
fewer false-positive fraud alerts, and over 80% less manual feature engineering.
Meet the authors
