Agentic Credit at DBS Is Live... Supervisory Technology Is Advancing at One-Fifth the Pace of the Models It Has to Supervise, and This Week Every Inst

Welcome to Agentic Newsletter for Financial Services, the first newsletter dedicated to the transformation in financial services driven by the advancements in AI, shaping the Agentic economy and crafted by a team of more than half a dozen agents that keep evolving and adapting.

These Agents scan the world on a daily basis, filtering hundreds of news articles. We work together to bring you a weekly curated selection coupled with my human insights.

Thanks for reading and being part of our fast-growing subscription community. This is edition No. 59, past 40,000 subscribers. Let's stay up to date and build a more human world powered by agentic AI.

This week, the agentic AI story split cleanly into two camps: institutions proving out scale, and regulators moving to keep pace with it.

DBS disclosed one of the more substantial multi-agent production deployments in banking, 80 agents preparing credit approvals for its largest corporate clients, while keeping a human checkpoint on every output. Chief data and transformation officer Nimish Panchmatia put a number on the governance gap driving that caution: supervisory technology is advancing at roughly one-fifth the pace of the models it has to supervise. Across the Caucasus, ABB Bank in Azerbaijan showed what happens when fewer legacy constraints meet AI ambition, more than 180 models across 86 use cases, and it is now buying NVIDIA HGX B200 infrastructure to train its own local models rather than stay dependent on vendors.

On the infrastructure front, zerohash became another major infrastructure provider shipping an agentic payments product and joining the x402 Foundation, betting its Know Your Agent (KYA) screening and usage-metered settlement can become the default compliance layer. Bloomberg moved to control the plumbing behind private-markets data with its acquisition of Canoe Intelligence, a bet that the next moat in finance is data normalization. And a Clearwater Analytics survey put a hard number on a problem most firms are not solving fast enough: a 23-point gap between how executives rate their data's completeness and its accuracy, even as 95% raise AI budgets.

On regulation, Germany's BaFin gained real enforcement power this week under the KI-MIG law, though its own director-general was careful to frame the mandate as sample-based monitoring, not wall-to-wall supervision. And a new research paper offers financial institutions a concrete answer to Satya Nadella's warning about vendor lock-in: freeze the model, keep the institutional memory external and portable.

This newsletter is a human-AI collaboration involving seven AI agents orchestrated by Nicolas C. and my own curation and refinement.

1. Deployment & Use Cases

DBS Chooses Caution: Agents Can Wait Until Controls Catch Up

DBS has orchestrated as many as 80 AI agents to prepare credit approvals for its largest corporate clients, a live production workflow, not a pilot, while still requiring a human to check every agent's work before it acts. Chief data and transformation officer Nimish Panchmatia says supervisory technology is advancing at roughly one-fifth the pace of the underlying models. Personal and team agents, like Microsoft Copilot, stay walled off from production entirely; enterprise agents that touch production, including the 80-agent credit workflow, are built in-house under standard software development lifecycle controls. Watch whether supervisory technology closes that gap before more of the fleet moves off the human checkpoint.

  Emerging-Market Banks Embed AI Into Core Systems, ABB Bank as an Example

ABB Bank in Azerbaijan has built more than 180 AI models across 86 use cases, reporting AZN 50 million in value generated over three years and roughly 90% paperless operations, while pairing automation with human-in-the-loop controls rather than removing oversight. The bank is now acquiring Dell servers built on NVIDIA's HGX B200 platform to train local large language models, moving from consuming AI tools to building its own. 

The Caucasus, Central Asia and MENA region carries fewer legacy constraints than mature markets, letting banks like ABB modernize faster. The AZN 50 million figure is self-reported and unaudited, treat it as a claim.

  2. Agentic Commerce, Rails & Infrastructure

On-Chain Rails: zerohash Launches Agentic Finance Suite & Joins the x402 Foundation

zerohash launched its Agentic Finance Suite, giving payments companies and card networks the infrastructure to let AI agents hold, move and stream money directly, built on Know Your Agent screening across 40+ protocols and 100+ assets. The suite's Payments Streaming Engine lets an agent sign one spending-capped authorization while zerohash meters usage and settles in increments as a session runs. 

On the infrastructure front, just zerohash became another major infrastructure provider shipping an agentic payments product and joining the x402 Foundation, betting its Know Your Agent (KYA) screening and usage-metered settlement can become the default compliance layer. It joins Fireblocks, Mastercard, and Stripe, all of which shipped comparable agent-payments infrastructure earlier this year, so the real question is no longer whether this space consolidates but whose compliance layer wins out.

  3. Market Impact, Funding & Innovation

Bloomberg Buys Canoe Intelligence to Fix Private Markets' Data Fragmentation

Bloomberg has agreed to acquire Canoe Intelligence , folding Canoe's AI-driven document processing for private markets into Bloomberg's portfolio and analytics infrastructure; deal terms were not disclosed. Canoe processes more than 1.5 million documents monthly across 44,000 funds for 500+ institutional clients representing over $11 trillion in assets under service. 

The intention is to enrich ASKB, Bloomberg's agentic AI conversational interface, which will surface Canoe's normalized private-markets data alongside public-market data so investors can query both through natural language. Data infrastructure has quietly become where the real competitive moats in finance are built, and Bloomberg is betting the next moat is feeding a clean, unified dataset into its own agent. 

   The AI ROI Gap Asset Managers Can't Ignore

A Clearwater Analytics study of 178 senior fund executives found 95% raised AI budgets in the past year and 85% plan to raise them by at least 50% more over the next 12 months. But a 23-point gap has opened between how firms rate the completeness of their data, 79%, versus its accuracy, 56%, meaning the same executives funding AI expansion are unsure whether their data can support it. 

Clearwater is itself an agentic investment-management platform vendor, so the study's framing carries a commercial angle even though the survey was independently fielded. Budget growth will not fix a broken data pipeline; the 23-point gap, not the 95% spending figure, is the number a board should be asking about.

4. Regulation & Governance

BaFin's New Mandate: Policing AI in Banking and Insurance to Protect Rights and Ensure Fairness

Germany's BaFin gained expanded authority over AI use at banks and insurers when the KI-MIG law took effect on July 29, operationalizing EU AI Act enforcement inside German financial services. From August 2, any AI-driven customer interface must clearly identify itself as non-human under Article 50 transparency rules. 

BaFin can order fines, but director-general Jens Obermöller was explicit that the mandate is sample-based monitoring, not an exhaustive audit regime; full enforcement against high-risk AI, meaning credit-scoring and insurance risk-pricing, begins in December 2027. Sample-based monitoring still means real exposure; institutions using AI in credit-scoring or chatbots should not assume low audit probability equals low risk.

 5. Research

Frozen Models Learn On the Job: Distilling Deployment Feedback Into Retrievable Rules

A new paper from Valentin Tablan, Scott Taylor and Kristoffer Bernhem shows frozen-weight AI agents can still improve continuously by pairing them with an external memory that distills each interaction into retrievable natural-language rules. Tested on a banking benchmark using only one-bit success/failure feedback, the approach raised single-trial success to 1.6x a baseline using verdicts alone, and 2.6x using corrections. 

Microsoft's Satya Nadella made a related argument this month in his 'Reverse Information Paradox': keep the harness, context and memory separate from the model so any one model can go away without taking accumulated enterprise learning with it. The authors' cross-model transfer result, memory built under Mistral Large improving performance when read by Claude Sonnet 5, shows this kind of institutional learning can survive a model swap.