On 19 August 2026, DBS announced that it had rolled out a system of specialised AI agents for drafting credit memos to some 1,500 corporate relationship managers and credit risk managers worldwide. A pilot with 150 participants preceded it. The bank does not say what that pilot produced.
That absence is the most interesting thing in the announcement. A bank that scales tenfold after a completed pilot is likely to have numbers; publishing them would have cost nothing.
What: several specialised agents handle more than 70 sub-tasks, according to the bank, and produce a review-ready first draft of the credit memo.
Who: around 1,500 relationship managers and credit risk managers, following a pilot with 150 participants. The bank describes the rollout as global.
Target: at least 30 per cent less time. Preparing credit memos accounts for up to 40 per cent of a relationship manager's time, DBS says.
Decision: assessment and the final call remain with the human. Bankers and risk managers work the draft up iteratively.
Not disclosed: technology or model vendor, pilot results, take-up.
A goal, not a measurement
The release contains exactly one passage with percentages. Preparing credit memos and related activities can account for up to 40 per cent of a banker's time, it says, and: “The goal is to reduce time spent by at least 30%.” It is a goal. Neither the release nor the bank's interviews nor the coverage report a measured result from the pilot.
Two things matter for reading it. The base first: if the upper bound of 40 per cent holds and the target is met exactly, that is 12 per cent of total working time. Quoting the 30 per cent without that base suggests two and a half times as much. And the base itself is soft: the 40 per cent is phrased as an upper bound with a modal verb and carries no source, presented neither as measured nor as estimated.
One can read this generously. A bank that states a goal rather than a success is claiming less than the industry norm, in a sector that offers no shortage of unevidenced productivity claims. One can also read it the other way: the pilot is complete, its result remains unpublished, and the only quality judgement in the release is the word “successful”, which is nowhere substantiated. Both readings are tenable. Only the second half is demonstrable: the number is missing.
Rolled out is not in use
The tense of the release is unambiguous and works in the bank's favour. It says the capability has been rolled out, not that it will be. The global reach is stated explicitly. A completed move from pilot to production is therefore documented, and that is more than announcements of this kind usually deliver.
The wording still needs pinning down. Rolled out to 1,500 staff means made available, not used. The release says nothing about take-up, about the number of memos actually drafted or about the intensity of use. The 1,500 are those with access. In a field where the gap between availability and use is the real difficulty, that distinction should not be blurred.
No comparable case
I looked for a case that combined the same three elements: a named number of users, a date and a documented production deployment for agentic credit memo drafting in corporate lending. No publicly documented case of that kind could be found at any of the banks examined, among them Citigroup, HSBC, Standard Chartered, JPMorgan Chase, BBVA, Santander, Commerzbank, ING and Deutsche Bank. That is a statement about the public record, not about what banks are actually doing; internal rollouts without a press release are the norm.
The closest candidate is Citigroup, which unveiled an agent platform in April 2026. The figure of 180,000 staff cited there refers to AI tools generally rather than to agents, no credit memo use case is named, and the bank expressly describes the position as an early stage. The product category is also available off the shelf now: S&P Global, according to its own announcement, launched an agentic credit memo generator in June 2026. That demonstrates market readiness, not production use inside a bank.
One figure now circulating in search results comes with a health warning. A double-digit productivity gain attributed there to several European and American banks traces back to a vendor blog without a primary source, one that names no bank at all. The names appear only in the search summaries. None of it is citable.
Europe protects natural persons, and companies fall outside
For European banks, the obvious question is whether such a system would fall under the high-risk rules of the AI Act. On the wording, the answer is remarkably clear: no, at least not automatically.
Annex III, point 5(b) of the regulation covers AI systems intended for the creditworthiness assessment and credit scoring of natural persons. The recitals say what is being protected: those persons' access to finance and to services such as housing, electricity and telecommunications. This is consumer protection. An agent assessing only corporate creditworthiness is not caught by the wording. That provision was, incidentally, left untouched by the Digital Omnibus amendments in force since 27 July 2026.
A second, independent carve-out applies as well. Under Article 6(3), a system listed in Annex III is not high-risk where it does not materially influence the outcome of decision-making, including where it is intended to perform a preparatory task for an assessment. A system producing a draft that a human evaluates and owns falls almost exactly within that exemption.
Two qualifications attach to that. The first sits in the same paragraph: a system listed in Annex III is always high-risk where it performs profiling of natural persons. Once an assessment does not merely involve guarantors, shareholders or sole traders but profiles them, no exemption applies. The second concerns formalities: a provider taking the view that its system is not high-risk must document that assessment before the system goes into service, and must register it. “Not high-risk” is therefore not a free pass but a self-classification that has to be justified. Transparency obligations and the AI literacy duty remain untouched in any event. And that documentation duty falls on the provider of the system: anyone deploying a third-party product must first establish who holds that role.
Three supervisory regimes, three gaps
The comparison is where it gets interesting. In Singapore, where DBS is supervised, principles on fairness, ethics, accountability and transparency have applied since 2018. A more specific set of guidelines on AI risk management, intended to cover generative systems and expressly AI agents, was put out to consultation by the Monetary Authority of Singapore in November 2025; as far as public sources show, it has not yet appeared in final form. DBS is deploying while the more specific framework is still in draft.
In the United States, the agencies replaced their long-standing model risk management guidance in April 2026. A footnote in the new text states that generative and agentic AI models are novel and fast-moving, and therefore fall outside its scope. Quoting only that clause turns the meaning on its head, however: the following sentence makes clear that institutions should determine appropriate governance and controls for tools not covered, based on their own risk practice.
The European Union is the only one of the three with a written regulation and fixed dates, and the only one whose gap runs along a segment boundary: corporate lending is not caught by the high-risk provision. Obligations for Annex III systems apply from 2 December 2027 following the Digital Omnibus deferral.
Recommendations
Immediately: a 30 per cent saving on a sub-process that takes at most 40 per cent of the time is at most 12 per cent of total time. Before any internal investment decision, have every percentage traced back to its base before it enters a business case. That applies to vendor figures as much as to your own business lines' estimates.
This month: the meaningful measure is not the number of people with access but take-up, and the share of drafts that survive without substantial rework. If you are running a pilot should capture both from the outset. Neither can be reconstructed afterwards.
Before deployment: an agent confined to corporate lending is not caught by the wording of Annex III, point 5(b), and the exemption for preparatory tasks applies in addition. Both fall on the provider and must be documented before the system runs, and the system must be registered. With a third-party product, establish first who holds that role. The test is not whether natural persons appear but whether the system profiles them.
For planning: DBS describes, elsewhere and not in this announcement, a registry in which its agents are identified, traced and evaluated. Whichever tool is chosen, those are the questions a supervisor will ask: which agents are running, who owns them, how their performance is measured and how the decision path is documented. Answers to those questions do not emerge as a by-product of a roll-out.
Glossary
Credit memo: the written paper with which a relationship manager takes a corporate loan to the credit committee. It sets out the borrower’s financial standing, the security, the structure and the covenants, and it carries the recommendation on which the committee decides.
Annex III point 5(b): the AI Act’s high-risk category for creditworthiness assessment and credit scoring. It expressly covers “natural persons”; on the wording, corporate lending does not fall within it.
Article 6(3) AI Act: the second, independent route out of high-risk classification. A system that does not materially influence the outcome of decision-making is not high-risk despite Annex III. Anyone relying on it carries the burden of showing that it does.
SR 11-7: the 2011 US guidance on model risk management, replaced in April 2026. Its successor expressly puts generative and agentic models outside its scope, while requiring firms to determine appropriate controls for them all the same.
Christian Schablitzki
Strategy & Management Consultant · Agentic AI expert for financial institutions
More than 20 years in investment banking and derivatives trading, followed by over 10 years advising financial institutions. Currently Partner at Infosys Consulting in Germany. Certified in Google AI, Generative AI Leader (Google Cloud) and IBM RAG and Agentic AI.
LinkedIn profile →Keep reading – every 14 days in your inbox.
Capital markets insights, regulatory updates and AI trends. Concise, well-founded, free.
GDPR-compliant. Unsubscribe at any time.