A client’s legal structure sets the pace of Deutsche Bank’s credit decisions

Marie-Jeanne Deverdun, who joined Deutsche Bank’s management board in May as Chief Technology, Data and Innovation Officer, gave a concrete time figure in an interview published by Reuters on 26 October. Artificial intelligence, she said, had cut some credit decisions for smaller corporate clients from weeks to minutes, and some decisions for multinational groups from months to weeks, with the project live since the summer. TheStreet read it as a productivity number and ran its piece under the headline “Deutsche Bank just put a hard number on AI productivity”.

The reason Marie-Jeanne Deverdun herself gives points somewhere other than computing power. The minutes are reserved, in her words, for corporate clients “which are less complicated in their setup and their legal structure”. What it apparently measures above all is which clients the bank has in mind for a fast track. Where that track starts, how large the loans on it are and whether a person takes the final decision cannot be read from the bank’s public statements on the credit project. Nor does it have to: the European Banking Authority (EBA) guidelines on loan origination call for an internal definition, not a published one. For competitors and supervisors, the real news therefore lies in that boundary. What follows draws on the bank’s public statements and on the applicable rules. Deutsche Bank’s internal processes are not public.

Three claims by the bank concern different spans

The minutes are the third time claim Deutsche Bank has made about its credit process this year. Its prepared remarks for the first quarter, published on 29 April, said the bank was using AI “to significantly accelerate the credit process in the Corporate Bank”, without a figure. Christian Sewing supplied one at the annual general meeting (AGM) on 28 May. In the Mittelstand, Germany’s mid-sized companies, the chief executive said, the bank was cutting the time from first contact to paying out a loan by up to 95 per cent. In October, Marie-Jeanne Deverdun talked about the decision rather than the payout, and, according to Reuters, about weeks becoming minutes for some smaller corporate lending.

The three claims concern different spans of the same process. The first gives no figure at all. The second measures the elapsed time from a client’s first enquiry to the money arriving in the account. The third measures the time to a credit decision, and Reuters does not say where that span begins. It is tempting to read the numbers as stages of a programme announced in the spring and delivered in the summer. The two figures, however, cannot be converted into each other, and the Q1 remarks give none. Going from three weeks to ten minutes would be a fall of more than 99 per cent on an illustrative calculation, whereas the AGM figure is up to 95 per cent for a longer span in the Mittelstand segment. Both numbers can be true at once. Neither says anything about the other.

For anyone benchmarking their own bank, this dictates where to start. Before asking how fast a credit process is, a bank needs to decide which span it is timing: the decision, the commitment or the payout. Without that step, a head of corporate banking ends up comparing an in-house elapsed time with someone else’s decision time, and reads a difference in measurement as a gap in performance.

Fast decisions for small business clients have existed for years

The strongest objection to the excitement over minutes is that they are nothing new in small business lending. Germany’s savings banks rate borrowers up to EUR 750,000 largely automatically with their KundenKompaktRating. That is a rating rather than a credit commitment, but it takes a laborious step out of the process. In March 2018, according to IT-Finanzmagazin, Commerzbank reported that processing its online working-capital loan, which had previously taken six to eight days on average, could in the best case be completed within 24 hours. Deutsche Bank itself offers business clients a loan of up to EUR 250,000 that can be applied for directly online, with payout possible within a few days. Automated assessment sped up small loans long before language models were being discussed in credit processes.

The October disclosure does differ from these products in two respects. It comes from a significant institution supervised directly by the European Central Bank (ECB), whose corporate business, according to its annual report, runs from small business clients through the Mittelstand to multinational groups. And by Marie-Jeanne Deverdun’s own account, the bank sets the boundary of its fast track by the client’s structure, whereas online loans set it by amount. An amount is stated on the application. Whether a legal structure is simple is not stated anywhere. The bank has to establish it.

The EBA guidelines leave it to the bank to draw the line

As a significant institution, Deutsche Bank is not subject to the German minimum requirements for risk management (MaRisk), which have excluded significant institutions since the ninth amendment. What applies instead is guideline EBA/GL/2020/06 on loan origination and monitoring, which the ECB has said it complies with for the banks it supervises. The guidelines have applied since 30 June 2021, so the rule itself is not new. They cover automated models used in creditworthiness assessment and in credit decisions, including models that only prepare a decision. Their use should be “appropriate to the size, nature and complexity of the credit facility and the types of borrowers” (paragraph 41). Sections 4.3.3 and 4.3.4 require a bank to understand its model, ensure the quality of input data, be able to trace results, and have mechanisms in place for overrides and escalation.

Once a model takes the decision itself, the guidelines become more specific. The credit risk policy then has to define the “products, segments and limits for which automated decision-making is allowed” (paragraph 38(g)), and the decision-making framework has to set out the powers and limitations “of any automated models for credit decision-making purposes” (paragraph 65). That is precisely the line competitors want to know about. The guidelines require it as an internal definition that supervisors can examine. They do not require it to be disclosed to clients or the market. That Deutsche Bank keeps the line to itself is therefore an observation, not a failing.

How far banks use this latitude can only be read from self-reported information. In workshops held in 2025, thirteen significant institutions told the ECB that AI models in credit scoring supported decisions taken by people, “except for small retail loan originations, which are sometimes automated”. The ECB published the results in November 2025, stating explicitly that they were not a supervisory assessment. Whether Deutsche Bank was among the thirteen is not known. If its minutes are automated decisions, the open question is whether they go beyond the “small retail loan originations” those institutions described in 2025. If they are decisions taken by a person and prepared by AI, the minutes are saved at the credit analyst’s desk. Reuters leaves both readings open. Christian Sewing did tell the AGM that the decision stays with people, but he said so in the passage on controls, about credit and compliance checks, not about the credit project. An earlier article on AI use at significant institutions looks at how the ECB frames the human role in these processes more broadly.

A simple structure only emerges from a check

Anyone who has managed credit lines for trading counterparties knows that the hard part is rarely the limit itself but the question of who counts as part of the counterparty. If legal structure is the criterion, the bank moves exactly this question to the start of the credit process. Since 31 March 2026, section 18 KWG has allowed a bank to grant a loan exceeding “1 500 000 Euro oder 10 Prozent seines Kernkapitals” (EUR 1.5 million or 10 per cent of its core capital) only if the borrower discloses its financial circumstances. The threshold used to be EUR 750,000. Below it, section 18 KWG requires no disclosure, and the question of structure arises from group formation, anti-money laundering checks and the creditworthiness assessment under the EBA guidelines. The borrower is not necessarily the individual company, because section 19(3) KWG treats a group of connected clients as a single borrower. Which companies belong to such a group is set by Article 4(1)(39) CRR, supplemented since 2024 by Delegated Regulation (EU) 2024/1728, with criteria for control and eleven cases of economic dependency.

The EBA guidelines make the same point in their treatment of the creditworthiness assessment. For medium-sized and large enterprises, they explicitly require banks to “analyse the organisational structure, business model and strategy of the borrower” (paragraph 144(b)), while paragraph 121(b) names only the business model and strategy for micro and small enterprises. Where a borrower belongs to a group, banks should collect information on the connected clients, “especially when repayment is reliant on cash flow emanating from other connected parties in the same group” (paragraph 93), and the group-level assessment applies in substance to micro and small enterprises as well (paragraph 124). The guidelines follow the EU definition of small and medium-sized enterprises, under which much of Germany’s Mittelstand probably counts as medium-sized and so falls under the stricter requirements. What Reuters means by “smaller corporates” cannot be mapped onto either class.

Anti-money laundering checks come on top. The German Anti-Money Laundering Act requires a bank to establish a counterparty’s ownership and control structure at the start of the business relationship, not with every credit decision. From 10 July 2027 the EU Anti-Money Laundering Regulation (AML Regulation), Regulation (EU) 2024/1624, applies, and it counts indirect holdings along every chain. Shares are multiplied “along the chain” and added up across all chains, and “all shareholdings on every level of ownership” count. German law already covers indirect holdings. What is new is the explicit calculation rule. In all likelihood, little changes for a company with a single shareholder. For a holding structure with three tiers, the work grows before the bank has read a single balance sheet.

Classifying a structure as simple is therefore a check in its own right and, on one reading, the one that decides who reaches the fast track. The Greensill case shows how opaque a group can be. On 11 September 2020, BaFin ordered a special audit of Greensill Bank. According to the German government, it cited the complexity of the Greensill Group’s business model and the lack of transparency in GFG Alliance’s financial circumstances.

In consumer lending, legislators draw the line

In retail banking, the answer to who draws the line for automated credit decisions is different: legislators have drawn it, and in several places. Article 22 of the General Data Protection Regulation (GDPR) gives a data subject the right not to be subject to a decision based solely on automated processing that produces legal effects concerning them. From 20 November 2026, section 30(6) of the German Federal Data Protection Act (BDSG) allows a borrower under a general consumer credit agreement to request human intervention where the creditworthiness assessment involves automated processing. The EU AI Act classifies creditworthiness assessment of natural persons as a high-risk use. Earlier articles cover the details, on the Singapore bank DBS and its rollout of credit agents and on the AI case studies published by the UK Financial Conduct Authority (FCA).

Corporate lending has no such statutory line. Article 22 GDPR protects natural persons, which a limited company is not, and the EU AI Act does not cover the creditworthiness of legal persons. Here the line is drawn by the bank, in its credit risk policy and its decision-making framework, and supervisors check whether it has done so in a way they can follow. This need not be a flaw, since a company deals with its bank differently from a consumer and usually has its own advisers. It does shift the burden, though. Where the law asks the questions in retail banking, a bank in corporate lending has to ask them itself and keep the answers ready for inspection. For sole traders and freelancers, who fall under Article 22 GDPR as natural persons and whose treatment under the EU AI Act is unresolved, the line runs right through small business lending, wherever a fast track also covers sole traders or freelancers.

Client-side agents raise the question of who is authorised to act

In the same interview, Marie-Jeanne Deverdun said the bank was preparing for “an agentic economy where our agents interact with the agents of our clients with clear accountability and the right controls”. According to her, the bank’s own agents are managed through an internal platform called ADA, which develops, monitors and controls agents and allows the bank to switch between models. Reuters does not say whether ADA also runs the credit project. In the United States, signals currently point the other way. According to an Axios report published on 22 October, large banks are restricting access by third-party AI agents to client accounts, with Morgan Stanley blocking them and others moving towards read-only access.

On one reading, this has a practical consequence for corporate lending. If a client’s agent submits the documents in future, the question of who owns the company is joined by two more: on whose behalf the agent acts, and with what authority. The fast track then requires not only a simple structure but also authority that can be evidenced. The article on the UK Treasury consultation looks at how consent, authentication and liability might be handled for payments made by agents.

Recommendations for corporate banking boards, credit risk and compliance

Glossary

Christian Schablitzki – the agentic banker

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 a Partner at Infosys Consulting in Germany. Certified as Claude Partner – Claude Code (Anthropic), as well as in Google AI, Generative AI Leader (Google Cloud) and IBM RAG and Agentic AI.

LinkedIn profile → · agenticbanker.ai →