Deutsche Bank is developing a trade surveillance system with Google Cloud built on a large language model (LLM). It is intended to detect anomalies in orders, trades and market movements close to real time and escalate them to a human compliance officer for review. The work became public through a Bloomberg report of 25 February 2026.
For institutions running their own market abuse surveillance, the case is instructive on two counts. It shows where a language model actually bites inside a regulated control function, and it demonstrates on a concrete example how quickly metrics change banks as they travel through the coverage.
What: LLM-based trade surveillance to detect anomalies in orders, trades and market movements
Who: Deutsche Bank together with Google Cloud
Status: in active development; the sources speak consistently in the future tense, with no production use evidenced
Second phase: LLM-based monitoring of staff communications, rollout announced for 2026
Sourcing: Bloomberg report of 25 February 2026; neither db.com nor cloud.google.com confirms the account officially
Personnel: Bernd Leukert left the bank at the end of June 2026; Marie-Jeanne Deverdun succeeded him
Whose figures are actually in circulation
Many retellings place two values next to Deutsche Bank's name: up to 40 per cent fewer false positives and savings of up to $5m a year. Both figures belong to Nomura.
They come from the same Bloomberg article, which covers three banks. Nomura is in talks with another, unnamed institution about jointly trained surveillance models and expects that venture to cut false positives by 30 to 40 per cent and save up to $5m annually. This is an expectation for a project that does not yet exist. It is attributed to Tahir Zafar, then international head of AI strategy at Nomura and since moved to JPMorgan.
The only figure attributable to Deutsche Bank by name reads differently: the bank has cut the false positives that trigger deeper probes by more than 25 per cent. That statement sits in the present perfect and therefore reads as an outcome already achieved. It is an assertion by the then chief technology officer rather than an audited result, with neither a methodology nor a review report published.
What the system is meant to do
The professionally interesting part is the division of labour. The language model performs the analysis and proposes a route; the decision stays with the human. Bernd Leukert described it as follows: “The LLM can do the analysis and help recommend the route which the compliance officer can validate and then close the alert. The ultimate decision stays with the compliance officer.”
The work therefore targets the most expensive problem in trade surveillance. Rules-based systems generate false positives in volume, and each one consumes handling time in compliance. A model that pre-sorts alerts and supplies a recommendation attacks exactly that, while leaving decision rights where they are.
On the scale of the existing set-up the report gives two figures: the current system scans more than 40 internal and external channels to monitor front-office staff, and around 200 legacy surveillance servers have been retired. Sid Nadella, director and global head of capital markets solutions at Google Cloud, frames the direction: “Banks are worried about data loss prevention. Monitoring communications, making sure there is no problematic activity, has always been part of it. That can be amplified by using AI.”
The second phase and its German particularity
Alongside trade surveillance proper, LLM-based monitoring of staff communications is planned, with rollout in 2026. It is intended to pick up abnormal behaviour such as forwarding confidential information to personal email addresses.
Here the case leaves purely technical ground. In Germany, introducing technical systems capable of monitoring employee behaviour or performance falls under works council co-determination. Alongside that sit the General Data Protection Regulation requirements on necessity and proportionality. None of the sources assessed addresses this, so the point is my own assessment rather than a statement by the companies involved. For institutions planning something comparable, it will likely shape the timetable more than the technology does.
How far the work has come
To gauge maturity it pays to read the sources' choice of tense closely. They speak consistently in the future or conditional: the system is being developed, the intention applies once it is operational, the rollout is slated for later in the year. Anomaly detection in orders and trades is furthest advanced, yet nowhere is it described as live.
Two earlier efforts are routinely blended into this account and belong apart. In September 2023 Bloomberg reported on a Deutsche Bank pilot analysing tone in telephone calls, also with Google Cloud and also featuring Bernd Leukert. A Google Cloud blog post of February 2024 describes a data infrastructure for trade surveillance with around 30 per cent IT cost savings. Both evidence a multi-year partnership without belonging to the 2026 LLM story.
Goldman Sachs appears in the same report, though far more vaguely: the firm has been looking into using agentic AI to analyse trades and search for suspicious signals. No executive is quoted by name and no figures are given.
What this means in practice
Four starting points for leaders in trading, compliance and market abuse surveillance.
Now: Whether a 25 or 40 per cent reduction counts as large depends entirely on the starting point. Without knowing your own rate and the handling time per alert, you cannot assess a proposal or build a business case. That measurement takes a few weeks and is useful regardless of any technology decision.
In planning: The model described analyses and proposes; the compliance officer decides and closes the alert. That split belongs in the target architecture in writing before the first vendor speaks. It determines which evidence you can later put in front of a supervisor.
Now: Monitoring staff communications requires works council agreement in Germany and needs a data protection justification. Bring the works council in only after vendor selection and you risk a project halt with budget already committed. That sequence shapes the timetable more than model quality does.
Ongoing: This case shows exactly how two values from an article covering three banks end up attributed to one. Anyone lifting such figures into a board paper should check which institution stated them and whether they describe an outcome or an expectation. Secondary sources routinely make the two indistinguishable.
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 fortnight in your inbox.
Capital markets insights, regulatory updates and AI trends. Concise, well-founded, free of charge.
GDPR-compliant. Unsubscribe at any time.