Deutsche Bank's six-year cloud migration is what made its document automation possible

The migration came first, the automation came after

The headline result in this post is a generative AI document pipeline running at 97% accuracy with 40% less handling time. But Leukert spends most of the piece describing everything that happened before that number was possible.

Deutsche Bank started migrating to Google Cloud in 2019, moving around 260 applications, beginning deliberately with non-critical ones to build confidence in a regulated environment. The AI work is presented as the payoff of that groundwork, not the starting point.

This leap paid off. We have moved around 260 applications to Google Cloud, including business-critical ones, which allowed us to quickly deploy generative AI technologies.Montana Labs

The implicit argument is that the automation gains were gated on infrastructure readiness. A bank cannot run document extraction at scale on systems it hasn't yet moved, secured, and stabilized.

The S4/HANA move is the load-bearing claim

The most concrete engineering detail is the migration of SAP S4/HANA — 17 financial reporting systems including the strategic general ledger, planning, and forecasting — from on-premise to public cloud, described here as one of the most complex migrations in financial history.

The reported outcomes are specific: data processing improvements of up to 50%, and recovery time reduced by a factor of 16 to 20. Those are operational numbers about resilience and throughput, not model accuracy figures.

For an automation story, this is the more instructive part. The faster recovery and processing headroom are what let the bank layer AI onto core finance without treating it as a science experiment bolted to fragile infrastructure.

Where the automation is actually running

Leukert names three concrete uses of generative AI. Document processing automates thousands of documents daily — extracting information from customer orders and legal documents — at 97% accuracy with 40% less handling time.

A Digital Assistant currently serves employees in research and origination & advisory, curating content for reports and analysis, with stated potential to expand bank-wide. And developers are using generative AI to write code, find bugs, and improve documentation.

These are narrow, bounded deployments — specific document types, specific employee groups — rather than a claim of firm-wide transformation. The 97% figure also quietly concedes a 3% error rate that a regulated bank must still handle through review, which is why the framing stays on assistance and handling-time reduction rather than full replacement.

The 6,000-person training program is part of the automation cost

Deutsche Bank built a Cloud Engineer Program and trained more than 6,000 employees in cloud and AI skills, working with Google Cloud on hackathons and migration sprints. This is presented alongside the technology, not as a footnote.

The pairing is the useful signal for anyone planning similar automation. The bank treated staff capability as a prerequisite for deploying AI tools, not an afterthought — the same people expected to use the Digital Assistant and trust the document pipeline had to be brought up the curve first.

What this sequencing implies for automation projects in regulated firms

The specific implication of Deutsche Bank's account is that the automation wins were downstream of years of unglamorous work: risk-controlled migration, a historically complex core-finance move, and mass retraining.

The document-processing accuracy and the developer tooling are real, but they arrive last in the narrative for a reason. Teams looking at the 97% and 40% numbers should read them as the end of a multi-year sequence, not a plug-in capability — and note that even the bank keeps its live AI use scoped to defined document types and departments rather than the whole institution.