The starting point: Manual work at its limit
DFKP (Deutsche Firmenkredit Partner) is one of Germany’s largest independent financing brokers. ISO-certified and specialized in the German SME sector, the company arranges corporate financing from around €10,000 to several million euros – through a network of more than 200 banking and financing partners. Its core business: aggregating, processing, enriching documents and thereby relieving customers and banking partners of manual work.
Documents are not a side issue; they are the process. Annual financial statements, business management reports (BWAs), trial balances, bank statements, tax records, articles of association, identity documents: without them, there is no pre-scoring, no qualified credit assessment, no offer from a financing partner.
“We turn documents from a bottleneck in the analog process into an enabler in the digital process.” – Maximilian Dassler, Specialist in Digital Credit Processes, DFKP.
The problem at a glance
Between 2,000 and 3,000 documents per week – more than 100,000 per year – had to be manually reviewed, classified and processed by a dedicated team. The range of document classes alone comprised around 200 different types, with 10 to 15 core classes accounting for the lion’s share of the volume. Two to three full-time positions were tied up exclusively with this work – capacity that was lacking in subsequent processes.
Adding to the challenge was the heterogeneity of incoming channels: documents arrived by email, via the customer portal and through API – each channel with its own structure, each document requiring individual effort. Bank statements were the biggest pain point: 300 to 400 different banking systems, no standardized format, and even correctly identifying the billing period required an enormous amount of manual work. The underlying issue: there was no scalability. More volume automatically meant more staff, with all the consequences: longer onboarding periods, quality fluctuations due to turnover, bottlenecks during volume peaks.
The solution: A two-stage AI pipeline with proprietary routing
DFKP evaluated various approaches: enterprise OCR solutions, proprietary ML models, and several IDP providers. The requirements were clear: map heterogeneous document classes, provide an all-in-one solution for separation, classification and extraction, data protection as a prerequisite, fast time to production, and the ability to manage and retrain the system independently.
A decisive test approach tipped the scales in favor of Buildsimple: from the outset, DFKP fed in difficult cases – not synthetic sample documents, but real, heterogeneous records from day-to-day operations. The platform delivered.
Architecture: How the pipeline works
Stage 1: Foundation model:
The trained base model handles classification and extraction. Separate confidence thresholds are defined for each document class. Documents that exceed the threshold are processed fully automatically – without manual review.
Stage 2: LLM for complex cases:
If a document falls below the threshold, a large language model takes over. It can handle unusual characteristics, missing structural elements and atypical formatting. Routing takes place not only at class level, but also within a class – some forms of annual financial statements pass through directly, while others require the LLM.
Control remains with the customer:
Responsibility for the business control logic rests entirely with DFKP. Buildsimple provides the platform – DFKP determines which documents are processed how, which error rates are acceptable, and when each model is used. Thresholds are calibrated in regular reviews using real production data.
Human in the loop: Quality assurance as an integral part of the system
Full automation is not a goal at any cost. Documents that do not meet the minimum thresholds are sent to a dedicated quality team at DFKP. Buildsimple’s intuitive processing interface makes review efficient.
Corrected results are automatically fed back via API and serve as training data. Advisors in subsequent processes can flag documents that appear to have been classified incorrectly. The system learns during ongoing operation – the straight-through processing rate rises continuously.
Implementation: 3 months from kickoff to production operation
The POC began in early 2025 and was not a sandbox operation from day one. Real documents, real volume, real error rates. The project started with the 10 to 20 document classes that cover 85–90% of total volume – classification and extraction were trained together from the outset.
What made the unusually short ramp-up time possible: DFKP brought pre-labeled data from its own day-to-day operations. This eliminated the usual cold start entirely – classification was running virtually from day two of the POC. At the same time, the team was able to operate the tool independently from the beginning: upload themselves, configure themselves, adjust themselves. Buildsimple experts were closely involved in complex edge cases, but business control remained with DFKP throughout. Today, the pipeline has been running reliably for more than a year and is considered a best practice for introducing IDP systems in regulated environments, both internally and externally.
Results: Measurable, concrete, business-critical
70% reduction in manual effort
80–85% straight-through processing rate
3x document throughput
30–50% fewer errors
Less effort, greater depth: With the same headcount, the team processes three times the number of documents. At the same time, more data points are now extracted than ever before – not fewer. Fields that were previously omitted for time reasons now flow automatically into the Finance Engine.
Better matching, higher conversion: A large proportion of the data points on which automated pre-scoring is based comes directly from the extracted documents. More data points, higher quality, better matching between borrower and financing partner, and therefore a measurably higher conversion probability with the banking partner.
Document quality as a competitive advantage: Banking partners send poorly prepared documents back. In financing brokerage, this is not an edge case but everyday business. Today, DFKP differentiates itself through document quality – structured, complete, and transferable directly into banks’ end systems. Some partner banks already have a fully digital interface, in some cases with live offers in real time.
Measurably shorter time to offer Advisors in subsequent processes are no longer held up by slow document processing. The time from document receipt to a qualified credit assessment has decreased measurably – a direct advantage for the end customer.
Team effects: From clerk to case manager
The introduction met with remarkably little resistance. Two factors made the difference: the tangible elimination of repetitive work and Buildsimple’s processing interface, which is more intuitive than the previous Salesforce custom view.
Freed-up capacity was not cut, but repurposed: in a pilot project, lead management staff now also handle smaller customers themselves – customer cases that previously would not have gone through the full advisory process. The role profile has been expanded and responsibility has increased.
The consequence: DFKP does not optimize for the maximum automation rate, but for economically sensible automation. Freed-up capacity flows into handling the genuinely complex cases – that is the real scaling gain.
Conclusion
In three months, DFKP built a production-ready IDP pipeline that now processes 80–85% of documents fully automatically – more reliably, faster and with greater data depth than the manual process could ever provide. Buildsimple did not simply automate a process; it created the foundation for a new level of digitalization in the credit brokerage business: from document intake through to the bank’s end system, in an end-to-end, quality-assured workflow.
