Intelligent AI automation in core processes_
Efficient, over 99% correct, legally compliant

Companies can offer their employees the benefits they deserve with billyard! The platform makes management of benefits such as subsidies for food, mobility, fitness, health and other benefits child’s play – and legally compliant to boot. Companies and employees alike can manage benefits easily and intuitively with legal and tax aspects integrated into the platform (certified according to IDW-PS 860). billyard is leading the way in digitising all processes related to employee benefits.

Context

An important employee benefit that can be managed through billyard are meal allowances. Companies can subsidize employees with tax benefits by covering their meal costs. billyard digitizes and simplifies the administrative effort while also taking into account tax regulations. The employees are then automatically billed via payroll accounting.

To minimise the effort for everyone involved, employees submit proof of purchase for their relevant meals in an app. To do this, they can take a photo of the proof of purchase on their mobile phone or upload it as a file.

The document’s data must be transferred to billyard’s system: document issuer with address, date and time, the items, the chargeable amount and the document number. billyard has to then double-check the data taking into account all rules and laws, such as: is it food, has the receipt number already been submitted, is the time appropriate for a breakfast or lunch subsidy?

An example: A two-kilo bag of potatoes is not eligible for the lunch subsidy (stock purchase), but 300 grams of potatoes as ingredients for a meal are. Subsidies that do not comply with the law can have considerable consequences for the employer in the context of tax audits.

The challenge

These requirements are already complex enough for an automated system – but on top of that, purchase receipts do not have a uniform format. Although they must contain certain data, the arrangement, format and design do not follow set standards. In addition, users upload the receipts to the system as photos (pixel graphics), which initially makes them non-machine-readable.

Legal certainty and correctness of the eligibility check are a critical part of the core business. That’s why previously all receipts had to be checked by humans in a user interface developed for this purpose. This was both labour- and cost-intensive. Hence the goal: to develop a system that delivers automated results with very high quality.

Our service

With our experience from numerous successful AI projects, cronn’s AI team was exactly the right partner for this challenge. In an agile development process, a solution was created that automates the work steps with the help of AI: from recognising key figures from photos of receipts, to the legal check and the determination of the grant amount. This is achieved with the help of software for optical character recognition (OCR) and large language models (LLMs), which we have further prepared for these special tasks using machine learning.

The system also checks with about 99.9% accuracy whether an item is tax-eligible. This precision is crucial because incorrect grant decisions can have legal and financial consequences for employers. For those fringe cases where the system cannot determine with certainty cronn has developed a new user interface for the human reviewers, and here also there has been a significant increase in efficiency.

Innovative methods and full transparency

cronn’s system ensures extremely high accuracy and avoids errors and hallucinations caused by LLMs. To do this, our team built and trained a robust machine learning pipeline. These are operated on-premises to have full control over models and data protection. This approach guarantees resilience, sovereignty and economy in resources.

cronn’s innovation here is to intelligently combine AI tools – all the tools used are interchangeable components. The pipeline consists of an OCR system for character recognition, LLMs that have been optimized for this specific use case and a specially trained classifier which evaluates the eligibility of individual items.

Measurable economic impact

The new AI solution and the new user interface for manual checks have a direct positive effect on billyard’s profitability:

• significant cost reduction thanks to partial and full automation
• strong acceleration of processing
• cost-stable scaling, regardless of document volume
• significant increase in efficiency for human inspectors

The development approach follows the principle of “profit-oriented agile software development”. Each optimization is transparently evaluated for its economic potential and prioritized if it saves costs or improves quality.

In addition, the cronn team attached great importance to complete transparency of the pipeline. While AI solutions are often black boxes, cronn created a system with powerful observability for billyard. We measure exactly where the system is good or can be optimized. In addition, the quality of the output is automatically measured and compared with the results of the human reviewers.

cronn’s solution saves billyard time and money and the collaboration will continue to drive efficiencies in a business-critical core area. And most importantly, the architecture was designed to ensure full data protection and data sovereignty.