Introduction of AI in input processing

The ARD ZDF Deutschlandradio Contribution Service in Germany from Europe region has released this tender for Introduction of AI in inbound processing. Subject of the procedure: The ARD, ZDF, and Deutschlandradio Contribution Service is procuring an AI-based service for the automated processing of incoming written documents (paper/scan, email, fax), which will be operated as a managed service (IaaS, PaaS, SaaS) on a sovereign cloud. The AI ​​service classifies the documents according to the Contribution Service's document type classification system (DAK/SDAK), identifies the subject matter, and extracts the necessary data for further processing. The results, each with a confidence score, are returned to the existing processing workflow as structured JSON objects. The AI ​​service is being implemented not as a replacement, but as a bypass for the existing inbound processing. The routing of the documents is handled by intelligent switches provided by the Contribution Service; the core systems remain unchanged. The scanning or digitization of the mailings, the input and output interfaces of the overall inbound processing, and the intelligent switches themselves are not part of this tender. The scope of services is divided into three components: Component 1 - Cloud Infrastructure: Provision of a sovereign, highly available, and scalable cloud environment as IaaS, including on-demand scalable CPU, GPU, storage, and network resources for AI workloads. It is intended to conclude a contract for Components 1 and 2 with a fixed term of four years. Component 2 - Managed Service: Operation of the AI ​​platform as PaaS and SaaS, in particular monitoring, logging, incident and problem management, patch and release management, backup and recovery, capacity management, and provision and operation of a standardized AI serving stack (including container orchestration and inference). The broadcasting service retains control over the business logic, agents, and the selection and switching of AI models. The intention is to conclude a contract for modules 1 and 2 with a fixed term of four years. Module 3 – Development of a customized input management solution ("AI Service"): Conception, development, integration, and deployment of a customized solution based on the provided platform, including the setup of classification and extraction for each issue, proof of the agreed recognition quality, training, documentation, and acceptance testing support. The AI ​​Service, including all work results (e.g., source code, configurations, prompt artifacts, data models, interface descriptions, tests, documentation), will be transferred to the Broadcasting Service with unlimited usage rights in terms of time, location, and content. The Broadcasting Service will also be granted the right to use, operate, modify, and further develop the AI ​​Service without restriction, and to have it operated or further developed by third parties.No dependency on the original contractor may arise, and switching providers must be possible in principle. The solution should primarily be implemented using open-source components with permissive licenses. Additionally, a service quota should be provided for other matters of comparable complexity and for optimizations after go-live or acceptance testing. Scope and volume: The focus is on approximately 1.8 million informal letters per year, which are currently neither automatically classified nor analyzed for content. In addition, there are approximately 4.7 million formal letters from five categories: exemptions, deregistration, third-party debtor declarations, clarification procedures due to death, and bank and address changes. For sizing purposes, a minimum of 6.5 million mailings per year should be assumed (approximately 25,000 mailings per working day, peak value approximately 26,700). Handwritten mail accounts for approximately 20%. In some cases, special categories of personal data pursuant to Article 9 GDPR (e.g., health data) will need to be processed. Key functional requirements include, in particular: - AI-supported text recognition (OCR) across the entire document content, including multi-page documents and attachments; - content-based classification and issue recognition using text understanding instead of form anchors, including recognition of multiple issues per submission and separation of multiple documents; - extraction of identification fields (e.g., contribution number, name, address) and the specialist data required for each case, also from attached supporting documents; - recognition of handwritten content, including free-form marginal notes, checkboxes, and verification for the presence of signatures; - confidence scores for each result and field; no inclusion of unverifiable content; protection against prompt injection; - integrated human-in-the-loop path with configurable control rules and an accessible (BITV 2.0), German-language, fully keyboard-operated post-processing interface; - inclusion of further document classes without requiring the contribution service to rebuild the system itself; no training of model weights with personal data of the contribution payers; - Model-agnostic architecture (preferably open-weight models), REST interfaces, container/microservice architecture, UTF-8, reporting (including precision/recall/F1 for each document class), and integration with the broadcasting service's monitoring tools. The description of the services to be tendered continues in section 5.1 of this announcement.Scope and volume: The focus is on approximately 1.8 million informal letters per year, which are currently neither automatically classified nor analyzed for content. In addition, there are approximately 4.7 million formal letters from five categories: exemptions, deregistration, third-party debtor declarations, clarification procedures due to death, and bank and address changes. For sizing purposes, a minimum of 6.5 million mailings per year is assumed (approximately 25,000 mailings per working day, peaking at approximately 26,700). Handwritten mail accounts for approximately 20%. In some cases, special categories of personal data pursuant to Article 9 GDPR (e.g., health information) will be processed. Key functional requirements include: - AI-supported optical character recognition (OCR) across the entire document content, including multi-page documents and attachments; - content-based classification and issue recognition using text understanding instead of form anchors, including the recognition of multiple issues per mailing and the separation of multiple documents. - Extraction of identification fields (including contribution number, name, address) and the specialist data required for each case, also from attached supporting documents; - Recognition of handwritten content including free margin notes, checkboxes, and verification for the presence of signatures; - Confidence scores for each result and field; no inclusion of unverifiable content; protection against prompt injection; - Integrated human-in-the-loop path with configurable control rules and an accessible (BITV 2.0), German-language, fully keyboard-operated post-processing interface; - Inclusion of further document classes without requiring the contribution service to rebuild the system itself; no training of model weights with personal data of the contribution payers; - Model-agnostic architecture (preferably open-weight models), REST interfaces, container/microservice architecture, UTF-8, reporting (including precision/recall/F1 for each document class), and connection to the contribution service's monitoring tools. The description of the services to be tendered continues under section 5.1 of this notice.Scope and volume: The focus is on approximately 1.8 million informal letters per year, which are currently neither automatically classified nor analyzed for content. In addition, there are approximately 4.7 million formal letters from five categories: exemptions, deregistration, third-party debtor declarations, clarification procedures due to death, and bank and address changes. For sizing purposes, a minimum of 6.5 million mailings per year is assumed (approximately 25,000 mailings per working day, peaking at approximately 26,700). Handwritten mail accounts for approximately 20%. In some cases, special categories of personal data pursuant to Article 9 GDPR (e.g., health information) will be processed. Key functional requirements include: - AI-supported optical character recognition (OCR) across the entire document content, including multi-page documents and attachments; - content-based classification and issue recognition using text understanding instead of form anchors, including the recognition of multiple issues per mailing and the separation of multiple documents. - Extraction of identification fields (including contribution number, name, address) and the specialist data required for each case, also from attached supporting documents; - Recognition of handwritten content including free margin notes, checkboxes, and verification for the presence of signatures; - Confidence scores for each result and field; no inclusion of unverifiable content; protection against prompt injection; - Integrated human-in-the-loop path with configurable control rules and an accessible (BITV 2.0), German-language, fully keyboard-operated post-processing interface; - Inclusion of further document classes without requiring the contribution service to rebuild the system itself; no training of model weights with personal data of the contribution payers; - Model-agnostic architecture (preferably open-weight models), REST interfaces, container/microservice architecture, UTF-8, reporting (including precision/recall/F1 for each document class), and connection to the contribution service's monitoring tools. The description of the services to be tendered continues under section 5.1 of this notice.- AI-powered optical character recognition (OCR) across the entire document content, including multi-page documents and attachments; - Content-based classification and issue recognition using text understanding instead of form anchors, including recognition of multiple issues per submission and separation of multiple documents; - Extraction of identification fields (e.g., contribution number, name, address) and the specialist data required for each case, also from attached supporting documents; - Recognition of handwritten content, including free-form marginal notes, checkboxes, and verification for the presence of signatures; - Confidence scores for each result and field; no inclusion of unverifiable content; protection against prompt injection; - Integrated human-in-the-loop path with configurable control rules and an accessible (BITV 2.0), German-language, fully keyboard-operated post-processing interface; - Inclusion of further document classes without requiring the contribution service to rebuild the system itself; no training of model weights with personal data of the contribution payers; - Model-agnostic architecture (preferably open-weight models), REST interfaces, container/microservice architecture, UTF-8, reporting (including precision/recall/F1 for each document class), and integration with the broadcasting service's monitoring tools. The description of the services to be tendered continues in section 5.1 of this announcement.- AI-powered optical character recognition (OCR) across the entire document content, including multi-page documents and attachments; - Content-based classification and issue recognition using text understanding instead of form anchors, including recognition of multiple issues per submission and separation of multiple documents; - Extraction of identification fields (e.g., contribution number, name, address) and the specialist data required for each case, also from attached supporting documents; - Recognition of handwritten content, including free-form marginal notes, checkboxes, and verification for the presence of signatures; - Confidence scores for each result and field; no inclusion of unverifiable content; protection against prompt injection; - Integrated human-in-the-loop path with configurable control rules and an accessible (BITV 2.0), German-language, fully keyboard-operated post-processing interface; - Inclusion of further document classes without requiring the contribution service to rebuild the system itself; no training of model weights with personal data of the contribution payers; - Model-agnostic architecture (preferably open-weight models), REST interfaces, container/microservice architecture, UTF-8, reporting (including precision/recall/F1 for each document class), and integration with the broadcasting service's monitoring tools. The description of the services to be tendered continues in section 5.1 of this announcement. with tender notice no. BW 04/26 published on 13 Oct 2026. The tender is due to expire on 12 Nov 2026. This tender is sourced from Vergabe and is eligible for suppliers interested in Customer Relation Management software package. Interested bidders can obtain further information including complete bid documents or ways to participate by registering on Tender Impulse and referencing via Tender Impulse ID 15465618.