ZTABS deploys self-hosted AI infrastructure for Berlin enterprises across BaFin-supervised financial institutions, healthcare organisations, manufacturers and government agencies. Self-hosted deployment runs open-weight language models on customer-controlled hardware. Whether data remains within the organisation's security boundary depends on the actual configured full-stack data paths; on-premises hardware alone does not inherently guarantee zero external transmission.
Berlin self-hosted AI demand spans financial services (compliance automation, credit analysis, risk modelling), healthcare (clinical documentation, research analytics, patient engagement), manufacturing (quality AI, engineering documentation, process optimisation) and government (document analysis, citizen services). ZTABS is remote-first by design, with an office in Houston, Texas. Our self-hosted AI deployment services cover infrastructure deployment across markets.
Deployment scenarios and requirements
| Need | What we build | Key requirement |
|---|
| Financial services | Compliance AI, credit analysis, risk modelling | BaFin supervision, GDPR, EU AI Act |
| Healthcare | Clinical NLP, research analytics, patient tools | MDR where applicable, GDPR |
| Manufacturing | Quality AI, engineering documentation, process support | ISO quality systems, works council |
| Government | Document analysis, citizen services, analytics | BSI requirements, GDPR |
What self-hosted means
Self-hosted deployment runs open-weight language models on hardware the organisation controls. Self-hosting addresses scenarios where cloud AI APIs are contractually prohibited, where data classification requires controlled infrastructure, or where organisational policy favours infrastructure the customer controls.
What self-hosting does not automatically provide. On-premises hardware does not inherently guarantee GDPR compliance, EU AI Act compliance or zero vendor risk. GDPR applies based on controller and processor activities, not server location alone. The EU AI Act regulates AI systems based on risk classification, not deployment location. Self-hosted deployment supports compliance requirements but is not a substitute for comprehensive compliance programmes.
EU AI Act phased implementation. The EU AI Act (Regulation 2024/1689 as amended by Regulation 2026/1744) entered into force August 2024. Prohibited practices from February 2025. General-purpose AI model obligations from August 2025. Annex III listed high-risk systems from 2 December 2027; regulated-product high-risk systems under Annex I from 2 August 2028. These are distinct categories: Annex I covers products subject to EU harmonisation legislation, Annex III lists specific high-risk use cases. Article 50 transparency and Article 73 serious incident reporting apply based on classification and intended purpose; not all AI is high-risk and obligations differ by role.
Works council co-determination. BetrVG section 87(1)(6) gives works councils co-determination rights over technical devices with monitoring capability. AI systems that have the capability to monitor employee behaviour or performance may trigger co-determination regardless of whether monitoring is the primary purpose. Works council engagement is required before deployment where applicable.
Model licences. Open-weight models have specific licence terms. Llama 3 uses Meta's custom Community License. Open-weight means trained weights available for download; it does not mean open-source or full training-corpus disclosure.
Planning self-hosted AI for your Berlin organisation?
Tell us about your infrastructure, security requirements and use cases. We start with environment assessment and workload analysis.
Financial services deployment
Berlin's BaFin-supervised financial institutions create demand for self-hosted AI supporting compliance, credit analysis and risk modelling.
BaFin supervision. BaFin supervises banks, insurance companies, financial services institutions and securities trading. DORA directly covered entities are subject to DORA from 17 January 2025; certain German national-extension institutions have BAIT transitional application through 31 December 2026 before full DORA alignment from 1 January 2027. MaRisk continues alongside DORA. AI systems used in material risk management must meet applicable model governance requirements. Self-hosted deployment supports governance and auditability.
DORA compliance. The Digital Operational Resilience Act applies to financial entities in the EU from January 2025. DORA establishes ICT risk management, incident reporting, digital operational resilience testing and third-party risk management requirements. Self-hosted deployment addresses certain third-party dependencies but does not eliminate all DORA obligations; the institution's ICT risk management framework must account for the AI system.
Credit analysis. Self-hosted language models support credit assessment by processing financial statements, industry reports and application documents within the institution's infrastructure. EU AI Act Annex III includes AI systems for creditworthiness assessment of natural persons as high-risk, with a specific exception for systems used for financial-fraud detection. Credit decisions require qualified credit personnel within the institution's lending authority.
Compliance automation. Self-hosted AI supports AML/CFT compliance by processing transaction monitoring, customer documentation and sanctions screening. EU Anti-Money Laundering framework obligations apply. AI provides analytical support; reporting decisions are human determinations by authorised compliance personnel.
Healthcare deployment
Berlin's healthcare organisations create demand for self-hosted AI supporting clinical documentation and research.
EU MDR considerations. The EU Medical Device Regulation applies to software that qualifies as a medical device based on intended medical purpose. Whether a particular clinical AI system requires MDR classification depends on its intended purpose, claims made and function. Engineering validation is distinct from clinical authority and regulatory classification.
Clinical documentation. Self-hosted language models support German and English clinical documentation, coding assistance and research analytics. Clinical decisions require qualified healthcare professionals. German medical terminology processing and DRG coding assistance require evaluation against the engagement's clinical data.
Research analytics. Self-hosted AI supports analysis of research datasets within the organisation's infrastructure. Research use requires applicable ethics approval and data governance compliance. GDPR research exemptions under Article 89 have specific conditions including appropriate safeguards.
Manufacturing deployment
Berlin's manufacturing sector creates demand for self-hosted AI supporting quality systems and engineering documentation.
Quality AI. Self-hosted language models support quality documentation processing, non-conformance analysis and CAPA (Corrective and Preventive Action) management. Quality decisions require qualified personnel within the manufacturer's quality management system. AI provides analytical support and documentation assistance.
Works council requirements. Manufacturing deployments must address BetrVG 87(1)(6) co-determination where AI systems have monitoring capability for employee behaviour or performance. BetrVG 90 additionally requires timely information and consultation on AI planning that affects work processes and the working environment, distinct from the 87(1)(6) co-determination right. Not every AI system requires works council approval; the trigger is monitoring capability, not AI use generally. Process AI that analyses employee-generated data or work patterns may trigger co-determination. Engagement with the works council before deployment is required where applicable.
Engineering documentation. Self-hosted AI supports technical documentation processing, specification analysis and process documentation within the manufacturer's infrastructure. German-language engineering documents require model evaluation against the specific technical vocabulary and document formats. Technical standards references, safety documentation and regulatory compliance records require domain-specific terminology accuracy that must be benchmarked against the manufacturer's actual documentation corpus.
Government deployment
Berlin's government agencies create demand for self-hosted AI within controlled environments.
BSI requirements. The Federal Office for Information Security (BSI) establishes security requirements for government IT systems. Government data classification determines security controls. Self-hosted deployment supports classified processing requirements where mandated.
Digital services AI. Self-hosted AI supports government document processing, citizen service automation and internal analytics. German-language capability is essential. Government decisions and citizen-facing responses remain with authorised officials.
Full-stack data path mapping
Self-hosted deployment requires mapping every component that processes data. The full stack includes the inference engine, tokeniser, embedding model, vector database, key-value cache, authentication layer, telemetry pipeline, backup systems, administrative interfaces, support channels, model update paths, tool-calling endpoints and RAG data sources. Each component has its own data flow and potential egress path. Data access propagates through retrieval results, cached responses, trace logs, exported analytics and adapters; this requires active management at each layer.
Supply chain and update management
Approved supply chain. Model weights and dependencies follow approved procurement with provenance review. Hash verification and scans reduce risk but cannot prove the whole stack is uncompromised. Staging with evaluation and rollback capability precede production deployment.
Illustrative Example
Example deployment: Berlin financial institution self-hosted compliance AI
A Berlin-based BaFin-supervised bank needs self-hosted AI to support AML/CFT compliance screening, credit document analysis and regulatory reporting within its controlled infrastructure.
The deployment would install open-weight language models on GPU infrastructure within the bank's environment. AI models would process transaction monitoring alerts, extract structured data from customer documentation and support compliance analysts with case preparation. Credit analysis models would process German and English financial statements.
Assumptions: the bank's infrastructure team manages the environment, existing banking and compliance systems remain operational alongside the AI enhancement, works council engagement under BetrVG 87(1)(6) is completed before deployment where the system has monitoring capability, EU AI Act Annex III high-risk classification for creditworthiness assessment of natural persons (except financial-fraud detection) is addressed, all credit decisions remain with authorised officers, and all compliance reporting decisions remain with authorised personnel.
Residual risks include model accuracy variation across German legal and financial document types, works council negotiation timeline, EU AI Act compliance documentation requirements, GDPR deletion limitations for trained models, GPU capacity planning and operational knowledge transfer. No guarantee of specific false positive reduction, processing speed improvement or regulatory outcome.
This is a hypothetical scenario for illustration. Actual scope, timeline and cost depend on your requirements.
Cost drivers for self-hosted AI
Self-hosted AI deployment has no published cost guide on this site. Cost varies significantly across engagements.
Model and workload factors. Model size, context requirements, KV cache, precision, batching and concurrency.
Infrastructure factors. GPU procurement, redundancy, power, cooling, network configuration and disaster recovery.
Regulatory factors. BaFin documentation, GDPR compliance support, EU AI Act documentation for applicable risk categories, works council engagement support, DORA ICT risk documentation where applicable.
Integration factors. Connection to banking, compliance, clinical and manufacturing systems. Each integration's API access and data entitlements must be confirmed.
Operations and handover factors. Documentation, monitoring, model update procedures, capacity planning and knowledge transfer.
Need to scope self-hosted AI for your Berlin organisation?
Share your environment, security requirements and use cases. We provide a deployment assessment covering infrastructure, compliance and operational scope.
Proposal evaluation criteria
- Relevant self-hosted AI deployment experience across financial services, healthcare or manufacturing environments with supporting evidence
- Full-stack infrastructure design including data flow mapping, security architecture and regulatory documentation
- Integration plan covering actual API access, licence entitlements and data permissions
- Delivery milestones with measurable acceptance criteria including works council engagement timeline where applicable
- Operations ownership covering model updates, patching, rollback and monitoring with clear responsibility allocation
- Provider team identifies who delivers engineering, trains users, delivers documentation and provides support; customer team identifies who provides environment access, validates domain-specific outputs (compliance accuracy, credit analysis quality, clinical NLP relevance), accepts milestones and handles post-handover operations; both parties agree staff and user languages, documentation ownership and scope, translation responsibilities and handover materials
How delivery works
ZTABS is remote-first by design, with an office in Houston, Texas. CET/CEST timezone coordination for Berlin clients; scheduling and staff language commitments are agreed per engagement.
Environment assessment. Infrastructure review, security boundary analysis, integration targets and workload benchmarking.
Model selection. Candidate models evaluated against domain-specific test sets including German and English capabilities. Multilingual model capability is distinct from staff language availability.
Security and regulatory readiness. Applicable controls implemented. EU AI Act documentation, GDPR documentation and works council engagement support prepared where applicable.
Pilot and acceptance. Pilot targets a single use case. Rollout extends with rollback capability.
Training and handover. User training, operations transfer and complete documentation.
When ZTABS is not the right fit
- Turnkey GPU hardware procurement for this engagement. We design requirements; procurement follows the organisation's channels.
- Qualified compliance officer, credit analyst, clinical authority or works council negotiator for this engagement. These require authorised personnel.
- Ongoing managed hosting for this engagement. Internal operations post-handover; retainers separately agreed.
- BaFin licence application or regulatory representation for this engagement. Regulatory representation requires separate qualification.
- On-site embedded team for this engagement. ZTABS is remote-first by design. On-site separately agreed.
White-label AI products
ZTABS builds custom white-label AI products for clients to brand and resell. Berlin enterprises can commission white-label self-hosted AI platforms for distribution.
Data governance
Each engagement agrees data classification, processing boundaries and retention requirements. GDPR obligations apply including lawful basis, data minimisation, purpose limitation and rights fulfilment. Data access propagates through retrieval results, cached responses, trace logs, exports and adapters; this requires active management at each layer. Training or fine-tuning on customer data requires documented authorised purpose, training-data scope, custodian approval and purpose limitations. Revocation is agreed and verified across active retrieval paths, vector stores, caches, exports, logs, backups and adapters as applicable; propagation lag is disclosed and affected access restricted while propagation is incomplete. Lawful retention, legal holds and regulatory preservation obligations are distinguished from active retrieval and use. Model weights may retain residual statistical patterns not individually reversible through source deletion; the engagement does not promise instant erasure of every backup or model weight. Residual-risk acceptance documents technical limitations but does not waive statutory rights or legal obligations. Separate legal review addresses GDPR data-subject rights, regulatory retention and applicable remedies.
Monitoring and operations
Infrastructure monitoring. GPU utilisation, memory, latency, throughput, storage and network health.
Security monitoring. Access logs, authentication events, anomaly detection. BSI reporting requirements where applicable.
Model quality. Evaluated against held-out test sets with domain-specific criteria including German language performance, financial compliance accuracy and EU AI Act monitoring requirements for applicable risk categories. Drift detection and periodic re-evaluation. Monitoring reduces risk but cannot guarantee all regressions are caught.
Capacity management. GPU utilisation trends inform capacity planning.
Handover
Deployment handover includes infrastructure runbooks, model configuration documentation, security and regulatory compliance documentation including EU AI Act documentation where applicable, monitoring dashboards, model update procedures, capacity planning guidance and escalation contacts. Post-handover support terms agreed per engagement.
Ready to deploy self-hosted AI in Berlin?
Share your infrastructure, security requirements and use cases. We start with environment assessment and workload analysis.