AI Agent Development Company in Hamburg

All Hamburg engagements are fixed-fee in EUR with milestones rather than open T and M, scoped per engagement rather than published as bands. A single-purpose production agent — one workflow, one outcome, integrated into one back-office system, with an evaluation harness and observability — is a ten to sixteen week engagement. A multi-agent orchestration (planner, researcher, writer, executor pattern) integrated into multiple back-office systems with retrieval over a substantial document corpus runs four to seven months. A high-risk EU AI Act use case (employee evaluation, content moderation at scale, biometric inference) adds the Annex III conformity workflow, technical documentation, and human-oversight design. Ongoing operations — model API spend, observability, regression testing on golden datasets — are scoped against traffic volume. All Hamburg engagements are fixed-fee in EUR with milestones rather than open T and M.

The EU AI Act entered into force on 1 August 2024 with a staged rollout: prohibited AI systems and AI-literacy obligations from February 2025, general-purpose AI rules from August 2025, and high-risk system obligations from August 2026 (with some exceptions extending to 2027). For most Hamburg enterprise agents — internal copilots, customer-service deflection, document-extraction workflows — the classification is limited-risk or minimal-risk, requiring transparency obligations (the user must know they are interacting with an AI) and basic governance. The classification trips into high-risk for specific Annex III use cases including employment decisions (recruitment scoring, performance management), creditworthiness assessment, critical infrastructure operation, and law-enforcement. Our discovery phase runs the classification, documents the reasoning, and engineers the controls the relevant risk tier requires — including the Article 12 logging, Article 13 transparency, and Article 14 human-oversight obligations for high-risk systems.

Yes, with strict guardrails. Aerospace maintenance is regulated by EASA Part-145 and Part-66 in Europe, plus the FAA equivalents for transatlantic operators, and any agent that surfaces a maintenance procedure to a certified mechanic must respect the source-of-truth constraint absolutely — the procedure in the Aircraft Maintenance Manual is authoritative, the agent is a navigation tool only. Our pattern uses tightly-constrained retrieval over the AMM, IPC, TSM, and SRM corpus, mandatory citation back to the specific page and revision, confidence thresholds that escalate to a human aerospace engineer when retrieval grounding is uncertain, and a documented audit trail of every query and response for EASA-acceptable record-keeping. We do not generate maintenance procedures — we surface authoritative ones — and we have engineered to comparable patterns for adjacent maintenance-document corpora in other regulated industries.

Yes. All Hamburg-facing agent output passes through a German-language quality gate before production. For Beiersdorf, Otto, Tchibo, or NDR-style brand-voice constraints we ship with a written brand-voice specification (tone, register, forbidden phrasings, preferred terminology mapped to the client's existing brand book), a regression test suite of prompts that historically broke brand voice and now must not, and an evaluation harness that human raters score on every release. The German-native model behaviour is currently strongest on Claude 3.5 Sonnet and Claude 3.7 Sonnet, with GPT-4.1 close behind and Mistral Large competitive for sovereignty-sensitive workloads. Self-hosted Llama 3.3 and Qwen 2.5 are acceptable when fine-tuned on the client's brand corpus and run through the same evaluation harness — but we will tell you up front when the cost of fine-tuning and self-hosting outweighs the saving over hosted-API spend.

Hamburg clients vary in their data-residency tolerance. For non-sensitive internal workflows (internal-knowledge copilots, draft generation for marketing) we use Anthropic Claude or OpenAI on the providers' US infrastructure with EU-US Data Privacy Framework as the transfer basis and the Hamburgischer Beauftragter's published guidance on third-country transfers documented in a Transfer Impact Assessment. For customer-facing workflows touching personal data we move to Anthropic Claude on AWS Bedrock in eu-central-1 Frankfurt or eu-west-1 Dublin (Bedrock Claude in eu-central-1 went GA in 2024), Azure OpenAI in Germany West Central or Sweden Central, or self-hosted Llama 3.3 and Qwen 2.5 on AWS Bedrock or Azure ML in EU regions. For maximum sovereignty (Airbus or HHLA regulated workloads) we self-host on customer-controlled EU infrastructure with no third-country transfer at all.

Every Hamburg agent ships with a four-part evaluation harness. First, a golden dataset of representative queries with documented expected behaviours (or, for open-ended workflows, a documented rubric) that runs on every model upgrade and prompt change. Second, regression tests on the historical bug catalogue — every defect we have shipped becomes a permanent test case. Third, online evaluation with sampled human review of production traces against the same rubric. Fourth, drift alerts on quantitative metrics (response latency, tool-call success rate, retrieval-grounding scores, escalation rate to humans) that trigger investigation when they move outside the trained band. Observability runs on Langfuse self-hosted in EU regions for sovereignty-sensitive clients, or LangSmith and Arize for clients that accept US-hosted observability with DPF terms. Monthly evaluation reports are German-language documents for the product owner and the AI-governance lead to read together.

Yes. Most Hamburg enterprise integration we do is into SAP S/4HANA and SAP CRM (the dominant ERP at Beiersdorf, Otto, Tchibo, HHLA), ServiceNow (the dominant ITSM at the larger banks and aerospace clients), Salesforce Service Cloud (where the customer-service workflow lives), and the Atlassian suite for engineering-facing agents. Integration is typically through the official APIs (SAP OData services, ServiceNow Table API, Salesforce REST and Bulk API) wrapped in tool definitions that LangGraph or the Anthropic Claude tool-use SDK can call. For SAP specifically, we use the SAP Business Technology Platform's AI Core and AI Launchpad where the client is already invested in BTP, and we engineer rate-limiting and quota management into the tool layer to keep agent traffic from degrading the production ERP.

A representative Hamburg agent project for a maritime, aerospace, or consumer-goods client runs fourteen to twenty-two weeks. Weeks 1 to 3 cover discovery, EU AI Act classification, value-of-time analysis, and the architecture decision record. Weeks 4 to 7 build the retrieval pipeline (document ingestion, chunking, embedding, evaluation of retrieval quality), the tool layer for back-office integrations, and the basic agent loop with a small golden-dataset evaluation. Weeks 8 to 12 expand the workflow coverage, harden the German-language quality gate, build the human-escalation path, and run shadow-mode deployment against current human operators. Weeks 13 to 18 cover production cutover in waves, observability and drift monitoring, and the documented evaluation report for the AI-governance lead. Weeks 19 to 22 cover hardening, EU AI Act conformity documentation if the use case is high-risk, and the handoff to client-side operations. We have shipped to this cadence for Hamburg consumer-goods and logistics clients.