Compact definitions of the key terms around AI automation, process automation, ERP/CRM integration and AI operations — AI agent, workflow, document automation, on-premise, managed AI operations, AI governance, monitoring, rollback, exception routing, drift, confidence threshold, GDPR. Practical, without marketing language — this is how we use these terms in DACH enterprise projects.
If a term remains unclear in our guides or pillar pages, you'll find the definition here.
A software component that is given a predefined goal and decides for itself which tools to use, and in what order, to reach it. The difference from classic workflows: the path isn't fixed — the goal is.
→ When an agent pays off, and when a workflow does · → AI Agents (Pillar 3)
Structures, responsibilities and processes that ensure AI solutions are operated in a GDPR-compliant, auditable way and in line with corporate and legal requirements (e.g. the EU AI Act). Typical building blocks: records of processing activities, risk classification, approval and escalation paths, audit logs.
The threshold above which the model decides for itself whether to complete a task fully automatically or escalate it to a human reviewer. Typical tiers: high confidence → full automation, medium → processing with pre-filled data, low → fully manual handling. A key element of any resilient AI solution.
Connecting external applications to a CRM system (Customer Relationship Management). Common target systems: HubSpot, Salesforce, Microsoft Dynamics 365 Sales, Pipedrive, Zoho. Typical AI applications: lead classification, automatic activity logging, enrichment of customer records.
Automatic capture, classification and processing of documents such as invoices, delivery notes or contracts. Combines OCR (text recognition), AI-based field extraction and ERP integration. Recognition rates of 90–95 percent for structured documents from known suppliers.
→ Where document automation actually works in practice · → Document Automation (Pillar 2)
Gradual decline in model quality after go-live, because the input data or the environment changes (new product lines, new language variants, new suppliers, new request types). Detected in Managed AI Operations through ongoing monitoring and corrected via model updates or retraining.
Connecting external applications to an ERP system (Enterprise Resource Planning) via documented interfaces — REST API, EDI, structured import formats. Common target systems in DACH companies: SAP, Microsoft Dynamics 365, BMD, DATEV, Sage, Oracle NetSuite. A prerequisite for productive AI automation in most business processes.
→ ERP/CRM integration without breaking your systems · → ERP/CRM Integration (Pillar 4)
Escalation logic that routes cases that can't be handled fully automatically to the right human reviewer, with the right reasoning attached. Multi-tiered: low confidence → pre-filled for review, medium → processing with a flag, high-confidence anomaly → escalation to the responsible department. Prevents escalations from becoming a dumping ground.
The EU General Data Protection Regulation (GDPR) applies to any AI solution that processes personal data — even indirectly (customer inquiries, delivery addresses, employee data). Practical requirements: a legal basis, data minimization, a record of processing activities, data processing agreements (DPAs), EU hosting where possible, a deletion and disclosure concept.
A defined communication point between two systems. An API (Application Programming Interface) lets applications read and write data in predefined formats, without direct database access. The common form today: REST over HTTPS. Still found in business: SOAP, EDI, structured CSV/XML import.
Ongoing operation of an AI solution after go-live: monitoring, drift checks, model updates, response to anomalies. Responsibility sits either with the internal team or with an external partner under a service contract with defined SLAs and a monthly operations report.
Ongoing tracking of the central operational KPIs of a production AI solution — turnaround time, escalation rate, error rate, processing volume, token cost. Forms the basis for drift detection, anomaly response and the monthly operations report. Stack typically built on Langfuse or OpenTelemetry plus classic infrastructure telemetry (Prometheus / Grafana).
→ Managed AI Operations in practice · → Managed AI Operations (Pillar 5)
The AI solution runs on your own infrastructure instead of in the cloud. Greater data sovereignty, no data flowing to external providers — at the cost of higher operating expense and a longer implementation time. Makes sense for highly sensitive data or strict regulatory requirements (healthcare, defense, critical infrastructure).
A time-boxed first undertaking (8–12 weeks) to test an AI solution on a real, clearly scoped use case. A prerequisite for deciding whether to scale up. What distinguishes it from a demo: real data flow, one measurable key metric, and named operational ownership.
A documented path back to the previous version of a model, configuration, or interface version, in case an update causes unexpected quality loss or outages. A requirement for any production AI solution: no update without a rollback path defined in advance. Closely tied to the eval set that's checked before every promotion.
Rule-based automation of recurring business processes with fixed, clearly defined steps. Robust, cheap to run, easy to audit — the right choice whenever the response can be described as a rule. In practice, it is the more common tool in business than the AI agent.
In a free initial call, we discuss which of these concepts are actually relevant for your company — and which aren't.
Request a free initial call