AI automation

Begin with the workflow

I map the inputs, decision points, exceptions, and verifiable outcome before choosing a model. The current effort is recorded first, so saved hours and changes in the number of errors can be assessed after deployment.

AI is useful where fixed rules fall short but a person can still judge output quality. Precisely defined actions stay in conventional automation, while models handle tasks such as classification, summarisation, retrieval, or drafting.

Agents, memory, and tools

Complex work can be divided among autonomous agents with separate roles, permissions, and queues. A coordination layer assigns tasks, manages retries, and pauses the workflow when information is missing or a risk threshold is reached.

I design memory systems that separate temporary context from reviewed long-term facts. I improve the system using measured outcomes, reviewed examples, and controlled instruction updates. An agent is not allowed to rewrite its own instructions without supervision.

Model Context Protocol (MCP) tools give agents a consistent interface to approved functions and data sources. Every tool receives defined input and output formats, permission boundaries, and failure behaviour.

Connections to existing systems

Agents can connect to CRM records, online shops, email, document stores, and internal data through existing APIs, webhook events, or n8n workflows. I check for a secure, supported integration route before adding custom connective code.

The data path is documented from source to final action. Read operations can then be separated from mutations, with each agent receiving only the access needed for its assigned work.

Control and data privacy

Higher-risk actions include a human approval point, while sensitive fields can be filtered or anonymised. Activity records make it possible to trace which data, instructions, and tools contributed to an outcome.

Storage location and retention follow the requirements of the process. Where needed, I select EU-hosted models, databases, and automation components, while limiting the information shared with external providers.

Typical scenarios

• Classifying incoming email, producing summaries, and preparing reply drafts.
• Enriching CRM records from approved conversations or documents.
• Checking product data and preparing content for an online shop.
• Searching contracts, reports, or internal knowledge with cited sources.
• Coordinating research, analysis, and approval across several agents.
• Handing repeatable actions to n8n or another execution system.

Principles

• People retain irreversible and reputation-sensitive decisions.
• Every action keeps a clear source and an auditable trail.
• Implementation starts with a small, safe scope and real examples.
• Quality and saved time are measured before the workflow expands.
• Permissions and memory are constrained by each agent's role.
• Failures stop safely and visibly instead of being concealed.