Six steps. Four agents. Twelve weeks.
The method behind every engagement. Six steps from a manual workflow to a productive digital colleague on your infrastructure. Plain language, no slides, no maturity models.
From process to agent.
Each step has a clear output. You see before we start what lands on the table at the end of the phase, and decide afterwards whether to continue.
Discovery
Map twelve workflows, prioritise top three. Sift WhatsApp groups, email threads, Excel tables. Collect pain points.
Definition
Fixed-price sketch per phase. DPA template, GDPR briefing for your DPO. Specify test set, define eval metrics.
Build
Pilot workflow live. Anthropic SDK on your infrastructure. Trust ladder L1 to L4. Test-set eval, logging, monitoring.
Validate
Three weeks production eval. Document edge cases, gradually raise trust level, gather error patterns for the runbook.
Scale
Rollout to all prioritised workflows. Per-team training, escalation runbook, cost caps, on-call rotation.
Handover
Handover to your IT. Nobody needs to call me when something breaks. Optional retainer for monthly monitoring.
Should I build custom at all, or will a SaaS tool do?
Honest comparison: 10 axes, decision tree, quadrant. So you do not buy the wrong tool.
How we find the right process.
Method describes how we build. Which processes are worth it and where bottlenecks live is the Operations Diagnostic Kit's job.
How we find bottlenecks
Five bottleneck markers (H, R, W, T, A) on the L3 BPMN frame of every top-15 process.
Markers in the diagnostic kit →Which processes are worth it
Priority scoring and quadrant map. Quick-Win, Strategic, Wait, Don't-Automate.
Quadrant in the diagnostic kit →Eight anti-patterns that kill audits
Big-bang RPA, tool choice before problem choice, black-box agents in customer-facing flows.
Read anti-patterns →Bella, Sven, Rita, Klaus.
Four patterns that recur in 80 percent of Mittelstand engagements. Names are placeholders, the patterns are real.
Multi-step processes with branches. Bella decides which sub-agent handles which piece.
Reads incoming mails, tickets, receipts. Classifies, drafts replies, escalates on edge cases.
Checks contracts, receipts, reports for completeness and compliance. Lists anomalies in prioritised form.
Recurring ops tasks. Consolidate reports, keep data in sync between systems, trigger-based actions.
Four levels from read-only to autonomous agent.
You decide when the agent reaches which level. Read-only first, then suggestion, then auto-action with review, then autonomous. Escalation under uncertainty is always built in.
Agent reads data, creates reports. No action, no write rights. Usable as a briefing tool for the human.
Agent drafts or recommends. Human validates before each action. Suitable for sensitive outputs.
Agent acts autonomously. Human reviews after the fact daily or weekly. Escalation under uncertainty.
Agent acts autonomously, human only spot-checks or reviews on escalation. Maximum leverage, highest eval bar.
On your infrastructure. EU-resident.
Anthropic SDK directly, no vendor platform in between. You keep the stack, you keep the data, you keep control.
- ModelClaude Sonnet, Opus, Haiku
Directly via Anthropic SDK, no vendor in the middle. EU-resident on demand.
- RuntimeYour own infrastructure
AWS Frankfurt, Azure West Europe, or on-prem. Code in your IT's git.
- ComplianceGDPR, GoBD, ISO 27001-ready
DPA signed on day one. Audit logging, cost caps, escalation.
- HandoverRunbook + on-call
Nobody needs to call me when something breaks. Per-team training, escalation docs.
Three stages, depending on activation energy.
Not every company starts at the same level. Three entry stages, same trust ladder principle, same handover.
n8n + OpenAI Node
When you have never built an agent. Low activation energy, low trust build. Good for L1 and L2 use cases with two tools in play.
Tasklet.ai or n8n + Claude
When the first pilot runs and the second workflow is up. More control granularity, own eval, MCP bridges to DATEV or SAP.
Claude Agent SDK / OpenAI Agents SDK
When you need custom logic, multi-step orchestration and audit trail. Code in your IT git, own eval suite, on-prem option.
What you get in writing.
These three promises are in every contract. If I do not keep them, you do not pay.
Fixed price per phase
No day rate, no open end. You know before we start what you get and what it costs. If I overrun, that is my problem.
DPA on day one
EU data residency, GDPR-compliant, GoBD audit logging. Briefing note for your DPO included. Reverse-charge invoicing to DE B2B.
Full handover
Nobody needs to call me when something breaks. Runbook, per-team training, on-call rotation. No consultant lock-in, no follow-on costs.
Before you book: read the pilot promise.
Fixed price. Success criterion in writing. Last instalment only on success. 48-hour cancellation. Code ownership from week one. Solo continuity plan.