Inbox agents
Read, categorize, route, archive. For email, ticketing systems, document intake. Client mapping, attachment processing, automated filing. The payroll use case above is in this class.
Service · AI Agents
Specialized AI agents that do real work.
An AI agent is not a chatbot. It reads data, makes decisions, performs actions. We build agents that work in your concrete workflow, not in a demo environment.
Integrated with Microsoft 365, Google Workspace, n8n, your own systems.
Fundamentals
More than a chatbot. Less magic than you think.
A chatbot responds. An AI agent executes. It reads incoming data (email, documents, database records), reasons about the task, calls tools (APIs, search functions, file systems), and delivers a result or triggers an action. At its core, an agent is an LLM equipped with tools.
- 01
Input
Email, file, API call, schedule, webhook.
- 02
Reasoning
The LLM analyzes, plans steps, decides.
- 03
Tools
Database, search, file system, external APIs.
- 04
Output
File, response, database entry, follow-up action.
Use case · Payroll accounting
How an AI agent organizes the inbox of a payroll accountant.
A concrete example from live operations. Self-employed payroll accountant, 40 clients, approximately 80 incoming client emails per day. Before: manual filing. After: two seconds per email.
Starting point
Before: 90 minutes a day on filing.
Every morning, 50 to 100 new emails from clients. Each with attachments: invoices, payroll tax statements, sick notes, contracts. Every file has to be recognized, renamed, and filed in the right client and subfolder. Routine work that comes back every day.
What the agent does
What the agent now handles.
- Polls the Outlook inbox every 60 seconds
- Identifies the client by sender address
- Unpacks ZIP attachments where present
- Reads PDF content via OCR
- Classifies the document by subject, file name, and content
- Renames the file to a standard scheme (for example, "2026-04_Payroll-Tax-Statement_Mueller.pdf")
- Files the document into the correct OneDrive subfolder
- Logs every step in an internal, audit-grade log
- Flags unclear cases for manual review instead of guessing
Technical stack
How we built it.
n8n as the workflow engine, Microsoft Graph API for Outlook and OneDrive, Claude for classification and renaming logic, Tesseract for OCR. All self-hosted, because payroll data is especially sensitive. No data leaves the controlled environment more than necessary.
Result
What measurably changed.
Instead of 90 minutes of daily filing work, now ten to fifteen minutes of review on the few unclear cases. Over 95 percent of emails run end-to-end. Filing errors have almost disappeared since go-live, because the agent names consistently and does not get tired. The freed time goes into onboarding new clients.
Scaling
Discuss a similar agent for your companyWhat this enables.
The agent runs today for one person with 40 clients. The architecture and code are built so the same logic is directly usable for three or ten more payroll accountants. The effort per new accountant: a few hours of configuration, not a rebuild.
Agent types
Four classes of agents we run in production.
Research agents
From 50 sources to a structured briefing in under 10 minutes. Market analyses, competitor updates, supplier research. The agent searches defined sources and delivers a consolidated result in the format you want.
Classification agents
Categorize and route tickets, inquiries, documents automatically. Above 95 percent accuracy, validated on your data. Saves manual triage and shortens response times.
Multi-step workflow agents
Multi-stage processes in which the agent moves data between systems. Fetch, enrich, validate, file. Example: quote generation from a brief, CRM, and product catalog.
How we work
From the first use case to a production agent in four to eight weeks.
- Step 01
Sharpen the use case
Two workshops with end users and IT. Which task exactly, which inputs, which outputs, which edge cases. The output: a precise spec.
- Step 02
Prototype
First working agent in two weeks, tested on historical data. You see early how the agent behaves and can iterate quickly.
- Step 03
Integration and hardening
Connection to your systems, error handling, logging, monitoring, rollback. A production agent must run cleanly even when the data arrives dirty.
- Step 04
Go-live and handover
Training for your admins, documentation, handover protocol. After go-live we shadow the first two weeks of live operation intensively, then transition to normal operations.
What we build with
We take what works.
Model and tool choice follows the use case. We are not an Anthropic shop or an OpenAI shop. The task decides which component goes where.
Reasoning models
- Anthropic Claude
- OpenAI GPT
- Google Gemini
Workflow engines
- n8n
- Make
- Custom implementations
Vector and storage
- Supabase
- PostgreSQL
- Pinecone
- Cloudflare D1
Frameworks
- LangGraph
- LangChain
- Vercel AI SDK
Hosting
- Cloudflare Workers
- Vercel
- On-premises
Integration APIs
- Microsoft Graph
- Google Workspace
- CRM systems
Data protection per GDPR. Self-hosting available where needed.
Right fit
This service fits you when:
Fits you
- You have a clearly bounded workflow with high volume.
- The task is rule-based or follows recognizable patterns.
- Your data is digitally available (email, file system, database, API).
- You want to relieve your people, not replace them.
- You are ready to invest in setup and hardening, not only in a demo.
Does not fit you
- You are looking for a general "AI assistant" without a specific task.
- The workflow is highly variable and creative (agents need structure).
- Your data exists only on paper or in people's heads.
- You expect 100 percent automation from day one.
- You want a chatbot, not an operational tool.
Frequent questions
Answers to questions we hear often.
- A simple workflow agent (for example the inbox agent above) starts at roughly EUR 8,000 to 15,000 net for the initial build. More complex multi-step agents with connections to several systems are typically between EUR 20,000 and 60,000. Plus ongoing hosting and model costs from EUR 50 per month.
- We build agents so that unclear cases are routed to a review queue instead of being guessed. That is the most important safeguard. In the payroll example, about 5 percent of emails land in the review folder daily instead of the automatic path.
- It depends on the setup. For sensitive use cases like payroll, we host everything ourselves (n8n on your server, local OCR). External LLM APIs are used only where it makes sense and is covered by a data processing agreement. When required, we can run entirely on local models.
- On typical workflows we reach accuracies between 92 and 98 percent after tuning. The remaining cases land in review. Accuracy is not binary, it depends on data quality and input variety. We measure this on your real data before go-live.
- We build agents so the reasoning model is swappable. Claude today, GPT tomorrow, Gemini after that, without rebuilding the workflow. This model independence is an architectural principle for us.
- Yes. We deliver code and workflows so your IT or another vendor can take them over. No proprietary black boxes, no lock-in. Clean documentation, ideally in n8n or standard frameworks.
- For operating, mostly no, a technically minded admin is enough. For continued development, someone who understands Python or JavaScript basics helps. If you do not have that, we take on maintenance as a service.
What pairs well
The agent rarely shows up alone.
Next step
Which recurring task in your company would you most like to never do again?
Tell us about it. We will tell you honestly whether an agent is the right path and what it would cost.
Discuss your use case