Services
AI & Software
Solutions
Training
Tools
Insights
References
Contact
Free initial consultation Ask a first question by email
About Deutsch
← AI & Software

Automation with approval

Recurring flows as workflows or agents with limited tools: preview, approval, audit log. Writing actions do not happen without a person agreeing.

Friday, 4:40 pm, the bulk mail goes to forty recipients, and the subject line is wrong. Anyone who has had to clean up something like that understands the concern about write access immediately. An agent you may allow to send needs three things: a preview, an approval and a log. Without them, every automation is an open window.

How I work: with source · with approval · in your system.

Starting point

Many flows are recurring and rule-based: capture invoices, build reports, move data between systems, pre-sort requests. The work is not hard, but it takes time and does not shrink as the business grows. At the same time, the concern about giving an AI write access grows. That concern is justified: an agent with access to a mailbox and a recipient list is, at worst, an open relay.

What I do

I build the flow as a workflow or as an agent with clearly limited tools. Each tool has a purpose, an allowed scope and a boundary. The agent can do only what it actually needs.

Before every writing action there is a preview: what will be written, where, with which values. A person confirms. Only then does a checkable automation execute, not the language model itself. The confirmation is bound to a token valid for that single operation, so a second click does not fire twice. Credentials sit in the automation, not in the AI tool. Recipients and targets get an allowlist, access gets a rate limit.

Every run is logged: time, operation, result, without sensitive content. That makes it traceable what happened if a dispute arises.

Result

You get running workflows or agents, operating documentation and the rules for approvals. Add the list of tools with their boundaries. What must not be automated is in there too.

Example

Invoice → posting → DATEV, in use. In my own ERP the service reads the document, suggests a posting from historical entries, and accounting reviews and approves it. Only then is the posting batch created, in the EXTF format. The suggestion speeds things up; the decision stays with people. Details are under posting matrix and assignment.

Report chain to an automation, demo. The request starts in the chat: “every Monday an actuals report as a spreadsheet to this address”. The assistant shows the request as a preview; after approval a job description goes as JSON to an n8n webhook, n8n acknowledges, builds the file and sends the mail. That chain is built and testable, but it does not run in production. The report rows are still placeholders, and the rhythm from the chat is currently driven by a fixed schedule rather than by the sentence. So I call it a demo chain, not a deployment.

DevLink, a related case. DevLink goes from a work package to a coding session on the developer machine and writes the resulting AI cost to the ticket. At the end of a session a hook reports back only token counts and metadata; no prompts, answers or transcripts leave the developer machine. Price changes do not apply retroactively, because each session stores its price version. Same idea as above, with cost instead of postings: make it measurable without giving up content.

Limit

A chain is only as good as its interfaces. Where the existing system has no clean API, it becomes either an adapter or a bad automation. And an agent that makes decisions with judgement does not belong in automation. Such steps stay out.

Then the unpleasant side: a webhook without protection is an open mail relay. Anyone who knows the URL can trigger reports to arbitrary recipients. Header authentication, a random path, a recipient allowlist and a rate limit belong before go-live, not after.

How it runs

  1. Capture the flow and define the approval point
  2. Define tools and limits, separate credentials
  3. Build preview and approval, design the log from the start
  4. Test with real cases, including failures
  5. Hand over operation, check allowlist and limits

Common questions

n8n or custom software?

n8n for flows that can be described as a chain and change often. Custom software when the logic is tightly coupled to the existing system or needs special rights.

What stays with people?

Every decision with judgement and every action that cannot be undone. Approval is not a formality; it is the safety mechanism.

How do you stop the agent from sending nonsense?

With a preview of the concrete values, an allowlist for recipients and a rate limit. Every run is also logged. Without these four, sending is open.

Can it be extended later?

Yes, if the tools are limited individually. New abilities arrive as a new tool with its own approval, not as broader access.

The next step costs nothing.

30 to 60 minutes, no strings attached: where is the biggest friction, what is the most effective lever, and is a collaboration a fit?