“Done” should mean done: AI agents safely in weclapp

In one of our agent tests, an AI agent got a simple task: rank customers by revenue. It queried the invoices with the field totalPrice. In weclapp, however, the field is called grossAmount. The wrong field name was silently dropped, the agent got no amounts, didn't notice and fetched every invoice individually. After 14 calls its budget was exhausted, and there was no ranking.

A person would have paused after the first empty result. An agent doesn't pause.

That's exactly why we have AI agents play through real tasks with wals.pro AI 4 weclapp before you work with it. We know the weclapp API from our own products such as Fulfillment 4 weclapp and 4Fulfillment and from hundreds of n8n workflows for weclapp customers. This knowledge is built into the guardrails with which AI assistants like Claude or ChatGPT work in your weclapp via wals.pro AI 4 weclapp. We show three of them here, and what you can take away if you connect an agent to weclapp yourself.

Clear feedback instead of silence

Models guess field names, that will keep happening. What matters is whether anyone objects. Someone on the team looks at the screen and sees that the column is empty. An agent has no screen. It only has the response, and a response without an error is a response without a problem for it.

That's why in wals.pro AI 4 weclapp no field disappears without comment anymore. If an agent queries a field that doesn't exist, it is told so explicitly, together with a pointer to the complete field list and, as the next step, the instruction to repeat the query with the right name instead of fetching records individually. For a typo like grossAmout, wals.pro AI suggests the right name right away. And because the agent can load linked data in the same query, invoices and customers arrive in one go.

One exception is intentional: if a field points to protected data, such as stored external connections, the agent gets a fixed explanation and no suggestion on how to get there after all. Which errors the weclapp API itself reports loudly and which stay silent, we show in The silent errors of the weclapp API.

For you, this means: an agent is only as good as the feedback it gets. Anything that is silently skipped must be spoken out loud by someone.

Verified against the real weclapp

An agent writes what it thinks is right. Whether weclapp sees it the same way only shows when saving. That's why wals.pro AI 4 weclapp doesn't rely on assumptions, but on weclapp itself.

For every entity that wals.pro AI may write to, there is a fixed write contract: which fields are required, which are allowed and which fields may appear in items. If a required field is missing, such as the customer on a quote, the agent gets a clear error message. If it sends a field that the contract doesn't know, it is reported back by name.

To make sure the contracts are correct, an automatic test runs on every change: every allowed field must actually exist in the official interface description of the weclapp API v2 or be on a documented exception list. We additionally check write paths live against a weclapp test system.

For you, this means: don't trust a preview that never asked weclapp. A real test against a sandbox is worth more than any simulation.

First the preview, then your approval

An agent acts as soon as it believes it has understood the task. That's why in wals.pro AI 4 weclapp no agent writes without asking.

Every change begins with a preview: this is what I want to change, on this record, with these values. The preview goes to you, together with the instruction to the agent to execute only after your explicit confirmation. The approval is a payload-bound one-time approval via SHA-256. It is valid only briefly, only once and only for exactly this change, this record, your tenant and you. If even a single value differs from the preview on execution, it is invalid. Every write is logged.

For document items, wals.pro AI goes one step further: after writing, it re-reads the items and checks, among other things, whether their number is correct.

That way you keep control, and the AI still takes the work off your hands.

For you, this means: separate what an agent wants to do from what it does. A human belongs in between.

Know the limits and say so

Agents don't know what an API can't do. wals.pro AI 4 weclapp knows and tells the agent.

One example is bills of materials. If you create a production order in weclapp via the API for an article with a bill of materials, the items stay empty: weclapp doesn't resolve the bill of materials in the process, and there is no dedicated endpoint for it in either v2 or v3. An agent without this knowledge would consider the empty order done. In wals.pro AI it gets the recipe along: first read the article's bill of materials, then take over each component with its factor, 1:1, without multiplying itself. The multiplication is done by weclapp.

Some things we deliberately don't offer, even though the API could do them. Marking an open item as paid without a bank transaction would be technically possible. In wals.pro AI this stays blocked until it is clearly regulated how such a step has to be approved. Your accounting should be right.

For you, this means: tell your agent what doesn't work and how to do it instead. An honest “not like this, but like this” is worth more than a false “done”.

Guardrails belong in the integration, not in the prompt

All three guardrails have the same core. An agent lacks the view of the interface, and when uncertain it guesses instead of asking. A better prompt changes little about that. What helps are rules that sit in the integration: clear feedback, checks against the real weclapp, an approval before every write and knowledge of the limits.

Christian Lanig from weclapp titled his talk at Community Day 2026 “Every request counts”. For agents, that applies twice over: every request counts, and every response must be correct.

Next steps

These guardrails are built into wals.pro AI 4 weclapp. The AI never writes without asking, but with preview, approval and log. You connect your AI assistant to your weclapp and work from day one with exactly the rules this article talks about: try wals.pro AI 4 weclapp.

Are you planning your own agent integration or do you want to know how AI can concretely help in your weclapp? Then let's talk: wals.pro/termin. You'll find more about the weclapp API itself in our series “weclapp API: The Handbook”, for example in The silent errors of the weclapp API.