Many conversations about AI in mid-sized companies start with the same question: can the model take over the whole process? Our answer from production is usually no, and that is not a shortcoming. The processes we run do most of the work automatically and deliberately leave the rest to a person. That split is exactly what makes them reliable. In this article we show where to draw the line and how to build approvals that do not slow anyone down.
Where language models are reliable, and where they are not
Models like Claude are very good at four kinds of tasks:
- Classifying: Is this email an invoice, a complaint or an inquiry? Which department owns it?
- Extracting: Pull invoice number, amount, supplier and due date out of a PDF, even when every PDF looks different.
- Drafting: Prepare a reply to a customer inquiry with the right data from your systems.
- Summarising: Condense a long ticket history into three sentences before a colleague takes over.
All four have something in common: there is an input, and a person can check the result in a few seconds.
Models are not suited to final decisions with consequences. Approving a bank transfer, issuing a credit note, committing to a price, cancelling a contract. The model has no sense of responsibility. It produces plausible output, and when it is unsure it sounds just as confident as usual. Anything that moves money or cannot be undone stays with a person.
Approvals that take seconds, not minutes
A bad approval looks like this: the employee gets a notification, opens the ERP, searches for the record, opens the PDF next to it and compares by hand. That takes minutes, and after a week everything simply gets waved through.
A good approval shows everything on one screen:
- the original, for example the invoice as an image
- the extracted data, already filled into the target fields
- the model's suggestion, with a short reason
- deviations highlighted, such as an amount that does not match the purchase order
- two buttons: Approve and Reject, plus a field for a correction where needed
In n8n this is a message in Teams, Slack or email, and the workflow waits for the answer. Approving takes well under a minute. And every correction is a data point you can use to improve the process.
Confidence thresholds instead of all or nothing
Not every case needs the same treatment. We work with three lanes:
- High confidence: The case runs through automatically. This only applies to steps that are reversible and low risk.
- Medium confidence: The case goes to approval.
- Low confidence: The case lands in a manual queue, the same as today.
The threshold does not come from the model alone. We combine the model's assessment with business rules: amount above a limit, new supplier, deviation from the purchase order, customer with special terms. In the beginning the thresholds are conservative and almost everything goes to approval. Only when the data shows that the model has been right in a category for weeks do we open the automatic lane a little further.
Start with one process and measure
Pick one process, not five. Good candidates have high volume, a clear input and visibly annoy the team. Incoming invoices, customer inquiries in support, order confirmations from suppliers.
Before you automate anything, measure the current state: cycle time, handling time per case, error rate, number of follow-up questions. Without these numbers you cannot say later whether it was worth it.
Then three stages:
- Shadow mode: The model suggests, the team works as before. You compare the suggestions with the real outcomes.
- Approval mode: The model prepares, the team approves.
- Partial automation: The safe cases run through, the rest stays in approval.
After each stage you compare against the baseline. It is unspectacular, but it is the difference between a pilot project and a process.
Guardrails, cost control and GDPR
We build guardrails in from day one:
- The model only gets the tools it needs for this one process, initially without write access.
- Every output is validated against a fixed schema before it is processed further.
- Retries are capped, and every workflow has a cost budget with an alert.
You keep costs under control by choosing the right model per task. A small, fast model is enough for plain classification. The detailed analysis gets the large one. Prompt caching and batch processing lower the cost further.
On GDPR: what matters is which data leaves the house. Send only the fields the task needs, and pseudonymise names and contact details wherever the result does not suffer. Sign a data processing agreement with the provider and check its commitments on data use. Anthropic offers corresponding contractual terms for the API. Log what was sent, so you can answer data subject requests.
The goal: a process the team trusts
Trust does not come from a high accuracy figure on a slide. It comes when the team can see every decision, correct every decision, and watch the corrections flow back into the process. A fully automated system that silently makes one mistake loses that trust for a long time. A process with approval that shows the mistake before it does damage earns a little more of it every week.
Takeaway
Human-in-the-loop is not a stopgap until the models get better. It is the operating model that works in mid-sized companies: the model does the preparation, people decide, and approval costs seconds. If you want to find out which of your processes makes a good starting point, book a free 30-minute intro call at /en#kontakt.