The automation that works on Tuesdays
Someone set up a clever Zap. When an order email arrives, it pulls out the customer name, product codes and quantities, and creates the order in your system. It worked well in testing. Then a customer replied from their phone, another pasted the order as a table, a third attached a PDF instead of writing it in the body, and one supplier changed their email template.
Now the automation fails several times a week, sometimes loudly, often silently. A person checks the task history every morning, fixes the broken runs by hand, and adds another parser rule that works until the next variation. The automation that was supposed to save time has become a small job of its own.
Why rule-based automation hits a wall
Tools like Zapier, Make and n8n are excellent at moving structured data between systems. Where they struggle is reading unstructured text. Their email parsers and text formatters depend on the input looking the same each time: the order number after a certain word, the quantity in a certain position. Humans do not write like that.
Every new rule makes the flow more fragile, because rules interact and nobody remembers why each one exists. The real limit is that you are asking a pattern matcher to understand language. That is the part AI does well, and the part rules never will.
What the breakages cost
| Failure | What it causes |
|---|---|
| Silent failure | An order or request never reaches the system, and nobody notices until the customer chases |
| Wrong field extracted | Incorrect quantities or addresses flow straight into orders |
| Daily manual fixing | Someone's morning goes on repairing runs |
| Rules nobody understands | Fear of changing the flow in case it breaks further |
| Automation abandoned | The team goes back to doing it all by hand |
Silent failures are the worst of these, because the business believes the automation is handling something it is not.
How we make the flow handle real-world input
- We collect a sample of real messages and attachments, including the ones that broke the flow, and define exactly which fields the next step needs.
- We replace the brittle parsing step with an AI extraction step, calling OpenAI or Anthropic Claude through the API, that reads the email body and any attachments and returns the fields in a fixed JSON structure.
- We validate the result before it goes further: product codes checked against your catalogue, customers matched to existing records, quantities within sensible ranges, required fields present.
- Anything that fails validation, or where the AI reports low confidence, goes to an exception queue, such as a Slack or Teams message or a simple review page, with the original email attached and the fields pre-filled for a person to correct.
- Everything that passes continues through your existing Zapier, Make or n8n steps, so the rest of the automation you built stays as it is.
- We keep a test set of tricky examples and rerun it whenever the prompt or model changes, so a fix for one format does not break another.
Where volume or complexity grows beyond what the no-code tool handles comfortably, we can move the extraction step into a small service of its own and let the no-code flow call it. You keep the flexibility without the fragility.
What the team sees afterwards
Orders arrive in the system whether the customer wrote a neat email, a scrappy one, or attached a spreadsheet. The morning repair job becomes a short look at the exception queue, and every item in it is there for a stated reason.
Nothing fails silently any more. If a message could not be processed, a person knows about it, with the message in front of them.
And because the rules have gone, changing the flow is no longer frightening. When a supplier changes their template, the AI step usually copes without any change at all. When it does not, the example goes into the test set and the fix is checked against everything that already worked.
Is this your situation?
- Your Zapier, Make or n8n flows break when emails or documents vary.
- Someone checks the task history each morning to repair failed runs.
- The flow has grown a long list of parser rules nobody wants to touch.
- You have found out about missed orders or requests from customers chasing.
- You have considered switching the automation off and going back to manual.