Start with the approval, not the feature.

We nearly built the wrong automation because we started with the feature instead of the approval decision.

The scene

A business wanted to automate a workflow that ran on WhatsApp: a customer asks for something, someone on the team checks a few things, and a reply goes out with a price or a confirmation. On paper it is a perfect automation. The messages are structured. The data lives in the ERP. An AI agent can read the request, look up the record, draft the reply.

So that is what we started designing. Intent detection, record lookup, reply templates, tone. Two weeks of good engineering conversation.

Then a simple question came up in a review: who is allowed to send that price?

The answer was: it depends. Under a certain amount, anyone. Above it, a manager. For certain customers, the owner personally. None of that was written down anywhere. It lived in people's heads and in the habit of walking over to someone's desk.

The feature we were building would have sent prices nobody had approved. It would have worked perfectly and been switched off within a week.

What changed in my thinking

Most automation and AI projects do not fail at the model. They fail at the approval. The expensive question is not "can the software do this?" It almost always can. The expensive question is "who has to say yes, at which point, and what do they need to see to say it fast?"

When you start with the feature, you design the happy path and discover the decision later, usually from the person who is now angry. When you start with the approval, the feature becomes obvious — and usually smaller.

In that project, the real automation was not the reply. It was the approval: put the request, the record and a suggested price in front of the right person on their phone, one tap to approve, and only then send. The AI's job became preparing a decision, not making one.

The rule

Before you automate anything, write down the decision it replaces, who owns that decision today, and what they would need to approve it in ten seconds.

Three questions, in order:

1. What decision does this step actually contain? If the answer is "none, it is pure data movement", automate it fully. If there is a judgement inside it, keep going.
2. Who owns that decision today, and does it depend on amount, customer or risk? Draw the thresholds. They are the real spec.
3. What is the smallest thing the owner needs to see to approve quickly? That screen, message or notification is the product. Build it first.

Do this and the AI conversation gets much calmer. You stop arguing about whether the model can be trusted, because the model is not being trusted with the decision. It is being trusted with the preparation, which it is very good at.

Next step

If you are planning an AI or automation project and you are not sure where the approval decisions are hiding, that is exactly what a strategy session is for. Bring the workflow. We will find the decisions before you build the features. If this is the problem you are dealing with, see AI automation consulting in Dubai.