Why our automation only suggests at first

The biggest risk of automation with AI isn't that it won't work. It's that it will work almost every time. A language model that answers ninety-nine emails correctly and on the hundredth promises a customer a discount the company doesn't give is worse than no automation - because after ninety-nine correct answers everyone stopped paying attention to it. That's why we deploy in four steps, and the automation has to earn its autonomy.

1. Review: what exactly should happen

Before we build anything, we go through the process with the people who do it today. Not as it's described in the company guidelines, but as it actually runs - including the exceptions nobody mentioned because they handle them automatically. The result is a description of inputs, outputs and rules, and an estimate of how many hours the process costs today. If the review shows that automation doesn't pay off, we say so.

2. Shadow mode: it runs, but sends nothing

The automation receives real data - the same emails, orders and comments as the team - and prepares outputs. But it doesn't send or change anything. Its suggestions are just saved next to what the person did. After a few weeks you can see where it hits the mark and where it gets things wrong. The customer knows nothing about it, and nothing in operations can break.

3. You approve

Once shadow mode shows that the suggestions make sense, the automation starts working for real - but every important step goes through a person. A finished draft reply or post arrives on your phone with Send and Edit buttons. Approving takes a few seconds instead of ten minutes of writing. And every edit is feedback: it is used to tune the rules and materials the automation works from.

4. Autonomy step by step

For types of tasks where the person has consistently just confirmed without changes, approval is switched off. Not all at once - step by step. Replies to positive reviews can run on their own, while negative reviews and complaints still go for approval. What gets autonomy is decided by results from operations, not by a promise at deployment. And if something changes, it's enough to switch approval back on for that step.

What it looks like in practice

This is exactly how the marketing automation runs at Fotobazoš. The system prepares social media posts from the catalog, sorts comments, reviews and messages, and writes draft replies. The owner approves them on her phone, and autonomy is switched on step by step as individual types of tasks prove themselves. Details are in the case study.

Isn't that slower?

On paper yes - weeks pass between launch and full autonomy. In reality it's the faster route, because nothing has to be pulled back and fixed. An automation that sends a customer nonsense in its first week usually ends up switched off, and nobody gives it a second chance. An automation that built trust gradually stays in operation and gradually takes over more and more work.