AI automation for businesses

Choosing the model first is like choosing the truck brand before knowing what you're hauling. The task says what has to be read, what must never be made up and which mistake is serious. A model that writes well but invents an order number is useless for that task, however slick the demo.

The scope names the inbox, the document type or the exception. “We want to use AI” isn't a scope. “Incident emails arrive classified with a draft reply, and nobody publishes without reviewing it” is. If it can't be explained with a real example from last week, we don't automate yet.

The example is the contract: this email yes, this email no, this field gets filled, this field is left empty on purpose. Without that sentence, the workflow fills gaps with inventions and the team stops trusting it by the third time.

An email comes in: client, site, symptom, and sometimes an order number that doesn't match. The workflow classifies it with the categories support already uses, not new “smarter” ones. It extracts the fields the ticketing system accepts and leaves empty the one that isn't in the message. Inventing the site to complete the record is a failure, not help.

The support agent sees the draft, the original email and which field was left empty. They accept, correct or discard. Until they accept, nothing is written to the system of record. The log keeps what came in, what was proposed and what was accepted. Without that screen, the automation is a black box and security is right not to want it.

The final click is automated only when the rule is boring and the cost of a mistake is accepted: archiving an exact duplicate, not issuing an invoice or closing an incident. If the mistake is expensive, the person stays in the loop. That sentence is written before the workflow is switched on, not after the first scare.

System passwords aren't pasted into a prompt or shared in a group chat. The workflow uses defined access, the minimum for the task, revocable and explainable. If it has to be cut off tomorrow, it's cut off without searching for which conversation held the key.

Permissions don't grow “because we already had access”. If the task goes from classifying to writing in the ERP, access is discussed again. IT and security see that decision on the same list as everything else that's switched on: what data goes out, what data never goes out and where the log is.

Switching models in a year shouldn't mean rebuilding the business logic or the credentials. The task and the rules belong to the company. The model is a replaceable part. The workflow code and its configuration are handed over in your name so that switch is never held hostage.

The workflow relies on what the system really allows. If the ERP only exports a file overnight, the automation works at that pace and doesn't promise real-time magic. Checking that limit is part of the scope, not a surprise in week four.

Nor do we ask for a data lake “to be ready”. The first inbox uses the fields of that inbox. Extending to another team is another engagement, once the first one has been used for a full week without going back to the old way.

We don't build an assistant with general access to emails, folders and personnel files “to see what people ask”. That access can't be explained, can't be revoked per task and mixes data that was never meant to be seen together. If the question doesn't fit into a task with fields, it isn't an automation: it's a hope.

Nor do we write unattended into billing, payroll or anything where a mistake costs money or a relationship. There, AI can prepare. Publishing stays with a person. And we don't train anything on personal data “because the model would come out better”. If we can't say what is stored and for how long, it isn't switched on.

Before building, we measure how much time the task takes and how often it goes wrong by hand. Afterwards, we check whether the team uses it or has gone back to the old way. Keeping a workflow alive out of pity is how companies pile up automations nobody dares touch that still hold a live credential.

The first workflow is deliberately narrow. One type of email, one type of document, one team. When that workflow has been used for a full week, we talk about the next one. The map of “the whole company on AI” is a slide deck, not a project.