I use AI in the systems around Virya. Quite a lot, actually. What I do not do is give an agent access to Gmail, Sheets, Drive and every social account and tell it to "grow the band". That sounds impressive right up to the first confidently invented email address.
Give the model the fuzzy bits
The useful role turned out to be much narrower. Models are good at language and ambiguity: classifying a reply, deciding whether a contact is a festival or a radio show, summarising a long thread, translating a message or preparing a draft around facts I already verified. The deterministic workflow still owns identity, status, retries and every external action.
Constrain the answer before using it
For example, a reply classifier may return positive, negative, neutral or unclear, add one sentence of summary and suggest the next step. It cannot invent a fifth state because the email "felt promising". The workflow validates the output before it touches the CRM.
Drafting works the same way. I pass the real contact type, the reason we are writing, the material being offered, the goal and the links that must appear. AI turns that into decent English or Polish. It may improve the sentence; it may not improve reality.
The send button stays human
The send button stays with a person. So do deletion, merging verified records, rights decisions and anything that could damage a relationship. A model can propose a collaboration email, but it cannot quietly grant paid advertising rights or decide that today is the perfect day to message five hundred people.
Confidence is not evidence
I also keep facts and suggestions in separate fields. A verified address is data. An address produced by a model is a candidate waiting for verification. Confidence is not evidence, no matter how many decimal places it has.
Failure must still make sense
When the model times out, returns broken JSON or contradicts the structured input, the workflow keeps the previous valid state and puts the record into review. The complete process still has to make sense on a day when the AI provider is unavailable.
Corrections are useful data too
The human corrections are useful too. I can compare what the model suggested with what we finally decided and see which signals actually led to good conversations. That is a much better way to improve the system than simply handing the model more permissions.
So yes, AI is part of the band machinery. It handles the fuzzy bits and saves us a lot of repetitive work. People still handle taste, relationships and decisions that matter. I think this is a healthier division of labour than pretending the robot is our new manager. :)