On model choice
The model is rarely why a project failed
Teams spend weeks comparing models and an afternoon deciding what the system is allowed to do. That ratio is backwards. In every system we have put into production, the questions that decided success were about boundaries: which sources count as authoritative, what confidence is high enough to act on, who approves an irreversible step.
Swap the model in a well-bounded system and it keeps working. Put the best available model into a system with no boundaries and it will eventually tell a customer something untrue with complete confidence.
On integration
Budget for the old system, not the new one
The AI part of an AI project is usually the smallest line item. The cost sits in the twenty-year-old database whose column names lie, the ERP whose API was designed for a different decade, and the permissions nobody has documented since the person who set them left.
This is why we start read-only. Reading an old system teaches you what it actually contains — which is reliably different from what the schema claims — before anything is allowed to write back.
On trust
An audit trail is a product feature
Logging tends to get treated as an operational nicety. In practice it is what changes the conversation inside the client's organisation. Before logs, the debate is philosophical: can AI be trusted at all. After logs, it becomes specific: this reply on Tuesday was wrong, here is the source it used, let us adjust the threshold.
The second conversation is productive. The first one never ends.