Over the past two and a half years we've used AI across nearly every layer of client work: coding assistants for development, AI-powered tools for managing timelines, communication, emails and documentation, and lately, clients vibecoding parts of what used to be entirely the job of developers. On top of that, AI is often no longer just how we build the product, it's inside the product we hand over, which changes what clients expect from us and how we manage those expectations. The do's and don'ts run through three stages. First, encounters with AI, where judgement and ownership stay with the person. Second, workflows, where one-off chats become something that repeats, and where vibecoding, nondeterminism and constantly moving model capability change how we work with internal software. Third, and where most of the difficulty sits, scaling it across a company: what can be shared and what cannot, which software this actually ends up replacing, and where the line between a domain expert and engineering is moving right now.