Our take
A conversational product has two jobs, and they are different work. The first is the product, the character, the quality, the behavior in the hard moments. The second is getting it to run inside systems your team does not own, under rules your client sets. Enterprise deals are won or lost on the second job, and it is usually the one nobody planned.
Connect on purpose
"We will integrate with their systems" is not a plan. Each edge is a choice, a direct API, the Model Context Protocol, a connector, or an embed, and the choice sets how much you control and how much a security team reviews. The teams that deploy cleanly pick the connection per edge, scope it to least privilege, and write down the option they rejected.
Deploy inside the rules
Where the data is allowed to live decides which model you can call. How traffic leaves the network decides where your logging and retries belong. What you can prove decides whether the tool ships at all. Residency, the gateway, the handoffs, and the compliance evidence are architecture from the first sprint, not a security review at the end. Design for them early and every new client becomes a fill-in-the-blank. Leave them to the review and each one is a rebuild.