Don't take the “fully automated support” promise seriously; the honest version is: automate the repeated questions so humans reach the real ones.
Step one: list the repeats. Watch a week of messages and calls. Usually ten or fifteen topics — shipping, returns, warranty, stock, how-to — make up most of the volume. That list is your widget's starting knowledge.
Step two: approved written answers. Write one precise answer per topic — the one your best agent gives. The Dorna widget answers from that knowledge, not from guesswork; so knowledge quality is the ceiling of answer quality.
Step three: the human boundary. Decide what always reaches a person: complaints, special billing cases, anything beyond the knowledge. Write into the widget's instructions what it should say for those — for example, announcing your support contact — so no customer hits a dead end.
Step four: the improvement loop. Once a week, review the Conversations page in the Studio: what was asked? Where was the answer weak? Add that topic to the knowledge. Teams that run this loop see the difference clearly within weeks — quality compounds.
And the success metric? Not “bot conversation count,” but two things: the share of questions answered correctly without human help, and the hours freed for your team.
See the practical setup in the widget install guide.