Before it submits something, I would want to review the details, know who will receive them and choose whether to proceed.
Adding an approval button is only useful if the person understands what they are approving.
And “done” should come with something I can verify.
I am evaluating AI‑agent prototypes that prepare or submit government‑related documents. My focus is on making the meaning of “done” explicit: the system must show what was prepared, who will receive it, and whether the user approved the final submission.
These are prototypes with simulated integrations.
Samarthan explores preparing a cybercrime complaint from a voice message.
JanSewak proposes guiding people through government websites and forms.
RailSetu looks at options involving seat changes on the same train when the full route is waitlisted.
My view: trust and human control will decide adoption.
Decision: I will adopt a review‑first workflow so that no AI agent can submit a document without a visible user review step. I also require explicit verification of success or failure, surfacing status, error messages, and corrective actions.