Low-fidelity wireframes for the 4 screens that carry this design's guardrail architecture —
intentionally boxy and grayscale, since this is a wireframe fidelity level, not a finished UI.
Scope note: see the
full case study (T5)
for the conversation-flow transcripts, agent architecture, evaluation approach, and risks
these screens map to.
Why these 4 screens specifically: each one exists to close a gap that shows
up repeatedly in travel support bots today — most either can't take real action at all (pure
FAQ bots) or take action with no visible reasoning, no confirmation step, and no memory once a
human takes over. These screens are the direct answer to those four failure modes, not
generic chat-UI decoration.
Screen 1
Reasoning You Can See
Maps to: Agent Architecture — intent + tool-call layer
Fixes: black-box answers
Support Copilot
"How much would I get back if I cancel?"
You'd get back ₹3,200 of ₹5,800 paid.
Intent: cancel_booking · 0.96 confidence
Why this amount? Economy Saver fare, cancelled 22hrs before departure → 55% refund per fare rule R-14. See full fare rule
Not what I expected →
Reasoning is one tap away, not buried or absent — same trust-layer pattern as Wayfare AI's "Why this?" (T2).
Most support bots today show a number with no rule behind it, so a user has no way to check it's right.
Screen 2
Confirm Before It's Irreversible
Maps to: Guardrail — confirmation gate
Fixes: silent autonomous actions
Confirm Cancellation
This can't be undone
FlightDEL → BOM, Tomorrow
Refund₹3,200 of ₹5,800
Arrives5–7 business days
Cancel
Confirm
Never fires on the first pass — every irreversible or money-moving action stops here first, no exceptions.
Industry bots today tend to over-correct one way or the other: too cautious to act at all, or acting without a visible stop-and-confirm step.
Screen 3
Context Follows You to a Human
Maps to: Agent Architecture — escalation router
Fixes: repeat-yourself handoffs
Connecting You
This needs a specialist — connecting you now. They already have everything below, you won't need to repeat it.
Handoff summary
PNR6XK92P
IssueRefund dispute > ceiling
TriedStandard refund quoted
Escalated becauseAmount over auto-approval limit
Structured handoff, not a transcript dump — the human agent opens straight into what matters.
"Please repeat your issue" after a bot handoff is one of the most common complaints about support bots today — this screen is the direct fix.
Screen 4
Audit Trail for Support Ops
Maps to: Guardrail — full audit trail
Fixes: invisible autonomous actions
Agent Action Log
issue_refund · ₹3,200OK
Confidence 0.96 · confirmed by user · 09:14
cancel_booking · PNR 6XK92POK
Confidence 0.96 · confirmed by user · 09:14
change_booking_dateREVIEW
Confidence 0.81 · fare diff. flagged · 11:02
Every autonomous action logged, not just conversations — this is what a wrong-action-rate audit (see T5's Evaluation Approach) actually reads.
Most bots today are a black box even to the ops team running them — no per-action log, only aggregate ticket-deflection dashboards.
Reading it: Screens 1 and 2 are what the traveller sees; Screens 3 and 4 are
what makes the system safe to operate at all — a handoff a human agent can act on immediately,
and a log support ops can audit. A copilot that only gets Screens 1–2 right still fails the
business the moment something needs to go to a human or get reviewed after the fact.