Skip to content

End-to-End Workflow

The workflow splits into a synchronous intake phase (user waits) and an asynchronous execution phase (user polls/receives updates). Phase B is the critical human gate — no execution occurs without explicit agent confirmation.

End-to-end workflow — detailed flow
Click any phase to explore
Phase A

Intake

The synchronous phase where the user's request is processed end-to-end before any response is returned. LLM extracts intent and parameters, deterministic services resolve the affected cards and policy, and a confirmation plan is generated for the agent to review.

A1Receive natural-language requestAssist
A2LLM extracts intent + parameters into typed JSON schemaAssist
A3Deterministic services fetch cards, policy, approval requirements, recipientsAssist
A4Generate plan preview, risk summary, exclusion reasons — present to agentAssist → Human confirmation
What happens
Agent types: "Update all Marketing cards to SGD 2,000 limit". LLM extracts operation=limit_update, team=Marketing, new_limit=SGD 2000. Card service returns 50 cards; policy check flags 2 as requiring Compliance approval (current limit > SGD 1,500).
Output
Plan preview: 46 eligible, 2 blocked (locked), 2 pending approval
01Ambiguous target group

Cannot resolve to a definite, bounded card set.

"Update cards for the new team" — no team name; system cannot determine which cards.
02Missing critical parameters

No limit amount or currency specified; clarification required.

"Raise the limit for Finance cards" — no new limit value; pipeline cannot continue.
03Policy conflict, no override path

Requested action is hard-blocked by policy with no self-serve override.

"Update limits for KYC-held accounts" — Policy §3.1 blocks this without Compliance sign-off.
04Repeated schema validation failure

LLM cannot produce valid structured output after two clarification rounds; offered structured form fallback.

Ambiguous currency (SGD vs. USD) still unresolved after follow-up — system escalates rather than guessing.