Business workflows that become conversations.

Qualify → collect → rate → compare → correct → consent

Guide a prospective customer through a preliminary quote while code owns eligibility, referral reasons, premiums, quote IDs, re-rating, and consent. The model collects and explains; it never fabricates a price or binds coverage.

Verify → cover → compare → elect → review → submit

Verify eligibility, collect household coverage, compare plan options, validate account limits and beneficiaries, disclose pending requirements, and submit only after review. Policy code owns plan terms, limits, payroll totals, and the final enrollment record.

Discover → select → quote → confirm → commit → verify

Understand the request, check availability, capture preferences, present priced options, and book. The model conducts the conversation; the flow guarantees no booking before explicit confirmation, no double charge, and quoted terms preserved through commit.

Diagnose → retrieve → propose → authorize → execute → confirm

Understand the issue, inspect account data, propose approved actions, place approval holds where policy requires, execute, and confirm. When human help is needed, the hand-off carries the complete case record instead of a transcript dump.

Intake → classify → extract → validate → route → adjudicate

Collect incident details in conversation, receive documents, use AI for perception — classification, OCR, extraction — then apply deterministic validation and business rules in codified steps. Exceptions route to human review instead of silently passing.

Gather → verify → enrich → approve → activate

Guide setup in one conversation: gather missing information, verify identity and eligibility, call backend services, handle exceptions, and activate. Durable session state means an interrupted customer resumes mid-onboarding instead of starting over.

Capture intent → policy check → approval → order → reconcile

Capture purchase intent conversationally, apply vendor and policy rules deterministically, route approvals with holds, create the request in your systems, and track status. Policy is enforced in typed handlers — not requested of the model.

Extract → reconcile → detect anomaly → approve → post

Use AI for perception over invoices and statements, normalise into structured facts, reconcile deterministically, surface anomalies for review, and post only after approval. One-shot and batch execution use the same flow model as conversations.