A back office that turns large-model conversation into a product you can sell and distribute: head office owns the conversation, agents own acquisition, members pay by tier, and commission settles itself every month. Change one prompt and every answer follows.
Project background
The system splits the business across three layers: head office holds the conversation capability, agents bring in users, and members pay by tier. The front end is the chat entry point customers use; the back office is where head office runs the product — 18 modules grouped into day-to-day operations, base information and system administration, covering the whole chain from acquisition and payment through conversation to commission settlement.
Conversation lives in prompt management. Each scenario has one prompt; attach it to an aggregation ID and it goes live. To change how the assistant answers — its tone, its boundaries — you edit the prompt, with no code change and no release. Content management holds the terms, privacy policy and FAQ; sensitive-word rules guard both directions; client image management controls what the front end shows.
Acquisition sits with the agents. Each agent has a channel and a set of promo posters, and the back office records the registrations, member payments and monthly new users that agent brought in. Two leaderboards — this month and historical months — show how each region and each agent is pacing.
Payment and commission are one line. Membership cards define tiers and entitlements; every member payment is logged. Assessment rules set the monthly new-user threshold and the commission rate; once met, payroll settles it month by month, with pending, paid and cancelled states all traceable. Conversation logs keep the original exchange, so a dispute can be checked rather than argued.
System architecture: three groups, 18 entry points
Modules are split by rhythm — what you open every day, and what you configure once. Day-to-day operations is the working surface for the ops team; base information is where product and content are configured; system administration is about people and permissions.
Roles and tiers
Visitor, member, super member, agent, staff (channel) and staff agent — six identities moving through one permission model. A visitor can try the product, a member unlocks the full conversation, a super member gets a higher quota. Agents and channel staff add promotion and settlement rights on top of membership, while a staff agent lets an employee recruit downline in a personal capacity.
Role and menu management turn “who sees which menu” into a configuration item, so adding a job function needs no code. Dictionaries and parameters hold the business enumerations and global switches that get adjusted often and should never be hard-coded.
Core feature modules
Operations: customers, promotion, payment, payroll
Registered-customer management filters by agent, phone, role and sign-up date, so you can see where each customer came from and which tier they sit in. Poster management is the agent’s asset library. Member payments and agent payroll keep the books aligned, and the two leaderboards roll forward month by month.
Base information: keeping the conversation on rails
Prompts, aggregation IDs, content and sensitive words decide how the assistant answers. Prompts are grouped by scenario, aggregation IDs switch the model source, sensitive words are screened on the way in and on the way out, and content management holds the agreements and FAQ — the ops team edits copy without waiting on a release.
Administration: permissions and parameters
Users, roles, menus, dictionaries and parameters form the permission floor. A role binds menus, menus control entry points, and dictionaries and parameters hold the values that change often but should never be written into code.
Monthly settlement: clear to calculate, easy to audit
At the start of each month the system runs the assessment rules: did the agent hit the new-user threshold, which commission band applies, how much is owed — each line sits in payroll. Operations confirms, finance pays, and the status moves from pending to paid. Get it wrong and a cancelled state lets you reverse it.
The historical ranking freezes each month into a snapshot. Agents change, rates are adjusted, and no settled month is rewritten — so when the books are reconciled, the numbers match.
Use cases
Channel distribution for knowledge commerce and AI apps
Head office builds the product and the content, agents drive acquisition, and revenue splits on results. The system carries the “who brought them, and what are they owed” question, so nobody reconciles in a spreadsheet — and the data stays with the company when an agent leaves.
Membership-based AI Q&A products
Free trial, member unlock, super-member quota — three price points backed by prompt and sensitive-word control. It is how large-model capability becomes a subscription you can sell, rather than a chat box anyone can use for free.
Internal promotion across multiple channel layers
Channel staff and staff agents let people inside the company promote too, sharing the same permission and settlement logic as external agents. Management cost does not climb as layers are added.
Need a similar distribution system?
The system can be adapted to your membership tiers, agent hierarchy and revenue-share rules, or you can take just one part of it and connect it to the membership and payment stack you already run. Tell us how you split revenue and we will come back with an approach and scope.