给澳洲移民与留学服务公司定制 AI 员工团队：挂在事务所自己品牌下的永远在线前台、客户门户与智能后台。由执业中的注册移民代理设计，合规长在结构里。
Custom AI staff teams for Australian migration and education practices: a branded, always-on front desk, client portal and smart back office. Designed by a practising registered migration agent, with compliance built into the structure.
澳洲的移民与留学服务公司同时被三样东西压着：接不完的询价、追不完的材料与期限、一年比一年重的合规要求。人手是最大的成本，也是最大的瓶颈。Subclass AI 为每家事务所定制交付一套三层 AI 系统，把前台接待、案件跟进、期限监控、政策盯梢这些吃人力的活交给 AI 员工团队，持牌代理只做判断和批准。少雇一个前台或案件助理，一年省下的工资是月费的好几倍。
Australian migration and education practices are squeezed by three things at once: enquiries they cannot answer fast enough, documents and deadlines they chase all day, and a compliance load that grows every year. Staff are the biggest cost and the biggest bottleneck. Subclass AI custom-builds a three-layer AI system for each practice, handing intake, case chasing, deadline monitoring and policy watching to an AI staff team, so the registered agent does only judgement and approval. One admin salary saved covers the retainer several times over.
我们卖的不是软件订阅，是给事务所配一个 AI 员工团队：一次性搭建费加月费托管。护城河有两条。第一条是 harness：针对移民业务流程定制的工具、记忆、规则和分工。第二条是身份：创始人自己就是执业中的注册移民代理（MARN 1800784），这套系统先在自己的事务所里跑通了，卖的是每天在用的东西，不是 PPT。
We don't sell a software subscription. We sell an AI staff team fitted to the practice: a one-off build fee plus a monthly retainer. The moat is double. First, the harness: tools, memory, rules and division of labour custom-built for migration workflows. Second, identity: the founder is a practising registered migration agent (MARN 1800784) who runs this exact system in his own practice daily. We sell what we use, not a deck.
挂事务所自己域名的品牌官网、中英双语结构化询价表单、付费咨询预约与收费公示。只收集情况、只答事务性问题，不给签证意见。7 乘 24 接住每一条线索，包括周六晚上那条。
A branded website on the practice's own domain, bilingual structured enquiry forms, paid-consult booking and a published fee schedule. It collects circumstances and answers practice questions only, never visa advice. Catches every lead, 24/7, including the Saturday-night ones.
申请人免密码登录（magic link），查案件进度、上传材料、看缺件清单与付款状态、留言提问。被专人照顾的体验，同时把追材料的电话省掉大半。
Passwordless magic-link portal where applicants track case progress, upload documents, see the missing-item checklist and payment status, and leave questions. The "someone is looking after me" experience, and most document-chasing phone calls disappear.
线索收件箱（AI 已起草回复待批）、案件状态看板、期限与有效期台账、政策变化预警、应收提醒、每日晨报。持牌代理手机上一屏看全事务所。
Lead inbox with AI-drafted replies awaiting approval, a case-state board, a deadline and expiry ledger, policy-change alerts, receivables reminders and a daily brief. The whole practice on the agent's phone.
三层不是三个产品，是一条流水线：Front Door 抓线索，HQ 转化成案件，Client Lounge 服务到下签好评。每一层都由同一套定制 harness（知识包加规则加工具）驱动。
Not three products but one pipeline: Front Door captures the lead, HQ converts it into a case, Client Lounge carries it through to decision and a five-star review. All three layers run on the same custom harness (knowledge pack plus rules plus tools).
| Core | Plus | Premium | |
|---|---|---|---|
| 月费Monthly | A$399 | A$699 | A$999 |
| 一次性搭建费One-off build fee | A$2,000 | A$3,000 | A$5,000 |
| 三层系统Three-layer system | 全部包含Included | 全部包含Included | 全部包含Included |
| AI 对话额度AI conversation allowance | 500 次/月500 / mo | 1,500 次/月1,500 / mo | 4,000 次/月，含专属语气训练4,000 / mo plus custom voice and tone |
| 自动化深度Automation depth | 询价接待、咨询预约、晨报Lead intake, consult booking, daily brief | 增加缺件催收、报价与咨询跟进、评价邀请Adds document chasing, quote and consult follow-ups, review invitations | 增加政策监控预警、递前核对清单、SMS/WhatsApp 多渠道、与现有行业软件对接Adds policy monitoring alerts, pre-lodgement checklists, SMS/WhatsApp channels, integrations with existing practice software |
| 修改与支持Changes and support | 每月 1 次调整1 change per month | 每月 2 次，优先支持2 changes per month, priority support | 不限次，季度业务复盘Unlimited, quarterly business review |
这一节做两件事：先像报价单一样逐项算清 AI 员工替掉了哪些重复人工、省下多少钱，再讲三种人各自怎么用。所有触点都设计成零学习成本，没有人需要看说明书。
This section does two things: first it itemises, like a quote, exactly which repetitive work the AI staff remove and what that saves; then it walks through how three kinds of people use the system. Every touchpoint is designed for a zero learning curve; nobody reads a manual.
这套系统的核心不是多一个工具，是把事务所里天天重复、不需要专业判断的人工活成建制地拿掉，把员工成本变成可控的月费。下表以一家 2 名持牌代理加 1 名行政助理、在办约 60 案的事务所为例。数字为典型估算，交付第 1 周按贵所实际情况逐项校准，之后每月运营报告持续复核。
The core of this system is not one more tool. It is the systematic removal of the repetitive, judgement-free manual work inside a practice, converting staff cost into a controlled retainer. The table below models a practice of two registered agents plus one admin assistant with about 60 active cases. Figures are typical estimates, calibrated line by line against your practice in week one of delivery and re-checked in every monthly report.
| 重复性工作Repetitive task | 通常谁在做Done today by | 每周耗时（估）Hours/week (est.) | AI 接管后，人只剩What humans keep | 省下Saved |
|---|---|---|---|---|
| 首次询价接待与背景信息收集First-contact intake and background collection | 前台与助理Reception, admin | 6 | 看一页结构化摘要Read a one-page summary | 5 |
| 咨询预约、确认、改期与提醒Consult booking, confirmations, reschedules, reminders | 前台Reception | 2.5 | 无Nothing | 2.5 |
| 报价单制作与发送Quote preparation and sending | 助理加代理Admin plus agent | 2 | 过目批准Review and approve | 1.5 |
| 报价与咨询后的跟进追单Follow-ups after quotes and consults | 助理Admin | 3 | 批准发送Approve the send | 2.5 |
| 签约建档与材料清单准备Case setup and document checklists on signing | 助理Admin | 2 | 无Nothing | 2 |
| 材料催收、收件核对与登记Document chasing, checking and logging | 助理Admin | 5 | 处理不合格材料Handle non-conforming items | 3.5 |
| 进度询问电话与主动进度更新"How is my case going" calls and proactive updates | 前台与助理Reception, admin | 5 | 个别例外沟通（门户自动更新）Rare exceptions (the portal updates itself) | 4.5 |
| 期限、成绩与 EOI 有效期台账Deadline, test-score and EOI expiry ledger | 助理与代理Admin and agent | 2 | 按预警行动Act on alerts | 2 |
| 政策新闻浏览与筛选Policy news scanning and filtering | 代理本人The agent | 3 | 读分级简报Read the graded brief | 2 |
| 应收提醒与催款Receivables reminders | 老板或助理Owner or admin | 1.5 | 批准催款函Approve the reminder | 1 |
| 晨间案件盘点与工作梳理Morning case round-up and planning | 代理本人The agent | 2.5 | 读晨报Read the morning brief | 2 |
| 完案好评邀请与收尾杂务Review invitations and closing chores | 助理Admin | 1 | 无Nothing | 1 |
| 合计Total | 35.5 | 约 6 小时判断与批准About 6 hours of judgement and approval | 约 29.5 小时/周About 29.5 h/week |
上表省的是行政工资。下表省的是持牌代理本人的时间：录入信息核对、一致性检查、时长计算、eligibility 初筛，这些活现在全靠代理肉眼一行行看。AI 的做法是预检并标红，代理只审标红项，结论永远由代理确认后才使用，对外不出任何自动判定。
The first table saves admin wages. This one saves the registered agent's own time: data verification, consistency checks, duration calculations and eligibility pre-screens, all currently done line by line with the agent's eyes. The AI pre-checks and flags; the agent reviews only the flagged items, every conclusion is confirmed by the agent before use, and nothing is auto-decided externally.
| 核查工作Review task | 每周耗时（估）Hours/week (est.) | AI 预检后，代理只剩After AI pre-check, the agent keeps | 释放Freed |
|---|---|---|---|
| 客户资料录入核对与前后一致性检查（多份表格间的地址、姓名拼写、日期矛盾）Data entry verification and cross-form consistency (addresses, name spellings, date conflicts across forms) | 4 | 只审标红的矛盾项Review flagged conflicts only | 3 |
| 伴侣签同居与关系时间线梳理（证据与陈述对齐、断档识别）Partner-visa cohabitation and relationship timeline (evidence aligned to claims, gaps identified) | 2.5 | 确认拼好的时间线与断档Confirm the assembled timeline and gaps | 2 |
| 工作经验时长计算与重叠核对（推荐信、工资单、职责期间逐段对齐）Work-experience duration calculations and overlap checks (references, payslips, role periods aligned) | 2.5 | 复核算好的时长与缺口Verify computed durations and gaps | 2 |
| 材料合格性初筛（对照清单查缺漏、格式与不合格件）Document eligibility pre-screen (against checklists: missing, malformed, non-conforming) | 3 | 处理例外件Handle exceptions | 2.5 |
| 偏远地区读书加分核查（就读时长加校区 postcode 对照现行法定名单）Regional study points verification (study duration plus campus postcode against the current legislative instrument) | 1.5 | 确认附来源的预查结论Confirm the sourced pre-check | 1 |
| 递交前整包一致性终查Final pre-lodgement package consistency check | 2 | 终审三档结论报告Final review of the three-grade report | 1.5 |
| 合计Total | 15.5 | 约 3.5 小时终审About 3.5 hours of final review | 约 12 小时/周About 12 h/week |
换算成钱，两本账加在一起：行政层每周净省约 29 小时，相当于 0.76 个全职编制；按现在的行情，案件助理年薪 A$65,000 加 super（12%，全成本约 A$72,800），一年省约 A$55,000。代理层每周释放约 12 小时核查工时，按保守的 A$150 咨询时薪、一年 46 周计，是约 A$80,000 的持牌产能，用来多接案或者早点回家。对比 Premium 档全年 A$11,988 加一次性搭建 A$5,000：光行政这本账，第一年就净省约 A$38,000，第二年起 A$43,000 以上，代理产能还没算进去。而且这个员工不请假、不离职、不用重新招聘培训，周六晚上照常上班。这就是报价的依据，也是客户算得清的账。
In dollars, both ledgers together: the admin ledger saves about 29 hours a week, roughly 0.76 of a full-time role. At today's market rate of A$65,000 plus 12% super (about A$72,800 all-in) for a case assistant, that is around A$55,000 a year. The agent ledger frees about 12 review hours a week; at a conservative A$150 consulting rate over 46 weeks, that is about A$80,000 of licensed capacity, spent on more cases or an earlier night. Against the Premium tier at A$11,988 a year plus a A$5,000 build: the admin ledger alone nets about A$38,000 in year one and A$43,000 plus from year two, before counting agent capacity. And this staff member takes no leave, never resigns, needs no re-hiring or retraining, and works Saturday nights. That is the basis of the quote, and it is a sum any practice owner can check.
设计原则：不做聊天机器人。我们刻意不在官网放聊天气泡，很多人反感网页 chatbot，看到就关；在移民行业它还是合规雷区。AI 全部藏在后台（分诊、摘要、起草、盯期限、写晨报），前台只有工具和人：结构化表单、一问一答、回电承诺。没有聊天机器人，这件事对事务所同时是体验卖点和合规卖点。
Design principle: no chatbot. We deliberately ship no chat bubble; plenty of people see one and close it, and in migration it is also a compliance minefield. The AI stays entirely backstage (triaging, summarising, drafting, watching deadlines, writing the daily brief) and the front of house is tools and humans: structured forms, a one-shot question box, and a call-back promise. "No chatbot" is a selling point to a practice on both experience and compliance grounds.
Subclass AI 卖的 harness 长什么样，不用想象。我自己的持牌移民事务所 Ozsky International 已经在 2026 年 8 月投产运行同一套结构，代号 Ozsky Engine。它既是内部生产线，也是产品本体：卖给移民事务所的，就是这套结构换上对方品牌和数据的版本。
What the Subclass harness actually looks like is not hypothetical. My own licensed migration practice, Ozsky International, has been running the same structure in production since August 2026 under the name Ozsky Engine. It is both the internal production line and the product itself: what a migration practice buys is this same structure carrying their brand and their data.
选移民代理业务做第一个实现，是刻意的。这个行业的合规压力、期限密度、错误代价，都高于普通中小企业：一条政策引错就是合规事故，赔的是客户的身份和事务所的执照。能在这里扛住的结构，卖回给同行不需要翻译，第二梯队再迁去会计与律所也只是换名词。
Proving it inside a migration practice first was deliberate. This industry carries more compliance pressure, denser deadlines and higher error costs than a typical SME: one wrong policy citation is a compliance incident, and the price is a client's status and the practice's licence. A structure that survives here sells back to the same profession without translation, and transfers to accounting and law firms in the second wave by renaming the nouns.
人类主席，唯一法律责任人与批准人
Human chairman, sole legal authority and approver
CEO agent，唯一交互入口
CEO agent, the single interaction surface
职能经理 agent
Functional manager agents
底层执行 skill
Underlying worker skills
每一层只跟相邻层说话。主席不点 skill，经理不越级，执行层不做判断。这条约束是整个系统可控的前提，也是可以卖给别人的部分。
Each layer talks only to its neighbours. The chairman never invokes a skill directly, managers never skip levels, and workers never make judgement calls. That constraint is what keeps the system controllable, and it is the part that transfers to other businesses.
主席只说目标，不说该叫谁。CEO agent 收到任何任务都走同一个五步协议，中间过程对主席透明，但不需要他参与。
The chairman states a goal, not a tool. The CEO agent runs the same five-step protocol on every task. The process is visible to the chairman but requires nothing from him.
这套系统没有独立数据库。每个案子的笔记既是人类阅读的工作记录，也是 agent 的状态存储。Agent 干活前先读笔记，干完把新结论写回同一份笔记。人和机器共用同一份真相，出问题时人能立刻看到系统以为的事实是什么。
There is no separate database. Each case note is simultaneously a human-readable work record and the agent's state store. An agent reads the note before acting and writes its conclusions back into the same note. Human and machine share one source of truth, so when something goes wrong the human can immediately see what the system believed.
等谁、下次跟进日、我方还欠什么。过了跟进日仍无动静的自动升回紧急。这条规则消灭了大部分沉默的漏案。
Who we are waiting on, the next follow-up date, and what we still owe. Any Pending item past its follow-up date escalates back to Urgent automatically. This single rule kills most silent drift.
各类成绩、评估、签证到期日与影响打分的生日全部登记并设提醒窗口。这类东西没人盯就会静静过期，代价是整条路径重来。
Test results, assessments, visa expiry dates and the birthdays that change a points score are all registered with reminder windows. Left unwatched these expire silently, and the cost is starting over.
同一个客户第二次被问起时，先读笔记再干活，干完追加回同一份笔记。系统越用越省力，而不是每次从邮箱重新翻一遍。
The second time a client comes up, the agent reads the note first and appends afterwards. The system gets cheaper to run over time instead of re-reading the inbox every round.
让生产者自己检查自己，是这类系统最常见的失效方式。所以验证总监被刻意设计成不参与生产、不给面子、绕开 CEO 直接向主席汇报。这一条对任何要靠 AI 对外说话的生意都成立。
Letting the producer check its own work is the most common failure mode in systems like this. The verification role is therefore built to produce nothing, defer to nobody, and report to the chairman directly. This holds for any business that lets AI speak on its behalf.
邮件、报价单、说明函，任何要出门的东西一律只出稿。发送和递交动作永远由持牌人本人执行，agent 没有发送权限。
Emails, quotes, letters: anything leaving the practice exists only as a draft. Sending and lodging are always performed by the licensed agent in person. No agent holds send authority.
任何政策结论必须带官网来源链接和查询日期。媒体与社区消息只当线索，未回官网核实的不能进入行动级预警。
Every policy conclusion carries an official source link and the date it was checked. Media and community chatter are leads only, and cannot reach action-level alerts without official verification.
政策引用是否现行、数字是否对得上、有没有保证成功一类的过头措辞、引用的规则是否适用于这个客户的具体情形。
Whether the cited rule is current, whether the numbers reconcile, whether there is over-promising language, and whether the rule actually applies to this client's circumstances.
一份强制清单管住破折号、状态图标、箭头、数学符号一类的机器痕迹。客户和雇主一眼看出是 AI 写的，专业度就打折，这是真实成本。
A mandatory list strips the machine tells: em dashes, status icons, arrows, mathematical symbols. Readers who spot AI-written text discount the sender immediately, and that is a real cost.
| 节奏Cadence | 执行者Runs as | 做什么What it does |
|---|---|---|
| 工作日 9:30Weekdays 9:30am | Trump | 晨报：全库状态巡检、当日与次日日程、近 24 小时需动作的邮件、昨日情报预警，输出最多五条今天必须动的清单，每条附建议承办人 Morning brief: status sweep, today and tomorrow's calendar, the last 24 hours of actionable email, and yesterday's intelligence alerts, condensed into at most five must-do items each with a suggested owner |
| 每日下午Daily afternoon | Murdoch | 扫官方政策源与媒体社区二级源，自动比对在手案件，分级归档，喂给次日晨报 Sweeps official policy sources plus media and community secondary sources, cross-matches live cases, files a graded brief, and feeds the next morning's brief |
| 每周Weekly | Cook | 收件箱周清，逐条搬到对应笔记，真正清空，有期限的同步进总览 Inbox clearing: each item routed to its note, the inbox actually emptied, deadlines synced to the index |
| 每月Monthly | Buffett | 经营复盘，四类金额汇总并交叉核对收费一致性，出该催谁与该报价谁的清单 Revenue review across four buckets, cross-checked against the fee schedule, producing who-to-chase and who-to-quote lists |
| 事件驱动Event-driven | Cook | 任何案件讨论结束自动归档：更新笔记、判定状态、同步总览 Any case discussion ends in automatic filing: note updated, state assessed, index synced |
降级设计：定时任务触发时本地设备通常是离线的，此时 agent 自动改读云端快照，并在产出开头标注基于哪一天的快照。任何一路数据源不可用都不阻塞出报，只标注今日缺失，继续跑完剩下的。这是把可靠性做进流程，而不是指望每台机器都在线。
Graceful degradation: when a scheduled job fires the local device is usually offline, so the agent falls back to the cloud snapshot and stamps the brief with which day's snapshot it used. No single unavailable source blocks the run; it is flagged as missing and the rest completes. Reliability is built into the process rather than assumed from the infrastructure.
要求使用者记住十几个工具叫什么名字、什么时候该用哪个，等于把调度工作退回给人。收敛到一个 CEO agent，使用者只需要会描述目标。
Asking a user to remember a dozen tool names and when to use each hands the dispatching job back to the human. Collapsing to one CEO agent means the user only needs to describe the goal.
Agent 的状态存在人能直接打开阅读和修改的地方，而不是黑箱数据库。信任来自可检查。
Agent state lives where a human can open, read and edit it, not in an opaque database. Trust comes from inspectability.
独立、绕开调度层直接汇报、只报不改。这个岗位的价值不在于它做了什么，而在于它能拦下什么。
Independent, reporting past the dispatch layer, reporting without editing. Its value is not what it produces but what it stops.
移民不是一个可以随手放 AI 客服的行业。谁能提供 immigration assistance 是 Migration Act 划的线，AI 用得对不对是 OMARA 盯的点。Subclass 移民版把这些规则做成产品结构，而不是合同末尾的免责条款。这一节同时是给潜在客户和合伙人的答案：为什么这个产品必须由持牌代理来设计。
Migration is not an industry where you casually drop in an AI chat agent. Who may give immigration assistance is a line drawn by the Migration Act, and how AI is used is a point OMARA now actively polices. The migration edition of Subclass turns these rules into product structure rather than a disclaimer at the bottom of a contract. This section is also the answer to why this product has to be designed by a practising registered agent.
Migration Act s276 把 immigration assistance 定义为运用移民程序知识提供建议、准备申请或代理；s280 限制只有注册代理、执业律师和豁免人员可以提供，无注册提供属违法，OMARA 明确 AI 生成的信息不构成任何豁免。所以 Front Door 只做三件事：结构化收集情况、回答事务性问题（收费、流程、材料清单）、预约付费咨询。这属于法条明确排除的 clerical 与信息层面。任何触及签证判断的问题，系统的唯一动作是把它变成一次咨询预约。
Section 276 of the Migration Act defines immigration assistance as using knowledge of migration procedure to advise, prepare applications or represent; s280 restricts it to registered agents, legal practitioners and exempt persons, and OMARA has confirmed AI-generated information enjoys no exemption. So Front Door does exactly three things: structured intake, practice questions (fees, process, document lists) and paid-consult booking, all within the clerical and informational carve-outs. For anything touching visa judgement, the system's only move is to turn it into a consult booking.
OMARA 2026 年 3 月发布的 AI 使用指引明确：代理对 AI 参与产出的移民协助承担全部责任，信息不准不因出自 AI 而免责。Subclass 的架构天生就是这个姿势：AI 永远只出稿，系统里不存在自动对外发送，发出与递交由持牌代理本人执行，每一步谁起草、谁批准都留痕。别家把人工审核当成本，我们把它当产品的骨架。
OMARA's guidance on AI use (published March 2026) is explicit: the agent remains fully responsible for immigration assistance produced with AI, and inaccuracy is not excused by its AI origin. Subclass's architecture is natively in that posture: the AI only ever drafts, no automatic outbound sending exists in the system, sending and lodging are performed by the registered agent in person, and every step records who drafted and who approved. Others treat human review as overhead; we treat it as the skeleton.
同一份指引要求：把客户个人信息录入 AI 平台前必须取得书面同意，并应事先向客户说明 AI 的使用方式。Subclass 交付时自带两样东西：委托协议里的 AI 使用披露与同意条款模板，以及系统里的同意状态记录。再加一层技术保险：送进模型前姓名与联系方式先脱敏替换，模型端根本没见过真名。
The same guidance requires written consent before a client's personal details are entered into an AI platform, and advance explanation of how AI is used. Subclass ships with both: an AI disclosure and consent clause template for the service agreement, and consent status recorded in the system. Plus one technical layer on top: names and contact details are pseudonymised before anything reaches the model, so the model never sees the real identity.
Code of Conduct 的保密义务加上 OMARA 关于云存储的执业指引，问的是同一组问题：数据存在哪、谁能访问、怎么控制、泄漏了怎么通知。Subclass 的标准答案：每事务所独立数据库、独立密钥、独立 AI 预算，字段级加密，运维访问走留痕的临时授权，数据主权归事务所，随时可整体导出。这些不是加购项，是移民版的出厂配置。
The Code's confidentiality duty plus OMARA's practice guidance on cloud storage ask one set of questions: where is the data, who can access it, how is that controlled, what happens on a breach. Subclass's standard answers: an isolated database, key and AI budget per practice, field-level encryption, logged temporary authorisation for any operational access, data ownership with the practice and full export at any time. None of this is an add-on; it is the factory configuration of the migration edition.
Code of Conduct s56(5) 要求客户档案在最后动作后保存 7 年，OMARA 可书面要求调阅。Subclass 的案件档案不可静默删除，保留期自动计算，OMARA 调阅时一键导出完整案件时间线（谁在哪天做了什么、批了什么）。合规义务在别处是负担，在这里是一个卖点：审计来了，你是全行业最不慌的那家。
Section 56(5) of the Code requires client files to be kept for seven years after the last action, and OMARA can demand access in writing. In Subclass, case files cannot be silently deleted, retention is computed automatically, and an OMARA request is answered with a one-click export of the full case timeline: who did what, and who approved it, on which day. Elsewhere this duty is a burden; here it is a selling point. When the audit comes, you are the calmest practice in the industry.
2026 年 4 月 1 日新规生效，OMARA 编制扩为三倍，制裁明显加速（部长公告口径：2021/22 以来已制裁 61 名代理），年度 CPD 与职业伦理培训要求加码。合规成本上升，恰恰是事务所愿意为带合规护栏的系统付费的原因。本节法规引用核对于 2026 年 8 月 6 日，以 OMARA 与立法官网最新公布为准。
New rules took effect on 1 April 2026, OMARA's staffing has been tripled, sanctions have visibly accelerated (61 agents sanctioned since 2021/22 on the minister's count), and annual CPD and ethics training requirements have been raised. Rising compliance cost is precisely why practices will pay for a system with the guardrails built in. Citations in this section were checked on 6 August 2026; the latest OMARA and legislation publications prevail.
栈（刻意选到成本近零、无供应商锁定）：
The stack, chosen for near-zero cost and zero lock-in:
原则一句话：让大多数交互根本不碰模型；碰模型的用最便宜的档；每一层都有硬顶。成本必须是设计出来的确定数，不是用出来的惊喜数。
The principle in one sentence: most interactions never touch the model, those that do use the cheapest tier, and every layer has a hard ceiling. Cost must be a number we designed, not a surprise we discover.
这是移民版最要命的风险，所以整个第 06 节都在管它：工具端零意见、触发即转咨询、每份产出归属于事务所的持牌代理。上线前请独立移民法律师对产品文案与合同做一次书面审阅，费用进搭建成本。
The existential risk of the migration edition, which is what all of section 06 manages: zero opinions from the tool, advice triggers become consult bookings, every output is attributed to the practice's registered agent. Before launch, an independent migration lawyer reviews the product copy and contracts in writing; the cost sits in the build budget.
模型涨价就换档，分层路由本来就可换模型；托管方变政策，静态架构一天可迁移。教训来自 Manus：永远不把客户资产押在别人的订阅上。
Tiered routing means models are swappable if prices rise, and a static architecture means hosting is portable in a day. Lesson learned from Manus: never park a client's assets inside someone else's subscription.
对外承诺（结果、费用、时限）一律 AI 起草、持牌代理确认；问答框零签证意见。这正是第 05 节验证岗与第 06 节合规结构解决的问题。
All external commitments (outcomes, fees, timing) are AI-drafted and agent-approved, and the question box gives zero visa opinions. This is exactly what the verification role in section 05 and the compliance structure in section 06 exist to solve.
我们卖的是有人负责的定制交付，不是工具。本地、上门、说人话，这是大厂不做的脏活。
We sell accountable, local, done-for-you delivery rather than a tool. On the ground, in person, in plain language: the unglamorous work platforms will not do.