怎么估算合规投入How to estimate compliance effort

合规投入是可拆的工作量，不是一笔神秘的开销。 本篇只算机制：哪些动作要人做、做几轮、多久重做一次。 本站不采价格数字，也不替你填单价——单价属于你的合同与薪酬口径。 算不清的根因通常是工作项没拆，而不是不会乘。

Compliance effort is countable work, not one mysterious overhead line. This guide computes mechanisms only: which actions need people, how many rounds, how often they repeat. No prices are recorded here, and this site does not fill in unit rates for you — those belong to your contract and payroll basis. When an estimate is unclear, the usual cause is unsplit work items rather than arithmetic.

• 要为一次 AI 功能上线准备预算与人力计划。
• 被问「合规要多久、要几个人」，但答不出依据。
• 已有数据流梳理与评测记录，想把它们接进预算讨论。
• 供应商递来认证材料，你要估自己这边还要做多少。

不适合：把估算当成法律义务的判断。义务范围要走当期法规与合同，本页不作法律结论。

• You need a budget and staffing plan for shipping one AI feature.
• Someone asks "how long and how many people", and you cannot show the basis.
• You already have a data-flow map and evaluation records and want them to feed the budget conversation.
• A vendor hands over certificates, and you must estimate what your side still has to do.

Not a fit: treating an estimate as a legal-obligation judgement. Scope follows current law and contract; this guide reaches no legal conclusion.

1. 一句产品定位：给谁用、做什么任务、是否涉及关键决策。
2. 一份数据流梳理表，含未知项（没有就先做，见对应指南）。
3. 现有控制清单：权限、日志、评测、护栏各自的状态。
4. 你自己的计量口径：人日、按次、按周维护，选一种并写清。
5. 一个可复核的假设清单，每条假设标明是示例还是依据。

1. One sentence of product framing: who uses it, for which tasks, whether it touches consequential decisions.
2. A data-flow map including its unknown cells (make one first — see the matching guide).
3. An inventory of existing controls: permissions, logs, evaluation, guardrails, each with a status.
4. Your own unit of measure: person-days, per-run, per-week upkeep — pick one and state it.
5. A checkable assumption list, each item marked as illustrative or as sourced.

义务档次会影响前三项的量级。欧委会的 AI Act 政策页说明该法案按四档风险划分， 并写明禁止性实践自 2025 年 2 月起适用（2026-11-06 直抓该页面）。 具体到你的产品落在哪一档、承担什么义务，需要按当期条文与你的业务判断，本页不作结论。

框架文件帮的是结构，不是范围。 NIST 的两个官方页都把自己描述为供组织自愿使用的工具： AI RMF 面向管理 AI 风险，Privacy Framework 面向管理隐私风险。 自愿不等于零成本——按它补结构同样消耗上面四项。

Risk tier changes the scale of the first three items. The European Commission AI Act page describes four risk levels and states that prohibited practices apply from February 2025 (page fetched 2026-11-06). Which tier your product falls into, and which duties follow, needs current text plus your business facts; this guide draws no conclusion.

Framework documents help with structure, not with scope. Both NIST pages describe themselves as voluntary tools: the AI Risk Management Framework for AI risk, the Privacy Framework for privacy risk. Voluntary does not mean free — building to their structure consumes the same four items.

1. 定范围：只算这一版要上线的功能，别把全公司合规计划混进来。
2. 列条目：把要求翻成你系统里的具体控制点，一条一行。
3. 标现状：每条标「已有 / 部分 / 没有」，依据写页面或配置位置。
4. 估工作量：对「没有」与「部分」的两类填轮次与人时，假设标明为示例。
5. 加维护：按你的发布节奏设定周期，写上触发重做的条件。
6. 列不确定：把会大幅改变结果的假设单列，标明需要什么信息才能收敛。

第 3 步省下不少时间。多数争论来自「你以为已经有了」。 工具侧的维护成本有明确开关。Microsoft Purview 文档指出： 血缘抽取可以关闭。扫描范围与频率要按计划设置。历史血缘要定期清理。 这些是工作量杠杆，不是价格——按你自己的账号与数据规模估。

1. Fix scope: cost only the feature shipping in this version; do not fold in a company-wide programme.
2. Enumerate lines: translate requirements into concrete control points in your system, one per row.
3. Mark status: label each line as present, partial or missing, citing the page or configuration location.
4. Estimate effort: fill rounds and hours for the partial and missing lines, marking assumptions as illustrative.
5. Add upkeep: set the cadence from your release rhythm and state what triggers a redo.
6. List the uncertainties: separate the assumptions that swing the result, and say what information would settle them.

Step 3 saves real time, because most disputes come from "I thought we already had that". Tool-side upkeep has explicit switches: the Purview documentation states that lineage extraction can be disabled, that scan scope and frequency need planning, and that historical lineage requires periodic cleanup. Those are workload levers, not prices — estimate them against your own accounts and data volume.

数字全为示例假设，用来说明算法，不代表任何厂商报价或任何团队的真实人天。

• 输入：一个给内部运营用的报表助手，会读客户邮件正文，只输出汇总，不直接改数据。
• 处理：按四项拆解（示例假设）—— 差距分析：32 个控制点，每条核对约 0.25 人日，合计 8 人日； 证据生产：交付物 6 份（梳理表、日志样例、评测记录、策略说明、权限表、复核记录），返工一轮，合计 12 人日； 复核轮次：法务、安全、业务各一轮，每轮 2 人日，合计 6 人日； 维护周期：按每季度一次重做，单次 3 人日，年 12 人日。
• 结果判断（示例）：首轮合计 26 人日，年维护 12 人日。 其中首要不确定项是「客户邮件正文是否构成需要额外义务的数据类别」—— 该条一旦判定为是，证据生产与复核轮次都要重算。
• 失败后下一步： 把该不确定项转成两条行动：向法务确认数据类别判定，向供应商确认其处理者角色； 在结论前，预算给出两种情形（是 / 否）而不是取中间值假装确定； 结论落定后更新假设清单与核验日期，保留一次估算的可追溯记录。

单价一律由你自己填：把「人日」换成你的口径再乘。 本页不给金额，也不假设汇率、税费或折扣。

All figures are illustrative assumptions that show the arithmetic; they are not vendor quotations or any team's real person-days.

• Input: an internal reporting assistant for operations staff. It reads customer email bodies, outputs only aggregates, and does not modify records.
• Processing, split into four items (illustrative assumptions): Gap analysis: 32 control points at about 0.25 person-days each = 8 person-days; evidence production: 6 deliverables (flow map, log sample, evaluation record, policy note, permission table, review record) with one rework round = 12 person-days; review rounds: legal, security and business, 2 person-days each = 6 person-days; upkeep: one redo per quarter at 3 person-days = 12 person-days per year.
• Judging the result (example): 26 person-days for the first pass, 12 per year for upkeep. The decisive uncertainty is whether customer email bodies form a data category that triggers extra duties — if they do, evidence production and review rounds both have to be recomputed.
• What to do after a failure: Turn that uncertainty into two actions: confirm the data-category call with legal, confirm the processor role with the vendor; before the answer lands, present two cases (yes / no) instead of a fake midpoint; once settled, update the assumption list and the verification date, keeping the estimate traceable.

Unit rates are yours to enter: convert person-days into your own basis and then multiply. This page gives no amounts and assumes no exchange rate, tax or discount.

• [ ] 每项工作量都能指出它对应哪个控制点或哪份交付物
• [ ] 所有假设标明「示例」或「依据某页面」，两类不混写
• [ ] 至少一条不确定项写明收敛所需的信息与责任人
• [ ] 维护项写清触发重做的条件，而不是只写一个周期
• [ ] 估算里没有出现任何具体价格数字，单价留给填表人
• [ ] 复核者能按你的表重算出同一个结果

• [ ] Every work item points to a specific control line or deliverable
• [ ] Assumptions are labelled illustrative or sourced, never mixed in one line
• [ ] At least one uncertainty names the information and the owner needed to resolve it
• [ ] Upkeep states what triggers a redo, not only a cadence
• [ ] The estimate contains no price figures; unit rates stay with the person filling it in
• [ ] A reviewer can recompute the same totals from your table

• 估完比实际低很多：多半漏了返工轮次，把「格式统一」也计成工作项。
• 两份估算是两个数量级：检查范围是否一个只算本版、另一个算了全公司计划。
• 被要求「给个准数」：给出区间与决定性假设，别用单点数字假装精确。
• 认证材料被当成省钱依据：管理体系认证说明流程存在，不替你省去证据生产（详见读第三方报告指南）。
• 工具开关没计成本：扫描频率、血缘抽取、日志保留都会带来维护工作量。
• 管理员配置项分散：企业版的角色、保留与审计设置常分布在多个界面， 按当期管理员指南列清单再估人时（站内既有来源，2026-11-03 复核）。

• Actual cost lands far above the estimate: rework rounds are usually missing, and format alignment is a real work item.
• Two estimates differ by an order of magnitude: check whether one covers this version and the other covers the whole programme.
• Asked for "one firm number": give a range plus the decisive assumptions; a single point is false precision.
• Certificates treated as savings: a management-system certificate shows a process exists; it does not remove your evidence work.
• Tool switches uncounted: scan frequency, lineage extraction and log retention all generate upkeep work.
• Administration scattered: enterprise roles, retention and audit settings often span several surfaces, so list them from the current administrator guide before costing the hours (existing in-repo source, rechecked 2026-11-03).

1. 把估算表与数据流梳理表对齐，两边条目号能互相指认。
2. 将证据交付物清单接进发布检查，避免临时补材料。
3. 需要向管理层汇报时，只报区间与关键假设，不报单点。
4. 每季度复核一次：义务档次、工具开关与发布节奏是否变了。

1. Align the estimate table with the data-flow map so row numbers reference each other.
2. Feed the deliverable list into release checks, so evidence is not assembled at the last minute.
3. When reporting upward, present ranges and decisive assumptions rather than a single figure.
4. Review quarterly: have the risk tier, tool switches or release rhythm changed.

2026-11-06 直抓了四处官方来源：

• 欧委会 AI Act 政策页
• NIST AI RMF 页
• NIST Privacy Framework 页
• Microsoft Purview Data Map 文档

第五条为站内既有来源，沿用 2026-11-03 的复核记录，本次未重抓。

以下未知：

• 你的产品落入哪一档义务、需要履行哪些具体条款——需要按当期法规文本与业务事实判断，本页不作法律结论。
• 各法域的执行时间表是否会调整——欧委会页面的表述有日期，本页未跟踪后续修订。
• 团队真实人日消耗——只有你自己的记录能给出，示例数字不构成参考基线。
• 供应商认证范围覆盖哪些服务与主体——需读其证书与适用性说明，本次未核。

On 2026-11-06 four official sources were fetched directly: the European Commission AI Act policy page, the NIST AI RMF page, the NIST Privacy Framework page and the Microsoft Purview Data Map documentation. The fifth is an in-repo source that keeps its 2026-11-03 recheck record and was not refetched today.

Unknowns:

• Which obligations apply to your product — this needs current legal text plus your business facts, and no legal conclusion is drawn here.
• Whether enforcement timetables in your jurisdiction shift — the Commission page carries dates, and later amendments were not tracked.
• Your team's real person-day consumption — only your own records can give that, and the illustrative figures form no baseline.
• Which services and legal entities a vendor certificate covers — the certificate and its scope statement would need reading.