Cost Benefit Analysis Template for AI ProjectsAI 项目成本效益分析模板：从投入到审批决策

Quick answer: what belongs in a cost benefit analysis template?快速答案：成本效益分析模板应包含什么？

This page provides the fields, formulas, review rules, and workflow for an AI data or automation project. It is deliberately broader than a basic ROI spreadsheet: a positive ratio is not useful if the options are incomparable, control costs are missing, time savings cannot be reused, or the proposed workflow fails a quality or security threshold.

本页提供适用于 AI 数据或自动化项目的字段、公式、复核规则与工作流程。它有意比基础 ROI 表格更完整：如果方案不可比、控制成本缺失、节省时间无法再利用，或拟议工作流未达到质量与安全门槛，那么一个正向比率并没有决策价值。

Use the template when a choice has material trade-offs当选择存在重大权衡时使用该模板

Use cost-benefit analysis when decision-makers must compare feasible ways to achieve an outcome and when costs, benefits, timing, risk, and uncertainty differ materially across those choices. It is appropriate for a pilot, platform purchase, workflow automation, data modernization, build-versus-buy decision, production expansion, contract renewal, or project redesign.

当决策者需要比较实现某项结果的不同可行方式，且这些方式在成本、收益、时点、风险与不确定性方面存在重大差异时，应使用成本效益分析。它适用于试点、平台采购、工作流自动化、数据现代化、自建与采购选择、生产扩展、合同续签或项目重构。

Organize the model as ten connected registers把模型组织为十个相互连接的登记表

Whether you implement the template in a spreadsheet, planning system, or financial model, separate the inputs into registers that can be reviewed independently and reconciled through stable IDs. Every cost and benefit should point to an option, period, source, owner, and assumption. Summary outputs should reference those registers rather than contain manually typed totals.

无论使用电子表格、规划系统还是财务模型，都应把输入拆分为可以独立复核、并通过稳定 ID 核对的登记表。每项成本与收益都应指向方案、期间、来源、负责人和假设；汇总输出应引用这些登记表，而不是手工输入总数。

Start the template with the decision, not the solution模板应从决策开始，而不是从解决方案开始

A model cannot be evaluated if the question keeps moving. Record the decision owner, approval date, affected workflow, population, time horizon, currency, price basis, funding constraint, objective, quality threshold, risk threshold, exclusions, and alternatives that remain open. State what happens under business as usual and when delay begins to have a cost.

如果问题不断变化，模型就无法被评价。应记录决策负责人、审批日期、受影响工作流、对象、时间周期、币种、价格基础、资金约束、目标、质量门槛、风险门槛、排除项与仍开放的备选方案。同时说明维持现状会发生什么，以及延期从何时开始产生代价。

Use one comparison frame for every option所有方案使用同一个比较框架

Measure business as usual with the same population, output definition, quality threshold, time window, and cost basis that will be used for future options. For an AI-assisted data workflow, baseline fields may include annual volume, eligible task categories, active labor by role, review and correction time, cycle time, queue time, first-pass acceptance, incidents, external spend, capacity constraints, and unmet demand.

测量维持现状时，应使用与未来方案相同的对象、输出定义、质量门槛、时间窗口与成本基础。对于 AI 辅助数据工作流，基线字段可包括年度任务量、合格任务类别、各角色实际人工、评审与修正时间、周期时间、排队时间、首次验收率、事件、外部支出、产能约束与未满足需求。

Include at least business as usual, process improvement without new AI, a limited controlled pilot, and scaled implementation when each is feasible. Do not give the preferred option a longer benefit horizon, lower quality standard, broader demand pool, or more optimistic adoption rule than its alternatives.

在各方案可行时，至少应包含维持现状、不新增 AI 的流程优化、有限受控试点与规模化实施。不能让首选方案拥有更长收益周期、更低质量标准、更广需求池或更乐观的采用规则。

Make every assumption traceable and testable让每项假设都可追溯、可测试

A number without provenance is not an input; it is an unsupported assertion. Give each assumption an ID and record the statement, value, unit, option, period, source, observation date, population, owner, confidence, affected calculation, sensitivity, validation method, due date, and status. Avoid hiding assumptions inside formulas because reviewers cannot see what changed.

没有来源的数字不是输入，而是缺乏支持的断言。应为每项假设分配 ID，并记录陈述、数值、单位、方案、期间、来源、观察日期、对象、负责人、置信度、受影响计算、敏感性、验证方法、截止日期与状态。不要把假设隐藏在公式中，否则复核者无法看到发生了什么变化。

Capture the complete incremental lifecycle cost记录完整的增量生命周期成本

Record costs caused by each option relative to the counterfactual, across implementation, operation, change, control, and exit. For each line, include cost ID, option, category, description, cash or internal resource, quantity, unit, unit price, period, one-time or recurring status, escalation rule, tax basis, uncertainty range, source, owner, and linked risk or assumption.

应记录每个方案相对反事实而新增的成本，覆盖实施、运营、变更、控制与退出。每一行应包含成本 ID、方案、类别、说明、现金或内部资源、数量、单位、单价、期间、一次性或持续性、价格升级规则、税务基础、不确定范围、来源、负责人，以及关联风险或假设。

Record benefits by mechanism, realization, and owner按机制、实现方式与负责人记录收益

A benefit register should explain how an output changes an outcome. Record benefit ID, option, affected population, baseline, future value, unit, eligible share, adoption, attribution, utilization, ramp-up, duration, evidence source, confidence, owner, first realization date, financial treatment, overlap with other benefits, and the operational metric that will verify realization.

收益登记应解释输出如何改变结果。应记录收益 ID、方案、受影响对象、基线、未来值、单位、合格比例、采用率、归因、利用率、爬坡、持续时间、证据来源、置信度、负责人、首次实现日期、财务处理、与其他收益的重叠，以及用于验证实现情况的运营指标。

Place every cost and benefit in the period it occurs把每项成本与收益放在实际发生期间

Choose monthly periods for short pilots or uneven contracts, quarterly periods for staged programs, and annual periods only when timing within the year does not affect the decision. Each period should calculate incremental cost, realized benefit, net cash flow, cumulative cash flow, and discounted value. Do not place full steady-state benefits in the first period if implementation, training, adoption, or operational learning creates a ramp.

短期试点或合同金额不均匀时使用月度期间，分阶段项目使用季度期间；只有当年内时点不影响决策时才使用年度期间。每个期间都应计算增量成本、实际收益、净现金流、累计现金流与折现价值。如果实施、培训、采用或运营学习形成爬坡，不能在第一个期间就计入全部稳定收益。

Use a small set of reconciled formulas使用一组可相互核对的公式

Keep inputs separate from calculated cells and protect formulas from manual override. Every metric should reconcile to the same cash-flow schedule. Define whether the model uses nominal or real values, whether tax is included, how partial periods are treated, and which discount rate finance has approved.

应把输入与计算单元分开，并保护公式不被手工覆盖。所有指标都必须核对到同一现金流计划。应定义模型使用名义值还是实际值、是否含税、如何处理不完整期间，以及财务批准了哪个折现率。

A metric can be mathematically correct and still misleading if its inputs use different scopes. Reviewers should trace the numerator and denominator back to the same option, population, currency, price basis, and horizon. Never compare three-year benefits with first-year cost or use gross time saving while recurring control work is excluded.

即使指标在数学上正确，如果输入范围不同，它仍然会误导。复核者应把分子与分母追溯到同一方案、对象、币种、价格基础与周期。不能用三年收益与一年成本比较，也不能使用毛节省时间却排除持续控制工作。

Keep risk, control cost, and non-monetized impact together把风险、控制成本与非货币影响放在一起

For each material risk, record cause, affected outcome, likelihood, consequence, preventive control, detective control, owner, response, control cost, residual exposure, monitoring signal, and stop threshold. Link the control cost back to the cost register and link any probability-adjusted expected loss to the benefit or risk-adjustment calculation.

对每项重大风险，应记录原因、受影响结果、可能性、后果、预防控制、检测控制、负责人、响应、控制成本、剩余暴露、监控信号与停止门槛。控制成本要链接回成本登记，任何概率调整后的预期损失则应链接到收益或风险调整计算。

Do not force every important effect into money. Record quality, privacy, fairness, security, continuity, employee impact, supplier dependency, environmental effect, and distribution across affected groups with a scale, direction, evidence, owner, and threshold. An option that fails a non-negotiable requirement should not win because of a high ROI.

不要强行把所有重要影响货币化。应以等级、方向、证据、负责人和门槛记录质量、隐私、公平、安全、连续性、员工影响、供应商依赖、环境影响以及受影响群体之间的分布。如果方案未达到不可妥协要求，就不能因为 ROI 高而胜出。

Test downside, base, upside, and switching values测试下行、基准、上行与转换值

Start with high-impact, low-confidence assumptions: eligible workload, adoption, total human time after change, usable capacity, error reduction, implementation duration, recurring usage cost, control workload, and benefit ramp. Define downside and upside values from evidence ranges rather than arbitrary percentage changes.

应从高影响、低置信度假设开始：合格工作量、采用率、变更后总人工、可利用产能、错误减少、实施周期、持续用量成本、控制工作量与收益爬坡。下行与上行值应来自证据范围，而不是随意设置百分比。

Show the recommendation under every scenario. If a small plausible change turns approval into rejection, the correct next step may be a representative pilot or another measurement—not a larger commitment. Record which evidence would most reduce decision uncertainty and how much it costs to obtain.

应展示每个情景下的建议。如果一个合理的小变化就会把批准变为拒绝，正确的下一步可能是代表性试点或进一步测量，而不是更大承诺。还应记录哪项证据最能降低决策不确定性，以及获取该证据需要多少成本。

Summarize the model in one reviewable decision table用一张可复核决策表汇总模型

The final summary should make comparison possible without hiding the supporting registers. Include each option's objective performance, total cost, realized benefit, net benefit, ROI, payback, NPV or BCR when appropriate, cash requirement, key non-monetized impacts, residual risks, evidence confidence, dependencies, and recommendation. Use the same metric definitions across all columns.

最终摘要应使方案可以直接比较，同时不能隐藏支撑登记表。应包括每个方案的目标表现、总成本、实际收益、净收益、ROI、回收期、适用时的 NPV 或 BCR、现金需求、关键非货币影响、剩余风险、证据置信度、依赖与建议。所有列必须使用相同指标定义。

Rankings are not sufficient. An option must first meet the objective and non-negotiable constraints. The preferred option should optimize value in a balanced judgement that considers monetized and non-monetized effects, affordability, delivery, risk, uncertainty, and evidence—not merely the highest BCR or ROI.

排名本身并不充分。方案首先必须满足目标与不可妥协约束。首选方案应在货币与非货币影响、可负担性、交付、风险、不确定性与证据之间形成平衡判断，而不是只选择 BCR 或 ROI 最高的方案。

Write a conditional recommendation that can be enforced写出能够执行的条件式建议

Record the preferred option, amount, funding source, accountable owner, approved period, required controls, expected benefits, unresolved assumptions, evidence to collect, next review date, scale conditions, and stop conditions. A staged approval can authorize discovery, pilot, limited production, and scale separately, allowing evidence quality to rise before commitment becomes difficult to reverse.

应记录首选方案、金额、资金来源、负责人员、批准期间、必要控制、预期收益、未解决假设、待收集证据、下次复核日期、扩展条件与停止条件。可以把调研、试点、有限生产与规模化分别审批，使证据质量在承诺难以逆转之前逐步提高。

Fill the template in evidence order, not spreadsheet order按证据顺序填写，而不是按表格顺序填写

1. Lock the decision frame.锁定决策框架。 Agree on scope, objective, counterfactual, options, horizon, and finance conventions. 对范围、目标、反事实、方案、周期与财务规则达成一致。
2. Measure the baseline.测量基线。 Reconcile operational, quality, demand, risk, and cost data with stable definitions. 使用稳定定义核对运营、质量、需求、风险与成本数据。
3. Create the assumption register.建立假设登记。 Expose sources, confidence, owners, tests, and affected calculations before forecasting. 在预测前公开来源、置信度、负责人、测试与受影响计算。
4. Build cost and benefit registers.建立成本与收益登记。 Use unique IDs, remove overlaps, and link every line to an option and period. 使用唯一 ID、消除重叠，并把每一行连接到方案与期间。
5. Schedule cash flow.安排现金流。 Model implementation, ramp, steady state, renewal, contingency, and exit when they occur. 按实际时点建模实施、爬坡、稳定、续约、应急与退出。
6. Reconcile financial metrics.核对财务指标。 Calculate net benefit, ROI, payback, NPV, and BCR from the same timed series. 从同一分期序列计算净收益、ROI、回收期、NPV 与 BCR。
7. Test uncertainty and risk.测试不确定性与风险。 Run evidence-based scenarios, calculate switching values, and apply constraints. 运行基于证据的情景、计算转换值并应用约束。
8. Review across functions.开展跨职能复核。 Finance, workflow, data, security, privacy, legal, procurement, and delivery owners challenge the model. 由财务、工作流、数据、安全、隐私、法律、采购与交付负责人质询模型。
9. Issue a versioned decision.发布带版本的决策。 Record approval, conditions, evidence gaps, next gate, monitoring, and model-refresh triggers. 记录批准、条件、证据缺口、下一决策门、监控与模型更新触发条件。

Keep one source of truth from forecast to realization从预测到收益实现保持单一事实来源

Give the model a version, owner, review status, approval history, change log, source index, and protected calculation layer. Record who may edit inputs, formulas, or decision thresholds. Preserve the version that supported each approval; do not silently overwrite it when actual results become available.

为模型设置版本、负责人、复核状态、审批历史、变更日志、来源索引与受保护的计算层。记录谁可以编辑输入、公式或决策门槛。当实际结果出现时，不得静默覆盖支持原审批的版本。

After implementation, add actual cost, usage, adoption, quality, incidents, time, and benefit realization beside the forecast. Explain variance by mechanism, update the remaining periods, and revisit the decision when a threshold or dependency changes. Evaluation is not an afterthought; its data and resource requirements belong in the original template.

实施后，应在预测旁加入实际成本、使用、采用、质量、事件、时间与收益实现情况。按机制解释偏差、更新剩余期间，并在门槛或依赖变化时重新审视决策。评估不是事后补充，其数据与资源要求应从一开始就纳入模板。

Avoid template errors that create false confidence避免制造虚假信心的模板错误

Turn the completed registers into an ROI scenario把已完成登记转化为 ROI 情景

Before opening the InfiniSynapse ROI Calculator, prepare total investment, recurring cost, expected realized benefit, horizon, and scenario assumptions from the template. Use separate base, downside, and upside cases. Keep the detailed source, owner, confidence, timing, and overlap controls in your working model because a calculator summarizes economics; it does not replace evidence governance.

打开 InfiniSynapse ROI 计算器前，应从模板准备总投资、持续成本、预期实际收益、周期与情景假设，并分别建立基准、下行与上行情景。详细的来源、负责人、置信度、时点与重叠控制仍应保留在工作模型中，因为计算器只负责汇总经济性，不能替代证据治理。

Move from template to calculation and approval从模板继续走向计算与审批

Frequently asked questions常见问题

Sources, scope, and adaptation notes来源、范围与调整说明

This template is an original organizational worksheet structure informed by official appraisal, cost-estimating, business-case, and AI risk guidance. It is not a substitute for the accounting, tax, legal, procurement, safety, or public-sector appraisal rules that apply to a specific decision.

本模板是一套原创的组织工作表结构，参考官方评价、成本估算、商业论证与 AI 风险指南。它不能替代适用于具体决策的会计、税务、法律、采购、安全或公共部门评价规则。

• HM Treasury, The Green Book (2026)英国财政部《绿皮书》（2026） — official guidance on rationale, options, lifecycle costs and benefits, uncertainty, discounting, sensitivity, switching values, and balanced value-for-money judgement.——关于理由、方案、生命周期成本与收益、不确定性、折现、敏感性、转换值及平衡价值判断的官方指南。

• HM Treasury, Guidance on Developing Business Cases英国财政部《商业论证编制指南》 — the Five Case Model and staged development of decision-ready spending proposals.——五维商业论证模型与可供决策的支出提案分阶段编制方法。

• U.S. GAO Cost Estimating and Assessment Guide美国政府问责局《成本估算与评估指南》 — best practices for scope, assumptions, source data, time-phased estimates, sensitivity, risk, documentation, and updates using actual cost.——关于范围、假设、来源数据、分期估算、敏感性、风险、文档以及使用实际成本更新的最佳实践。

• NIST AI Risk Management FrameworkNIST AI 风险管理框架 — voluntary guidance for governing, mapping, measuring, and managing AI risk throughout the lifecycle.——关于在生命周期中治理、映射、测量与管理 AI 风险的自愿性指南。

Apply the structure proportionately. Use the organization's approved currency, price basis, tax treatment, discount rate, accounting policy, approval thresholds, and control standards. Public-sector social cost-benefit analysis may require a wider perspective, prescribed discounting, distributional analysis, and statutory duties beyond this organizational ROI model.

应按比例应用该结构，并使用组织批准的币种、价格基础、税务处理、折现率、会计政策、审批门槛与控制标准。公共部门社会成本效益分析可能要求更广泛的视角、规定折现、分配影响分析以及超出本组织 ROI 模型的法定义务。

Fill, calculate, copy, and download an analysis worksheet在线填写、计算、复制并下载分析底稿

Use this compact worksheet to connect the decision and evidence notes to a four-period discounted cash-flow model. It is a working record, not an approval by itself.

用这份紧凑底稿把决策问题、证据记录与四期折现现金流连接起来。它是一份可供复核的工作记录，本身不代表项目已经获批。