Cost Benefit Analysis Example for AI Projects成本效益分析案例：AI 项目现金流、ROI、回收期与决策

Quick answer: what does this cost benefit analysis example show?快速答案：这份成本效益分析案例展示了什么？

The result is not an InfiniSynapse performance claim or a forecast for your organization. It is a transparent calculation model. Replace every assumption with evidence from your own workflow, keep cash savings separate from reusable capacity, and let sensitivity analysis determine whether the recommendation is robust.

该结果并非 InfiniSynapse 的性能主张，也不是对任何组织的预测，而是一套透明的计算模型。实际应用时应使用自身工作流证据替换全部假设，区分现金节省与可再利用产能，并通过敏感性分析判断建议是否稳健。

Define one decision before calculating value先定义一个决策，再计算价值

A data operations team receives 36,000 recurring requests per year. Analysts spend time clarifying requirements, preparing data, drafting queries, checking outputs, correcting errors, and delivering results. The decision owner must choose whether to continue the current workflow, improve intake and standardization without AI, or implement a controlled AI-assisted workflow for eligible requests.

某数据运营团队每年接收 36,000 个重复请求。分析人员需要澄清需求、准备数据、编写查询、检查输出、修正错误并交付结果。决策负责人需要在三个方案中选择：继续现有工作流、不使用 AI 而仅改善受理与标准化，或为合格请求实施受控 AI 辅助工作流。

The question is not “Does AI have value?” It is “Which feasible option best meets the defined outcome within the available budget and risk limits?” That wording preserves a credible counterfactual and prevents the preferred technology from becoming the premise of the analysis.

问题不是“AI 是否有价值”，而是“在可用预算与风险限制内，哪个可行方案最能实现既定结果”。这种表述保留了可信的反事实，避免把首选技术变成分析的预设结论。

Measure the current workflow with consistent definitions用一致定义测量当前工作流

The baseline uses twelve months of request records and a representative time study. A “completed request” means an output accepted by the requester after required review; abandoned requests and rework remain visible rather than being removed. Active labor includes analyst and reviewer time, while queue time is reported separately.

基线采用十二个月的请求记录与具有代表性的时间研究。“已完成请求”是指经过必要评审并被请求方接受的输出；放弃的请求与返工仍被保留，而不是从数据中删除。实际人工包括分析人员与评审人员耗时，排队时间则单独报告。

Compare three feasible options, including business as usual比较三个可行方案，包括维持现状

A fair analysis keeps business as usual as the benchmark and gives the non-technology option serious treatment. The process-only option improves intake forms, request classification, reusable query patterns, service levels, and review routing. The AI-assisted option includes the same process changes plus controlled drafting, retrieval, explanation, and review support for eligible tasks.

公平分析应把维持现状作为基准，并认真评估非技术方案。纯流程优化包括改善受理表单、请求分类、可复用查询模式、服务等级与评审路由。AI 辅助方案则在相同流程改进基础上，为合格任务增加受控的草拟、检索、解释与评审支持。

Option C proceeds to detailed analysis because it meets the outcome and control requirements and has a testable delivery path. Option B remains a useful fallback and should be retained in the decision table. Option A is not “free”: future contractor spend, delay, error, and unmet demand remain part of the counterfactual, even though they are not all treated as incremental project cost.

方案 C 因满足结果与控制要求且具有可测试的交付路径而进入详细分析。方案 B 仍是有价值的备选，应保留在决策表中。方案 A 并非“零成本”：未来外包支出、延迟、错误与未满足需求仍属于反事实的一部分，即使它们不全部计入新增项目成本。

Record the assumptions before showing the result先记录假设，再展示结果

The base case deliberately distinguishes total demand, technical eligibility, user adoption, gross time saved, and economically usable capacity. Multiplying total request volume directly by a demo time saving would overstate value because not every task is safe or suitable, not every user adopts the workflow, and not every recovered hour changes an organizational outcome.

基准情景有意区分总需求、技术合格率、用户采用率、毛节省时间与可经济利用产能。如果直接用全部请求量乘以演示中的节省时间，就会高估价值，因为并非每项任务都安全适用，也并非每位用户都会采用，而且并非每个释放小时都能改变组织结果。

Include implementation, operation, control, and exit costs纳入实施、运营、控制与退出成本

The example uses incremental cost: spending and internal resources caused by Option C relative to business as usual. Year-zero implementation is $145,000. Recurring annual cost is $78,000. The estimate includes work that is often omitted from vendor-price comparisons: data preparation, integration, evaluation, security review, human-review design, training, monitoring, incident response, and retirement planning.

本案例使用增量成本，即方案 C 相对维持现状而新增的支出与内部资源。第 0 年实施成本为 145,000 美元，年度持续成本为 78,000 美元。估算包含供应商价格比较中经常被遗漏的工作：数据准备、集成、评估、安全审查、人工评审设计、培训、监控、事件响应与退役规划。

Internal time is not automatically free. If subject-matter experts, security reviewers, engineers, and trainers are capacity constrained, their time has an opportunity cost and may also delay other work. The cost basis should state whether amounts are cash expenditure, allocated internal labor, or economic opportunity cost so finance does not add or omit them twice.

内部时间并非天然免费。如果领域专家、安全评审、工程师与培训人员的资源受限，他们的时间存在机会成本，也可能推迟其他工作。成本基础应明确每笔金额属于现金支出、分摊内部人工还是经济机会成本，避免财务重复计入或遗漏。

Convert eligible use into realized, non-duplicated benefit把合格使用转化为实际且不重复的收益

At steady state, 60% of 36,000 requests are eligible and 75% of those eligible requests use the new workflow. That produces 16,200 adopted requests. Human work falls from 22 to 9 minutes for those requests, recovering 3,510 hours. Only 65% of that capacity is expected to be redeployed into measured work, producing $136,890 in annual capacity value.

在稳定状态下，36,000 个请求中有 60% 合格，其中 75% 实际采用新工作流，因此形成 16,200 个已采用请求。这些请求的人工从 22 分钟降至 9 分钟，共释放 3,510 小时。预计其中只有 65% 能够重新投入可测量工作，因此年度产能价值为 136,890 美元。

Calculate net benefit, ROI, and payback step by step逐步计算净收益、ROI 与回收期

Year zero contains the $145,000 implementation cost. Year one benefit is reduced to 70% of the steady-state amount to reflect staged rollout, training, and adoption: $244,890 × 70% = $171,423. Years two and three use the steady annual benefit of $244,890. Recurring cost is $78,000 each operating year.

第 0 年计入 145,000 美元实施成本。为反映分阶段发布、培训与采用，第一年收益按稳定状态的 70% 计算：244,890 × 70% = 171,423 美元。第二年与第三年采用 244,890 美元的稳定年度收益，每个运营年度的持续成本均为 78,000 美元。

Simple ROI is a ratio over the selected horizon; it does not show timing. Payback shows when cumulative undiscounted cash flow turns positive, but ignores value after that point. Use both alongside a year-by-year cash-flow table, and use discounted measures when timing is material or the approval policy requires them.

简单 ROI 是所选周期内的比率，无法展示时点；回收期展示累计未折现现金流何时转正，却忽略该时点之后的价值。因此应把两者与逐年现金流表结合使用，并在时点影响重大或审批政策要求时使用折现指标。

Reconcile every result to a three-year cash-flow table把所有结果核对到三年现金流表

The cumulative balance remains negative at the end of year one, so calling the project “first-year positive” would be misleading if implementation cost is part of the investment decision. The project crosses zero about 3.7 months into year two under an even monthly realization assumption. In practice, use monthly or quarterly timing for large contracts, phased staffing, seasonal demand, or delayed benefits.

第一年末累计余额仍为负，因此如果实施成本属于投资决策的一部分，称项目“第一年为正”会产生误导。在假设收益按月均匀实现的情况下，项目约在第二年开始 3.7 个月后转正。对于大额合同、分阶段人员投入、季节性需求或延迟收益，应使用月度或季度时点。

Add discounted value without hiding the assumptions加入折现价值，但不要隐藏假设

For illustration, apply an 8% annual discount rate to years one through three while leaving year-zero cost undiscounted. This produces present-value benefits of $563,080 and present-value costs of $346,014. The resulting net present value is $217,067 and the benefit-cost ratio is 1.63.

作为演示，对第一年至第三年的金额采用 8% 年折现率，并保持第 0 年成本不折现。收益现值为 563,080 美元，成本现值为 346,014 美元，因此净现值为 217,067 美元，效益成本比为 1.63。

Eight percent is an illustrative corporate rate, not a universal recommendation. Use the rate, timing convention, inflation basis, tax treatment, and metric definitions approved by your finance team. Public-sector social appraisal may require a prescribed social discount rate and a wider perspective than organizational cash flow.

8% 只是示例性的企业折现率，并非通用建议。实际应使用财务团队批准的折现率、时点规则、通胀基础、税务处理与指标定义。公共部门社会评价可能要求规定的社会折现率，并采用比组织现金流更广泛的视角。

Test the assumptions that can reverse the decision测试可能逆转决策的关键假设

The base case is not a prediction. It is one internally consistent combination of assumptions. The downside case reduces eligibility, adoption, time saved, and usable capacity while increasing recurring cost. The upside case uses stronger coverage and reuse but does not remove review or control. These scenarios expose the range of plausible outcomes without assigning false precision to a single estimate.

基准情景不是预测，而是一组内部一致的假设组合。下行情景降低合格率、采用率、节省时间与可利用产能，并提高持续成本；上行情景采用更高覆盖率与利用率，但并不取消评审或控制。情景分析能够展示合理结果范围，避免对单点估算赋予虚假的精确性。

In the downside, steady annual benefit is only $99,550 before ramp-up, leaving little margin over $86,000 recurring cost and making the initial investment difficult to recover within three years. In the upside, steady annual benefit reaches $407,040. The recommendation should therefore depend on evidence for eligibility, adoption, minutes saved, utilization, and control cost rather than on a generic productivity claim.

在下行情景中，爬坡前的稳定年度收益只有 99,550 美元，相对 86,000 美元持续成本的余量很小，初始投资很难在三年内收回。上行情景下，稳定年度收益可达 407,040 美元。因此建议应取决于合格率、采用率、节省分钟、利用率与控制成本的证据，而不能依赖笼统的生产力主张。

Calculate switching values, not only optimistic ranges计算转换值，而不只展示乐观范围

A switching value is the point at which a key assumption changes the decision. At 100% adoption, the eligible workload could generate $182,520 of capacity value and $144,000 of other benefit, or $326,520 total steady benefit. To cover $223,000 of implementation plus first-year recurring cost without a ramp adjustment, adoption must reach approximately 68.3% if benefits scale proportionally.

转换值是关键假设变化并使决策发生改变的临界点。在 100% 采用率下，合格工作量可产生 182,520 美元产能价值与 144,000 美元其他收益，稳定收益合计 326,520 美元。若不考虑爬坡且假设收益随采用率同比例变化，要覆盖 223,000 美元实施与第一年持续成本，采用率需达到约 68.3%。

The 97.6% result is a warning, not a target: a one-year ramp makes full first-year recovery unrealistic under the base design. If policy requires twelve-month payback, the team must reduce implementation cost, accelerate adoption, increase validated benefit, change scope, or reject the project. It should not hide the mismatch by moving year-zero cost outside the calculation.

97.6% 是警示而非目标：在基准设计下，一年爬坡使第一年完全回本并不现实。如果政策要求十二个月回本，团队必须降低实施成本、加快采用、提高经验证收益、改变范围或拒绝项目，而不能通过把第 0 年成本移出计算来掩盖不匹配。

Keep financial results beside quality, risk, and evidence把财务结果与质量、风险和证据放在一起

A cost-benefit ratio does not prove that an option meets the objective. The decision table below shows why Option C is preferred conditionally rather than automatically. Option B has lower cost and lower delivery risk, while Option C has greater expected value but depends on representative pilot evidence and functioning controls.

效益成本比不能证明某个方案一定实现目标。下表说明为什么方案 C 是有条件的首选，而不是自动获批。方案 B 成本与交付风险较低；方案 C 的预期价值更高，但依赖代表性试点证据与有效控制。

Price controls and keep residual risk visible为控制定价，并保留剩余风险

AI project economics can look attractive when evaluation, review, security, monitoring, incident response, and model-change management are excluded. This example includes those controls in cost, but the decision still needs a risk register connecting each material failure mode to an owner, preventive and detective controls, response, residual exposure, and stop condition.

如果排除评估、评审、安全、监控、事件响应与模型变更管理，AI 项目经济性很容易显得优异。本案例已把这些控制计入成本，但决策仍需要风险登记表，把每项重大失效模式与负责人、预防与检测控制、响应、剩余风险及停止条件连接起来。

The NIST AI Risk Management Framework describes risk work through govern, map, measure, and manage. It is a useful voluntary reference for organizing responsibility and evidence, not a certification claim. Apply it proportionately to the decision, affected people, data, model behavior, and consequences.

NIST AI 风险管理框架通过治理、映射、测量与管理组织风险工作。它是一项有用的自愿性参考，可帮助组织职责与证据，但不代表认证。应根据决策、受影响人员、数据、模型行为与后果按比例应用。

Approve the next evidence gate, not an unchecked forecast批准下一证据门，而不是未经验证的预测

The recommendation is conditional because the largest value drivers—adoption, eligible coverage, human review, and capacity reuse—are not fully proven before production. Each gate should name the evidence owner, measurement window, sample coverage, approval authority, response when a threshold fails, and the date when the model is refreshed.

该建议采用条件式批准，因为采用率、合格覆盖率、人工评审与产能再利用这些最大价值驱动因素在生产前尚未完全验证。每个决策门都应明确证据负责人、测量窗口、样本覆盖、批准权限、门槛失败后的响应，以及模型更新日期。

Build your own cost-benefit analysis in nine steps用九个步骤完成自己的成本效益分析

1. Frame the decision.界定决策。 Name the owner, deadline, population, objective, budget, constraints, and available choices. 明确负责人、期限、对象、目标、预算、约束与可选方案。
2. Measure business as usual.测量维持现状。 Use consistent task, time, quality, demand, risk, and cost definitions. 使用一致的任务、时间、质量、需求、风险与成本定义。
3. Generate feasible alternatives.生成可行备选。 Include no-change, process-only, limited pilot, and scaled variants where credible. 在可信情况下纳入不变更、纯流程、有限试点与规模化方案。
4. Estimate lifecycle cost.估算生命周期成本。 Cover implementation, operation, controls, support, incidents, contingency, and exit. 覆盖实施、运营、控制、支持、事件、应急与退出。
5. Map benefit mechanisms.映射收益机制。 Connect each benefit to eligible use, adoption, attribution, utilization, evidence, and an owner. 把每项收益连接到合格使用、采用、归因、利用、证据与负责人。
6. Build timed cash flow.建立分期现金流。 Place costs and realized benefits in the period when they actually occur. 在成本与实际收益真正发生的期间记录它们。
7. Calculate consistent metrics.计算一致指标。 Reconcile net benefit, ROI, payback, NPV, and BCR to one cash-flow model. 把净收益、ROI、回收期、NPV 与 BCR 核对到同一现金流模型。
8. Stress-test and find switching values.压力测试并寻找转换值。 Change high-impact assumptions and calculate the threshold that reverses the recommendation. 改变高影响假设，并计算使建议逆转的临界点。
9. Write a conditional decision.形成条件式决策。 State the option, funding, owner, controls, evidence gaps, gates, monitoring, and stop rules. 明确方案、资金、负责人、控制、证据缺口、决策门、监控与停止规则。

Avoid the errors that make an attractive model unusable避免让漂亮模型失去决策价值的错误

Test your assumptions in the ROI Calculator在 ROI 计算器中测试你的假设

Use the InfiniSynapse ROI Calculator to model your own investment, operating cost, realized benefit, horizon, and return. Start with the base case, then create downside and upside copies. Preserve the source, owner, date, confidence, and calculation basis for every input outside the calculator so the result remains reviewable.

使用 InfiniSynapse ROI 计算器建模自己的投资、运营成本、实际收益、周期与回报。先建立基准情景，再复制为下行与上行情景。应在计算器之外保留每项输入的来源、负责人、日期、置信度与计算基础，确保结果可以被复核。

Continue from example to evidence and approval从案例继续走向证据与审批

Frequently asked questions常见问题

Sources, scope, and calculation notes来源、范围与计算说明

The figures in this page are a hypothetical instructional example created to demonstrate a transparent organizational cost-benefit model; they are not measured product results. The method draws on official guidance that emphasizes options, lifecycle costs and benefits, uncertainty, monitoring, risk, and documented assumptions.

本页数字为用于演示透明组织成本效益模型的假设性教学案例，并非经测量的产品结果。方法参考强调备选方案、生命周期成本与收益、不确定性、监控、风险及书面假设的官方指南。

• HM Treasury, The Green Book (2026)英国财政部《绿皮书》（2026） — appraisal of objectives, options, lifecycle costs and benefits, risk, discounting, sensitivity, and switching values. Its public-sector social-value perspective is broader than this organizational example.——关于目标、方案、生命周期成本与收益、风险、折现、敏感性及转换值的评价指南；其公共部门社会价值视角比本组织案例更广。

• 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 development of decision-ready spending proposals.——五维商业论证模型与可供决策的支出提案编制方法。

• NIST AI Risk Management FrameworkNIST AI 风险管理框架 — voluntary guidance for governing, mapping, measuring, and managing AI risk.——关于治理、映射、测量与管理 AI 风险的自愿性指南。

• 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 documented assumptions, credible cost estimates, sensitivity, risk, and updates with actual data.——关于记录假设、可信成本估算、敏感性、风险以及使用实际数据更新的最佳实践。

All calculations are rounded to the nearest dollar or one decimal place. The example assumes benefits occur evenly within each operating year for the payback interpolation. An 8% illustrative rate is used only in the discounted example. Your finance, accounting, tax, procurement, legal, privacy, security, and risk teams should review the actual model before a real commitment.

全部计算四舍五入到美元整数或一位小数。回收期插值假设收益在每个运营年度内均匀发生；8% 折现率仅用于折现示例。真实投资承诺前，应由财务、会计、税务、采购、法律、隐私、安全与风险团队审核实际模型。

See how the recommendation changes across three scenarios对比三个情景，观察审批建议如何变化

The worked example is one base case, not a promise. Switch between downside, base, and upside assumptions to see which result is robust and which depends on favorable execution.

文中的结果只是一个内部一致的基准情景，并不是承诺。切换下行、基准和上行情景，可以看出结论是否稳健，以及哪些结果依赖较理想的执行条件。