PDF · 中文
WHITEPAPER · 2026

企业级 AI Agent 落地:从试点到生产

一份给决策者的精简判断框架

作者:King Cao
适用对象:正在推进 AI Agent 落地,但试点多、上生产少的金融与大型企业
2026 年 11 月
企业级 AI Agent 落地白皮书  ·  中文版

摘要

AI Agent 项目卡住的地方,多数和模型没关系:点位选错、收益说不清、扩不出部门。

下面五条按落地顺序排,说的都是同一件事:怎么让一个能演示的 Demo 变成能进考核、进预算的链路。

  1. 先测链路,再选点位。 时间做加法,可靠性做乘法:局部环节提速 10 倍,整条链路最多省 20%;而一条 20 步的链路,每步成功率 95%,一次跑通只剩 36%。
  2. 首个试点只过三条:闭环、可度量、可逆。目标是用最低成本证明路走得通,不是创造最大价值。
  3. 甜点区是「有 SOP,但执行还靠人照着纸干活」的流程。 最不成熟的和最成熟的,都不适合先做。
  4. 用三层指标加同期 A/B 证明效果。 只测效率、用上线前后对比,结论都不可信。
  5. 第一天就画整合路径,并保留「停」的选项。 扩不出去,才是最大的失败。
图 01|从「能演示」到「能考核」的五个动作
图 01|从「能演示」到「能考核」的五个动作

01 · 为什么 POC 很多,上生产的很少

过去两年,几乎每家企业都做过 AI 试点。真正进入生产环境、进入部门考核、进入下一年度预算的,比例很低。

这个落差不是个别现象。麦肯锡 2025 年的全球调研(105 个国家、1,993 名受访者)显示,88% 的企业已在至少一个职能中使用 AI,但近三分之二仍未开始企业级规模化,任何单一职能里把 AI Agent 做到规模化的都不超过 10%。MIT NANDA 的统计更直接:企业已在生成式 AI 上投入 300–400 亿美元,而 95% 的试点没有产生可衡量的损益影响。

常见的原因就三个:

  • 只提效了一个点:局部指标很好看,整体交付周期没变,业务方感受不到。
  • 无法度量:没有基线、没有同口径对比,说不清到底好了多少,预算续不上。
  • 扩不出去:单点跑通了,横向推广时卡在数据授权、权限和部门 KPI 上。

三件都不是技术问题。先看清整条链路,再决定让 AI Agent 站哪一环,顺序反过来就会白做一轮。

行业层面的判断也指向同一处。Gartner 预计,到 2027 年底超过 40% 的 Agentic AI 项目会被取消,原因集中在成本失控、业务价值说不清、风险控制不足(Gartner,2025 年 6 月)。前两条正是本文第 2 节和第 5 节要解决的问题。

02 · 先测链路,再选点位

一张客服工单从进线到办结,要过好几道手。假设「拟答复」占 20%,其余是接入分派、身份核验与跨系统查询、质检复核、发送回访和返工。那么即使拟答复提速 10 倍、甚至 100 倍,整张工单的周期最多也只缩短 20%——因为这一环最多只占五分之一。瓶颈会立刻转移到下一环:答复写得快了,质检和回访排起了队,办结时长并没有同比例缩短。

试点通常从「写回复」开始,因为它演示效果最好:一段文字生成出来,肉眼立刻能判断像不像人写的。但演示效果和链路改善是两件事——前者在一个环节里就能被看见,后者只在整条链路上出现。

先测链路,再选点位。 看哪里占比最大、哪里是当前的真瓶颈,别看哪里最容易提效。

这里的「链路」指客户视角的端到端流程:从进线、分派到办结回访,跨几个系统、几个部门。部门内部的流程再顺,跨部门的交接处照样排队,而那一小段通常没有明确的责任人。

测算不需要工具,一张表就够。把最近六到十二个月的真实数据填进去即可:

环节平均耗时占总周期是否瓶颈可自动化程度
接入 / 分派填实际值
身份核验 / 跨系统查询填实际值
拟答复(写回复)填实际值
质检 / 合规复核填实际值
发送 / 回访填实际值
返工填实际值
图 02|单点提速 10 倍,整体为什么没感觉
图 02|单点提速 10 倍,整体为什么没感觉

多数企业填完会发现:自己最想先做的那一段,不是表里占比最大的那一段。「写回复」被最先选中,恰恰因为它的结果最好判断;可它只占两成,时间真正花得多的地方,多半在它前后。

这六段是按常见工单链路拆的,贵司的链路可能是四段,也可能八段,段数不必对齐。要保证的是把「返工」单列出来,它最容易被时间统计漏掉,也常常是真正的瓶颈;等待时间则摊回各段,别只记执行工时。

时间做加法,可靠性做乘法

上面那六个是业务环节,每个环节内部还要拆成系统动作:取单、验身份、查订单、查历史、查规则、生成答复、合规扫描、发送、回执、关单。一个跨系统闭环累计十几到三十步不等,这里取 20 步举例——即使每步成功率有 95%,端到端一次跑通也只有约 36%(0.9520),这还是在没有重试、没有人工兜底的前提下。

从这个 36% 出发,有两条路,都从每步 95% 起算,不要把它们读成一条连加的线:

  • 让每步更准:把单步失败率从 5% 压到 0.5%,端到端升到约 90%。
  • 让错的那次重来:每步加一次自动校验与重试,单步成功率升到 99.75%(1−0.052),20 步后约 95%。

第二条通常更便宜:它不需要模型更准,只需要每步的输出能被机器判定对错。

步数是这道乘法的指数。同样每步 95%,10 步约 60%,20 步 36%,30 步约 21%。所以要数的是贵司自己的步数,别照搬这个数。

图 03|20 步串联:从 36% 起步的两条路
图 03|20 步串联:从 36% 起步的两条路

可校验的输出,比更快的输出更重要。 这是自动重试、人工兜底和质检的共同前提,也正是第 3 节「闭环性」真正要问的那件事。

03 · 首个试点:三条准入,加上流程成熟度

一个职能要作为首个试点,需要同时满足三条。缺一条,项目大概率做成了也验不了。

判据一句话不满足会怎样
闭环性有明确输入,有可验收的输出做完说不清成败
可度量性已有现成 KPI 可作为基线无法证明改善,预算续不上
可逆性错了代价可控、能回滚、可追溯一次事故就终止项目
图 04|三条准入:客服全过,信贷审批暂缓
图 04|三条准入:客服全过,信贷审批暂缓

以金融行业为例:客服 / 坐席三项全过,是首选;审批流程时长好量、质量难量,次选;信贷审批出错即坏账、不可逆,暂不作为首个试点。

客服能成为行业默认起点,往往因为它是同时满足这三条的少数职能之一,跟价值大小没关系。

同样的判据,在不同部门的可落地程度差别很大,差别不在技术,在流程成熟度。AI Agent 需要两样东西:可执行的规则,和可校验的输出。流程不成熟的地方,连「什么叫做对了」都定义不出来。

五个可核查信号:有活着的书面 SOP、有系统留痕且数据可导出、有明确的角色与交接节点、有量化 KPI 与抽检评分、有 6 至 12 个月可回溯的历史数据。缺两条以上,建议先做流程治理。

图 05|甜点区不在最成熟的一端
图 05|甜点区不在最成熟的一端

最优区间也不在「最成熟」那一端。过度成熟的环节往往已被交互式语音应答(IVR)、规则引擎、机器人流程自动化(RPA)优化多年,剩下的都是规则覆盖不到的长尾个案,边际收益最低。甜点区很具体:量大、规则可枚举、有 SOP,但尚未被系统化覆盖,仍依赖人工判断边界。

04 · 客服不是一块业务,是四块

很多企业把「客服」当成一个整体来规划,结果在错误的那一块先投入。按规则确定性 × 情绪浓度排开,四段子链的难度差别很大。

顺序子链为什么排这里
①售后退换修规则可枚举,闭环清晰,情绪中等,最佳首切
②售前话术可训练,但要说服;KPI 是转化率,需谨慎设计
③技术支持依赖知识库,长尾复杂,做之前先做知识治理
④投诉个案差异大、情绪最高,最后做,或者不做

怎么分这四块:拿最近六到十二个月工单的「问题类型」字段归类,别照组织架构切——一张工单经常跨段,售前咨询转成售后投诉,所以统计单位是子链,不是部门。首切只取售后里最高频的三到五个场景,并在上线前把转人工的条件写死:规则库查不到、金额超阈值、客户明确要求谈人,任一条命中就转人。

图 06|售后 → 售前 → 技术支持 → 投诉
图 06|售后 → 售前 → 技术支持 → 投诉

先拿规则最硬、情绪最稳的一段把闭环跑通,再往难的方向扩展。

05 · 怎么证明有效:三层指标 + 同口径对比

「效果好不好」说不清,项目就活不过下一个预算周期。这不是修辞:麦肯锡 2025 年的调研里,只有 39% 的企业认为 AI 对 EBIT 产生了任何影响,其中多数不足 5%。必须同时看三层,只测效率是最常见的死法:

  • 效率:单均时长、单均成本、处理量。只看这一层,可能出现效率上升、质量下降,整体是负收益。
  • 质量:准确率、一次解决率、返工率、投诉率。
  • 风险:越权操作率、合规审计不通过率。一次事故,前面的收益全部归零。

三层定义清楚之后,还有两件事决定结论可不可信:样本够不够,口径是不是真的同口径。

样本量与观察窗口。「跑足周期再比」太含糊。要看出 3 个百分点的差异,两路合计至少需要数百单、连续 8 周以上;低于这个量级,抽样波动比 Agent 的真实效果还大,谁赢都可能只是运气。样本不够时,只报方向与置信区间,别报「提升 30%」这种点估计。

同口径的四个条件:同一批工单来源、同一时段、同一套评分标准、同一个复核人。换掉其中任何一个,两组之间的差就不再是 Agent 的贡献,而是口径差。

合规约束会压缩可分流范围。 金融场景里 A/B 不是纯随机:授信、定价、放款这类高风险决定不允许全自动执行,人工必须后置复核。能随机分流的只有「准备动作」——查询、拟稿、核对、比对;「决定动作」只能被辅助,不能被替代。这也意味着,能测到的收益上限,恰好就是这些准备动作在整条链路上的占比。

图 07|三层指标 + 同期分流 A/B
图 07|三层指标 + 同期分流 A/B

对比方式同样关键:必须做同期分流 A/B,不能用「上线前后对比」。前后对比会被季节性波动、人员流动、统计口径变更污染。做法是同一批工单随机分两路,一路走 AI Agent、一路走人工,跑足周期后比。

基线要来自书面 SOP,不能拍脑袋。 没有可复现的基线,任何「提升 30%」都不可信。流程制度在这里是度量基础设施,先有它,才量得出改善。

06 · 两个容易被忽略的前提

数据治理:以用促治,不要「先治理再上」。 先做两三年数据治理再上 AI,这条路几乎走不通,因为没有 AI 的时候,治理没有 ROI,业务部门不愿投入、工程没有终点。AI 是数据治理的触发器,不是它的结果。 有了明确的 AI Agent 目标,治理才收敛成「为了把一次解决率提到某个值,必须治理这几张表的这几个字段」。操作上别立项「数据治理工程」,在试点里附带一个最小治理集就够,只治理跑通这个闭环必需的数据。

规模化:先画整合路径,再选试点。 单点试点成功率通常不低,真正崩在横向推广,三个典型断点是:数据主权(试点的特批授权跨部门即失效)、KPI 冲突(客服优化一次解决率,后端部门的工单结构和 KPI 变差)、交接真空(部门内 SOP 清晰,部门之间没有 SOP,而 AI Agent 恰恰死在交接处)。

这个试点成功后,它的输出能不能被下一个部门接住?

这个问题要排在「选哪个部门」前面。否则做出来的是一个漂亮的孤岛,第二年就被叫停。

07 · 90 天路径

图 08|90 天,四个阶段,一个决定
图 08|90 天,四个阶段,一个决定
阶段时间关键动作交付物
测绘第 1–2 周画链路耗时表,对候选部门跑五问清单瓶颈定位 + 点位排序
定标第 3–4 周选定 1 个点位,定三层指标,取 6–12 个月数据指标定义 + 基线值
跑闭环第 5–10 周单点上线,人工兜底并行,异常全部记录首版闭环 + 事故台账
对比与决策第 11–12 周同口径 A/B 对比;画与下个部门的接口与数据授权决策报告:扩 / 改 / 停

第 11–12 周必须给出「停」的选项。 一个允许被停掉的项目,才会被认真评估;一个只能成功的项目,数据一定是修饰过的。

本文的边界。 安全合规(数据分级、脱敏、审计留痕)、可解释性(每一步为什么这样判)、模型风险管理(版本、漂移、回滚)是 Agent 进生产的前置条件,但都是独立议题,本文不展开:它们决定项目能不能上线,本文讨论的是上线之后能不能被证明有效。正在过安全评审的读者,建议把这三项单独立项,不要塞进试点里一起做。

附录 · 五问速判清单

决定某个部门是否适合做首个试点前,先答完这五个问题。答不上三个以上,先别立项。

已答问题为什么先问它
□有书面 SOP 吗?最近一次更新是什么时候?没有可复现的基线,任何「提升 30%」都不可信
□有系统承载吗?数据能导出、能回溯多久?导不出数据,就没有测算的起点
□多少人在做?人均日处理量多少?这个数决定 ROI 上限
□KPI 是什么?谁质检?标准能量化吗?量不出改善,预算续不上
□自营还是外包?按人头还是按单付费?这个口径决定收益怎么算

这五项不是评分表,是一票否决式的预筛:任何一项答不上,说明这个部门的「可度量性」还没成立。

本文的第三方数据均标注出处;未标注的比例与算式为推理示意,落地前请用本企业近 6–12 个月的真实数据替换。

参考来源

  1. MIT NANDA,The GenAI Divide: State of AI in Business 2025,2025 年 8 月。样本:300 个公开 AI 部署、150 位企业高管访谈、350 名员工问卷。本文引用「95% 的试点无可衡量损益影响」「300–400 亿美元投入」。
  2. McKinsey & Company,The State of AI in 2025: Agents, Innovation, and Transformation,2025 年 11 月 5 日。样本:105 个国家、1,993 名受访者。本文引用「88% 至少在一个职能使用 AI」「近三分之二未企业级规模化」「单一职能内 Agent 规模化不超过 10%」「39% 报告 EBIT 影响,多数不足 5%」。
  3. Gartner,Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027,新闻稿,2025 年 6 月 25 日。

第三方数据只用于说明行业分布,不构成对本框架结论的背书;本框架的比例、算式与图表示意,仍需以贵司数据替换。

引用格式:King Cao.《企业级 AI Agent 落地:从试点到生产》,2026 年 11 月。引用文中的比例与算式时,请注明它们是推理示意值,不是实测结果;第三方数字请回溯到「参考来源」的原始出处。

WHITEPAPER · 2026

Enterprise AI Agent Adoption: From Pilot to Production

A short decision framework for executives

Author: King Cao
For financial institutions and large enterprises with many AI pilots and few production deployments
November 2026
Enterprise AI Agent Adoption  ·  English edition

Executive summary

Most AI agent initiatives stall for reasons that have nothing to do with the model: the wrong starting point, no credible measurement, and no path beyond one team.

Five decisions, in the order they come up, for moving an AI agent from a demo that impresses people to a workflow that can be measured, budgeted and audited:

  1. Map the chain before picking a stage. Time adds up, reliability multiplies: one stage running 10× faster saves at most 20% of the cycle, while a chain of 20 steps, each at 95%, clears only 36% of the time.
  2. The first pilot must clear three gates: closed loop, measurable, reversible. Its job is to prove the path cheaply, not to maximise value.
  3. The sweet spot is a process with an SOP that people still execute manually. The least mature and the most mature processes are both poor starting points.
  4. Prove value with three metric layers and a same-period A/B test. Efficiency alone, or before/after comparisons, will not survive a budget review.
  5. Draw the integration path on day one, and keep "stop" as an option. Failure to scale is the real failure.
Figure 01 | Five moves from demo to production
Figure 01 | Five moves from demo to production

01 · Many PoCs, few production systems

Almost every enterprise has run an AI pilot in the last two years. Far fewer have put one into production, into a department's scorecard, or into next year's budget.

The gap is not anecdotal. McKinsey's 2025 global survey (1,993 respondents across 105 countries) finds 88% of organisations using AI in at least one function, yet nearly two-thirds have not begun scaling AI across the enterprise, and in no single function do more than 10% say they are scaling AI agents. MIT NANDA puts it more bluntly: of the $30–40 billion enterprises have spent on generative AI, 95% of pilots have produced no measurable P&L impact.

The usual causes are three:

  • Only one stage improved. The local metric looks great; overall cycle time does not move, so the business feels nothing.
  • No way to measure it. No baseline, no like-for-like comparison, no defensible answer to "how much better?" The budget does not survive.
  • No way to scale. The single point works, but expansion stalls on data rights, permissions and conflicting departmental KPIs.

None of these are model problems. Understand the whole chain first, then decide where the AI agent stands; reverse that order and the first round is wasted.

The industry-level forecast points the same way. Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls (Gartner, June 2025). The first two are what sections 2 and 5 are for.

02 · Map the chain before picking a stage

A service ticket passes through several hands before it closes. If drafting the reply is 20% of the cycle and the rest is intake, identity checks with cross-system lookups, QA, sending and rework, then making the draft 10× or even 100× faster shortens the whole cycle by at most 20% — that stage is one fifth of it. The bottleneck simply moves. Faster replies produce a queue at QA and follow-up, and resolution time barely changes.

Pilots usually start with drafting the reply, because it demos best: text appears on screen and anyone can judge whether it sounds human. But a good demo and a better chain are two different things — the first shows up inside a single stage, the second only shows up end-to-end.

Map the chain, then pick the pilot. Look at where the time actually goes and where the bottleneck sits, not where automation is easiest.

"The chain" here means the end-to-end path the customer experiences: intake, triage, closure, follow-up, across several systems and several teams. A team's own process can be smooth and still queue at the handover, and that short stretch usually has no owner.

You need no tooling, just one table filled with six to twelve months of real data:

StageAvg. timeShare of cycleBottleneck today?Automatable?
Intake / triagefill in
Identity + cross-system lookupfill in
Draft the replyfill in
QA / compliance checkfill in
Send / follow-upfill in
Reworkfill in
Figure 02 | A 10× speed-up that nobody feels
Figure 02 | A 10× speed-up that nobody feels

Most teams find, once the table is filled, that the stage they most want to automate is not the stage that eats the time. A drafted reply gets picked first precisely because its result is easy to judge; yet it is one fifth of the chain, and the hours sit on either side of it.

The six stages follow a common ticket workflow. Yours may have four stages or eight, and the count need not match. What matters is that rework gets its own line: it is the first thing a time study drops, and often the real bottleneck. Waiting time belongs inside each stage, not off to the side.

Time adds up. Reliability multiplies.

The six stages above are business stages, and each one breaks down into system actions: open the case, verify identity, look up the order, pull history, check the rules, draft the reply, run the compliance scan, send, receipt, close. A cross-system workflow runs anywhere from a dozen to thirty steps; take 20 as an example. At 95% success per step, only about 36% of cases finish end-to-end on the first pass (0.9520), before retries or human fallback.

From that 36%, there are two routes. Both start from 95% per step — they are not one cumulative line:

  • Make each step more accurate. Push the per-step failure rate from 5% to 0.5%, and the chain reaches roughly 90%.
  • Let the failed attempt run again. Add one automated check and retry per step: the step succeeds 1 − 0.052 = 99.75% of the time, and 20 steps land at roughly 95%.

The second is usually cheaper. It does not need a better model, only output that a machine can check.

Step count is the exponent here. At the same 95% per step, 10 steps give about 60%, 20 give 36%, and 30 give about 21%. Count your own steps rather than borrowing this number.

Figure 03 | 20 steps in series: two routes out of 36%
Figure 03 | 20 steps in series: two routes out of 36%

Verifiable output matters more than faster output. It is the common precondition for retries, human fallback and QA sampling — and it is what section 3 really asks when it asks about a closed loop.

03 · Three gates, plus process maturity

A function qualifies as the first pilot only if it clears all three gates. Miss one and you may build something you cannot evaluate.

GateIn one lineIf it fails
Closed loopDefined input, verifiable outputNo way to judge success
MeasurableAn existing KPI can serve as baselineImprovement cannot be proven
ReversibleErrors are cheap, recoverable, auditableOne incident ends the programme
Figure 04 | Three gates: contact centre passes, credit decisions wait
Figure 04 | Three gates: contact centre passes, credit decisions wait

In financial services: contact centres clear all three and are the natural first choice; approval workflows are measurable on time but weak on quality, so they come second; credit decisions are irreversible when wrong and should not be first.

Customer service is the default starting point because it is often the only place where all three gates hold at once. Value has little to do with it.

The same gates behave differently across departments, and the difference is process maturity, not technology. AI agents need executable rules and verifiable output. Where the process is immature, nobody can even define what "done correctly" means.

Five checkable signals: a living written SOP; a system of record with exportable logs; explicit roles and handoff points; quantified KPIs with sampled QA scoring; and six to twelve months of retrievable history. Missing two or more? Fix the process first.

Figure 05 | The sweet spot is not the most mature process
Figure 05 | The sweet spot is not the most mature process

The best target is not the most mature process, either. Mature areas have already been optimised for years by interactive voice response (IVR), rules engines and robotic process automation (RPA). What remains is the long tail of exceptions that rules never covered: the hardest work with the lowest marginal return. The sweet spot is specific: high volume, enumerable rules, an SOP in place, still executed manually, with humans judging the edge cases.

04 · Customer service is four businesses

Treating "customer service" as one block is why investment often lands in the wrong place. Ranked by rule certainty × emotional load, the four sub-chains are very different.

OrderSub-chainWhy it sits here
1After-salesReturn/repair rules are enumerable, closure is clear, emotion is moderate. Best first cut
2Pre-salesScripts can be trained, but persuasion is involved and the KPI is conversion
3Tech supportKnowledge-base dependent, long tail of complex cases; fix knowledge governance first
4ComplaintsCase-by-case, highest emotional load. Do last, or not at all

How to split it: bucket six to twelve months of tickets by their case-type field rather than by org chart — a single ticket often crosses sub-chains, a pre-sales question turning into an after-sales claim, so the unit of measurement is the sub-chain, not the department. For the first cut take only the three to five highest-frequency after-sales scenarios, and write the hand-over triggers down before launch: no rule matched, amount above threshold, or the customer asks for a person.

Figure 06 | After-sales → pre-sales → tech support → complaints
Figure 06 | After-sales → pre-sales → tech support → complaints

Prove the loop on the hardest rules and calmest conversations, then widen.

05 · Proving value: three layers, one fair comparison

If you cannot say whether it worked, the programme will not survive the next budget cycle. That is not rhetoric: in McKinsey's 2025 survey only 39% of respondents attribute any EBIT impact to AI, and most of those put it below 5%. Measure three layers together. Efficiency alone is the classic way to fail:

  • Efficiency: handle time, cost per case, throughput. On its own, efficiency can rise while quality falls, a net loss.
  • Quality: accuracy, first-contact resolution, rework rate, complaint rate.
  • Risk: out-of-scope actions, failed compliance audits. One incident wipes out every earlier gain.

Three layers are necessary but not sufficient. Two further questions decide whether the conclusion holds: is the sample big enough, and is the comparison genuinely like-for-like?

Sample size and observation window. "Run a full cycle, then compare" is too vague. To resolve a 3-point difference you need several hundred cases per arm and at least eight consecutive weeks. Below that, sampling noise exceeds the agent's real effect, and whichever arm wins may simply have been lucky. On a thin sample, report direction and a confidence interval rather than a point estimate such as "30% better".

Four conditions for like-for-like: the same case pool, the same period, the same scoring rubric, the same reviewer. Change any one of them and the gap between arms is a gap in measurement, not in the agent.

Compliance narrows what can be randomised. In financial services an A/B split is not purely random: high-stakes decisions — credit approval, pricing, disbursement — cannot run unattended, and a human must review afterwards. What can be randomised is the preparatory work: lookups, drafting, reconciliation, comparison. The decision can be assisted, never replaced. The measurable gain therefore tops out at whatever share those preparatory steps hold in the chain.

Figure 07 | Three metric layers + same-period A/B
Figure 07 | Three metric layers + same-period A/B

The comparison design matters just as much: use a same-period A/B split, not before/after. Before/after is distorted by seasonality, staffing changes and metric redefinitions. Route comparable cases randomly to an AI agent group and a human group over the same period, then compare.

Baselines come from the written SOP, not from estimates. Without a reproducible baseline, no "30% improvement" claim is credible. Process documentation is what gives an AI programme something to measure against.

06 · Two assumptions worth checking

Data governance follows use, not the other way round. Planning two or three years of data governance before AI rarely works: without an AI use case, governance has no ROI, business teams will not fund it, and the engineering work has no finish line. AI is the trigger for data governance, not its reward. A concrete AI agent target turns governance into "to reach this resolution rate, we must fix these fields in these tables": bounded, testable, fundable. Skip the standalone governance programme; attach a minimum viable dataset to the pilot, then widen it once the loop works.

Draw the integration path before choosing the pilot. Single-point pilots usually succeed; scaling is where programmes break. Three recurring failure points: data sovereignty (a one-off authorisation that expires across departments), KPI conflict (better first-contact resolution in service changes the downstream ticket mix and hurts another team's scorecard), and the handoff gap (SOPs inside a team, none between teams, which is exactly where AI agents die).

When this pilot succeeds, can the next team actually take its output?

Ask this before asking which department is keenest. Otherwise you build a beautiful island that gets switched off in year two.

07 · A 90-day path

Figure 08 | 90 days, four phases, one decision
Figure 08 | 90 days, four phases, one decision
PhaseWhenKey actionsDeliverable
MapWeeks 1–2Time the chain; run the five questions on candidate teamsBottleneck + ranked pilots
BaselineWeeks 3–4Pick one stage; define three metric layers; pull 6–12 months of historyMetric definitions + baseline values
Run the loopWeeks 5–10Go live on that stage with human fallback; log every exceptionFirst working loop + incident log
Compare & decideWeeks 11–12Same-period A/B; map the next team's interface and data rightsDecision report: scale / fix / stop

Weeks 11–12 must leave "stop" on the table. A project that is allowed to be stopped gets evaluated honestly. A project that can only succeed will produce polished numbers.

What this paper does not cover. Security and compliance (data classification, masking, audit trails), explainability (why a given step decided what it did) and model risk management (versioning, drift, rollback) are preconditions for putting an agent into production, but each is a discipline of its own and is out of scope here. They decide whether a project can go live; this paper is about proving value once it has. Teams currently in a security review should scope those three as separate workstreams rather than folding them into the pilot.

Appendix · Five questions before you start

Answer these before approving a first pilot. If you cannot answer three of them, do not start yet.

DoneQuestionWhy it comes first
□Is there a written SOP? When was it last updated?Without a reproducible baseline, no "30% improvement" is credible
□Is there a system of record? Can the data be exported, and for how long?No exportable data, no starting point for any estimate
□How many people do this work, and what is the daily volume per person?That number sets the ROI ceiling
□What is the KPI? Who does QA, and is the standard quantified?If improvement cannot be measured, the budget stops
□In-house or outsourced? Priced per head or per case?That basis decides how the saving is counted

These five are not a scorecard but a veto filter: any one left blank means the function is not yet measurable.

Third-party figures here are sourced; unsourced proportions and calculations are illustrative. Replace them with your own 6–12 months of operational data before committing to a target.

Sources

  1. MIT NANDA, The GenAI Divide: State of AI in Business 2025, August 2025. Sample: 300 public AI deployments, 150 executive interviews, 350 employee surveys. Cited here for "95% of pilots show no measurable P&L impact" and "$30–40 billion spent".
  2. McKinsey & Company, The State of AI in 2025: Agents, Innovation, and Transformation, 5 November 2025. Sample: 1,993 respondents across 105 countries. Cited here for "88% use AI in at least one function", "nearly two-thirds have not begun scaling", "no more than 10% scaling agents in any function", and "39% report any EBIT impact, most below 5%".
  3. Gartner, Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027, press release, 25 June 2025.

Third-party figures describe the industry, not this framework. The proportions, calculations and table values here still need replacing with your own data.

How to cite this paper: King Cao. Enterprise AI Agent Adoption: From Pilot to Production, November 2026. Proportions and calculations taken from this paper should be attributed as illustrative estimates rather than measurements; third-party figures should be traced to the original sources listed above.