11-19 更新 实时
首页/ 技术栈/ 正文

让 AI 走出演示,走进全球企业

技术栈 来源:AWS Machine Learning 2026-11-18 03:05 0 阅读 机器翻译
让 AI 走出演示,走进全球企业

AWS Forward Deployed Engineering 现已在 15 个国家上线,包括美国、墨西哥、巴西、加拿大、澳大利亚、新加坡、印度、韩国、日本、英国、德国、荷兰、法国、瑞典和沙特阿拉伯,面向商业客户和公共部门客户;并由数十个内置的、生产就绪的加速器提供支持。

我接触的每一家企业都用 AI 做出了能跑起来的东西。一个能读取理赔申请的原型。一个能处理银行卡争议的智能体。一个能在几秒内回答客户问题的模型。

原型能跑通。演示能打动利益相关方。然后企业试图把它投入生产,真正的问题才开始出现。生产数据比测试集更杂乱。审计人员想知道智能体为何做出每一个决策。使用量从试点小组扩展到整个公司。而最重要的规则,原来只存在于人们的脑子里。

这些都不是模型问题。它们是业务问题,而正是它们决定了 AI 能否真正产生回报。从演示到生产的这一段就是最后一公里,大多数 AI 都卡在这里。这也正是我们的团队投入时间的地方。

我们在 6 月推出 AWS Forward Deployed Engineering(FDE),正是为了弥合这一差距。此后,我们与金融服务、医疗健康、体育、媒体、零售等领域的商业客户和公共部门客户一起解决了这些问题。以下是我们学到的东西,以及它如何塑造了我们所构建的东西。

模型提供商、云提供商和咨询公司都在把工程师派驻到客户身边。在 AWS FDE,我们衡量成功的标准是工程师离开之后留下了什么:一套由你们团队自己拥有的基础,能够持续学习的工作流,以及值得信赖的治理机制。

在各类客户中,最难的问题出现在三个层面:基础、工作流和治理。每一层都比看上去更难,而且每一层都建立在前一层之上。无论你是从零开始,还是在现有系统上构建,情况都是如此。在全新建设(greenfield)的场景中,你可以快速推进,但尚无基础可依托。在存量(brownfield)系统中,智能体必须在数十年积累的数据、规则和集成体系内运作。

智能体要在规模上做有价值的工作之前,需要一个基础:身份、安全、上下文、治理和运维控制。这个基础必须从第一天起就生产就绪,并被后续每一个用例共享。上下文往往是最⼤的缺口。大多数客户的数据分散在各个系统中,也没有统一的本体(ontology)——即智能体进行推理所需的、描述业务如何运作的共享模型。

没有共享的基础,第二个试点的成本几乎与第一个相当。集成被反复重建,经验教训也无法沉淀下来。大多数组织都会走到一个节点:停下来先把基础建好,这样每个新用例都从上一次积累的成果出发,而不是从零开始。我们称之为 AI 重置。

Samsung Life Insurance 有四家关联公司,各自从零开始构建自己的 AI 单点方案。在 45 天内,FDE 团队构建了一个受治理的智能体平台,拥有单一控制平面并内置合规能力。如今,新智能体通过注册和策略分配即可部署,而无需重建。

接下来是工作本身:高流量、规则密集的流程,例如理赔处理、客户尽职调查(KYC)审核、信用卡争议、贷款发起、发票核对、许可审批以及客户服务。这正是 AI 能够带来即时且可衡量成果的领域。回报来得很快,但驱动这些流程的知识很少存在于系统之中,而是掌握在从事这些工作的人手中。

难的不是自动化本身,而是挖掘每个决策背后的业务规则。当发票与采购订单不一致时,以哪条记录为准?例外情况何时可以自动结清,何时又需要人工介入?经验丰富的员工凭判断回答这些问题,而这种判断很少被写下来。

有时,这些知识就埋在系统本身之中。

Rio Tinto Iron Ore 依靠一套已运行 30 年的 Oracle 系统,运转着全球最复杂的从矿山到装船的供应链之一。该系统拥有超过 5,000 张数据库表和 140 个应用。此前,梳理这套系统是一项耗时数月的工程。FDE 团队构建的智能体在数周内就完成了梳理,并揭示出 Rio Tinto 团队能够验证并据此开发的依赖关系。

梳理现状只是起点。最大的机会出现在组织不再追问 AI 如何让现有工作更高效,而开始追问它如何改变工作的完成方式之时。

治理不是项目的一个阶段,而是架构的一部分。随着智能体数量激增,安全风险、成本以及你必须能够解释的决策也随之增加。每个智能体都需要拥有独立于人类身份的身份、权限和审计轨迹。而且,由于底层模型具有概率性,智能体需要持续的监控,以确保其在上线很久之后结果依然可预测、可审计。

系统还必须在其周围一切都在变化的情况下持续运转。模型会漂移,业务规则会调整,数据格式会改变,监管会收紧。新的安全问题会出现,代码会逐渐过时。随着更新的模型到来,围绕它们的指令和工具也会变化。你需要评估和防护机制,让智能体在必须可预测的地方保持可预测,在可以发挥创造力的地方保持创造力。

当治理放在最后才做时,它就会变成阻断生产的关卡。而当它从第一天就被内置进去时,它恰恰是让你的业务能够快速前进的保障。

全球领先的专业服务公司 Marsh 需要能够经受其自身审计标准的智能体。FDE 团队从一开始就把这些管控内置进去,首个用例在项目启动约两个月后便进入生产环境。目标领域的效率提升最高达到 70%,部分流程步骤从数小时缩短到数秒。此后 Marsh 扩大了这项工作的范围,并持续对系统本身进行扩展。

我们的视角来自我们自己构建和运行这些系统的实践。AWS FDE 会将一支由 AWS 工程师、科学家和产品经理组成的小型团队,与你们的工程师以及一组专用 AI 智能体配对协作。我们的许多 FDE 来自打造 Amazon Bedrock、Amazon Bedrock AgentCore、Kiro、AWS Quick 和 AWS Security Hub 的服务团队。他们曾在 AWS 的规模上构建并运行 AI 产品,并将这些工程实践带入你的环境。我们的团队携手合作,以 45 天为一个冲刺周期,在你自己的 AWS 账户中把一个高价值的业务问题推进到生产环境。客户不应需要在速度、规模和管控之间做取舍。FDE 的构建目标就是同时交付这三者。

速度——你带着一个宽泛的目标或模糊的问题来找我们。我们与你的业务相关方一起将其细化,然后从第一天起就在你的环境中构建,基于你的数据,并通过你的身份与安全管控。我们与真正做这些工作的人并肩,以短周期迭代。每个周期都会把真实案例摆到他们面前,随着他们的反应,那些不成文的规则会浮现出来。我们把这些规则沉淀进一个动态上下文中,因此即使我们退出,系统仍会持续学习。从原型到生产,无需任何重建。整个过程你都能看到可运行的演示,评估会对照你在意的业务成果跟踪进展,并且在一个季度或更短时间内你就拥有一个生产系统。

规模——单点解决方案可以在某一个用例上快速推进,却无法为接下来的五个用例留下任何基础。每个 FDE 团队都以相同方式工作,使用 AI-DLC(AI 驱动的开发生命周期)——这也是我们自己的服务团队用来把宽泛问题转化为可落地成果的方法。我们部署一套有明确主张的 AWS 技术栈,它体现了我们在构建 AI 服务中学到的东西,并根据你的环境做适配:反映你业务运作方式的动态上下文、供你的工程师与智能体班组构建的工作台、运行在 Amazon Bedrock AgentCore 上的运行时,以及用于治理、成果、成本和智能体表现的控制塔。针对文档处理、客户情绪分析等常见任务的加速器负责处理标准化部分,这样每个新工作负载都能专注于你业务特有的内容。这套有明确主张的 AWS 技术栈、上下文和管控机制可以从一个用例延续到下一个用例。

你的上下文正是让规模化成为可能的关键。假设你的第一个用例是销售自动化工作流。由于它的上下文存放在你的 AWS 账户中,下一个用例——供应链优化——可以在此基础上构建,而不必从头开始。你的上下文、工具和 AWS 服务都集中在一处,因此每个用例的构建速度都比上一个更快。

控制——我们构建的东西归你所有。一切都运行在你的 AWS 账户中,处于你的身份、网络和安全控制之下。你的工程师从第一个冲刺周期起就与我们的团队并肩工作,在真实的业务问题上同时学习架构和交付方法。这样,在合作结束很久之后,他们仍能运行并扩展该解决方案。当我们退场时,你的团队通过控制塔来运行系统——这是一个统一视图,用于治理、保护和监控每一个智能体。我们衡量成功的标准,是你的团队在无需我们参与的情况下能构建多少。

原文较长,译文为节选;译文由英文原文机器翻译,术语按行业惯例统一,可在文末展开原文核对。

查看英文原文(Getting AI out of the demo and into businesses globally)

AWS Forward Deployed Engineering is now available in 15 countries including U.S., Mexico, Brazil, Canada, Australia, Singapore, India, South Korea, Japan, UK, Germany, Netherlands, France, Sweden, and Saudi Arabia for commercial and public sector customers; and powered by dozens of built in production-ready accelerators.

Every enterprise I talk to has built something with AI that works. A prototype that reads a claim. An agent that resolves a card dispute. A model that answers a customer question in seconds.

The prototype works. The demo impresses stakeholders. Then the organization tries to put it into production, and the real problems begin. Production data is messier than the test set. Auditors want to know why the agent made each decision. Volume grows from a pilot group to the whole company. The rules that matter most turn out to live in people’s heads.

None of these are model problems. They’re business problems, and they decide whether AI ever pays off. The stretch from demo to production is the last mile, where most AI stalls. It’s also where our teams spend their time.

We launched AWS Forward Deployed Engineering (FDE) in June to close that gap. Since then, we’ve worked through these problems with commercial and public sector customers in financial services, healthcare, sports, media, retail, and more. Here’s what we’ve learned, and how it shaped what we built.

The problem has three layers

Model providers, cloud providers, and consulting firms are all embedding engineers with customers. At AWS FDE, we measure success by what remains after our engineers leave: a foundation your team owns, workflows that keep learning, and governance you can trust.

Across customers, the hardest problems show up in three layers: the foundation, the workflows, and governance. Each one is harder than it looks, and each one builds on the one before it. That holds whether you’re starting fresh or building on existing systems. On a greenfield build, you can move fast, but there’s no foundation to stand on yet. On a brownfield system, agents must work within decades of data, rules, and integrations.

Layer one: A foundation that compounds

Before an agent can do useful work at scale, it needs a foundation: identity, security, context, governance, and operational controls. That foundation must be production-ready from day one and shared by every use case that follows. Context is often the biggest gap. Most customers have data spread across systems and no common ontology, the shared model of how the business works that agents need to reason over.

Without a shared foundation, the second pilot costs nearly as much as the first. Integrations get rebuilt, and lessons fail to carry forward. Most organizations reach a point where they stop and build the foundation first, so each new use case starts from what the last one built instead of from scratch. We call this the AI reset.

Samsung Life Insurance had four affiliates, each building its own AI point solutions from scratch. In 45 days, the FDE team built a governed agentic platform with a single control plane and compliance built in. New agents now deploy through registration and policy assignment instead of a rebuild.

Layer two: Automating workflows

Next is the work itself: High-volume, rule-heavy processes such as claims processing, know your customer (KYC) reviews, card disputes, loan origination, invoice reconciliation, permitting, and customer service. This is where AI can deliver immediate and measurable outcomes. The payoff is fast, but the knowledge that drives these processes rarely lives in a system. It lives with the people who do the work.

The hard part isn’t automation itself. It is uncovering the business rules behind each decision. When an invoice and a purchase order disagree, which record wins? When can an exception clear on its own, and when does it need a person? Experienced employees answer these questions from judgment, and that judgment is rarely written down.

Sometimes the knowledge is buried in the systems themselves.

Rio Tinto Iron Ore runs one of the world’s most complex mine-to-ship supply chains on a 30-year-old Oracle system. That system has more than 5,000 database tables and 140 applications. Mapping it had been a multi-month exercise. The FDE team built agents that mapped it in weeks, surfacing the dependencies that Rio Tinto’s teams could validate and build against.

Mapping what exists is only the start. The biggest opportunities emerge when organizations stop asking how AI can make existing work more efficient and start asking how it can change how work gets done.

Layer three: Governance that holds up as everything changes

Governance isn’t a phase of the project. It’s part of the architecture. As agents multiply, so do security risks, costs, and decisions you must explain. Every agent needs its own identity, permissions, and audit trail, separate from human identity. And because the models underneath are probabilistic, agents need monitoring that keeps their results predictable and auditable long after launch.

The system also must keep working as everything around it changes. Models drift, business rules shift, data formats change, and regulations tighten. New security issues appear, and code falls out of date. As newer models arrive, the instructions and tooling around them will change too. You need evaluations and guardrails that keep agents predictable where they must be and creative where they can be.

When governance comes at the end, it becomes the gate that stops production. When it’s built in from day one, it’s what lets your business move fast.

Marsh, a leading global professional services firm, needed agents that would hold up under its own audit standards. The FDE team built those controls in from the start, and the first use case reached production about two months after work began. Efficiency gains in the targeted areas have been as much as 70 percent, with some process steps moving from hours to seconds. Marsh has since expanded the work and keeps extending the system itself.

How we solve it: Speed, scale, and control

Our perspective comes from building and running these systems ourselves. AWS FDE pairs a small team of AWS engineers, scientists, and product managers with your engineers and a crew of specialized AI agents. Many of our FDEs come from the service teams that build Amazon Bedrock, Amazon Bedrock AgentCore, Kiro, AWS Quick, and AWS Security Hub. They’ve built and run AI products at AWS scale, and they bring those engineering practices into your environment. Together, our teams take a high-value business problem to production in 45-day sprints, in your own AWS account. Customers shouldn’t need to choose between speed, scale, and control. FDE is built to deliver all three.

Speed – You might come to us with a broad goal or an ambiguous problem. We refine it with your business stakeholders, then build in your environment from day one, on your data and through your identity and security controls. We build in short cycles alongside the people who do the work. Each cycle puts real cases in front of them, and the unwritten rules come out as they react. We capture those rules in a living context, so the system keeps learning after we step back. Nothing gets rebuilt between prototype and production. You see working demos throughout, evaluations track progress against the outcomes you care about, and you have a production system within a quarter or less.

Scale – A point solution can move fast on one use case but leaves no foundation for the next five. Every FDE team works the same way, using the AI-Driven Development Lifecycle (AI-DLC), the method our own service teams use to take a broad problem to a working outcome. We deploy an opinionated AWS stack that reflects what we’ve learned building AI services, adapted to your environment: a living context of how your business works, a workbench where your engineers and agent crews build, a runtime on Amazon Bedrock AgentCore, and a control tower for governance, outcomes, cost, and agent performance. Accelerators for common tasks, such as document processing and customer sentiment analysis, handle the standard pieces, so each new workload can focus on what’s specific to your business. The opinionated AWS stack, context, and controls carry from one use case to the next.

Your context is what makes scaling possible. Say your first use case is a sales automation workflow. Because its context lives in your AWS account, the next use case, supply chain optimization, builds on it instead of starting over. Your context, tools, and AWS services all sit in one place, so each use case is faster to build than the last.

Control – You own what we build. Everything runs in your AWS account under your identity, network, and security controls. Your engineers work alongside ours from the first sprint, learning both the architecture and the delivery approach on a real business problem. They can then run and extend the solution long after the engagement ends. When we step back, your team runs the system from a control tower, a single view to govern, secure, and monitor every agent. We measure success by how much your team can build without us.

After the first or second project, we want you running AI software factories of your own. One financial services customer did exactly that with AI-DLC, with agents and engineers building together. Throughput rose 53%, and code reviews got 56% faster.

You also keep your choices open. Models are improving fast. When a better or more cost-effective model arrives, you can adopt it through Amazon Bedrock without rebuilding, so you control both capability and cost. You can also bring your own coding agents. Our opinionated AWS stack, adapted to your environment, connects to your systems through APIs, open standards, and connectors, and it works in hybrid environments. You keep the tools and partners you already use.

A flywheel where everyone gets better

Your team gets better. Your capability grows with every use case. The living context, the opinionated AWS stack, and your team’s skills carry forward, so the second use case doesn’t rebuild what the first one built.

AWS Services get better. Our engineers bring what they learn from customers back into the products. You build on the latest services, not a custom solution that falls out of date.

All of AWS works around you. Customer development engineering explores early ideas, FDE takes the hardest problems to production, AWS Professional Services and AWS Partners extend the work across your organization, and AWS Support keeps it running. You get the full depth of AWS behind your AI, and you stay in control of how you use it.

What’s new, and how to start

AWS FDE is now available in 15 countries to commercial and public sector customers, and it’s simpler to buy. Each 45-day sprint has a set price under the FDE Terms and is billed through your AWS account.

We work with customers wherever they are in their transformation. For customers starting out, we build a secure, governed foundation and the first workflows on it. For those ready to change how work gets done, we rebuild workflows such as claims processing and KYC reviews, along with the ones their customers use directly, like shopping and issue resolution. Agents then run the work. For those held back by legacy applications, we work with their engineers to build software factories they can reuse for each new workflow. For those already scaling in production, we keep systems secure with agents that find and fix issues, and we lower the cost per outcome.

If you’re ready to start, talk with your AWS account team. Bring an executive who owns the outcome and engineers ready to build alongside ours. We’ve learned where projects stall before they start, so our agents take on much of the setup, like access, data, and security checks. Work can begin within days. Then we agree together on what the first sprint will deliver and how we’ll measure success.

The gap between AI that works and AI that changes business outcomes is real. It is an engineering, operational, and organizational one, and our teams help customers close it every day. The organizations that succeed with AI won’t necessarily have the most advanced models. They’ll be the ones that build the foundation, workflows, and governance to scale AI across the business.

We’re still early, and we learn something new with every customer. The models will continue to evolve, and so will the tools around them. What should last is a foundation you own, workflows that keep learning, and governance you can trust. That’s what we mean by what remains after our engineers leave.

To get started, contact your AWS account team or learn more about AWS FDE.


About the author

Dr. Asa Kalavade

Dr. Asa Kalavade

Asa is a technology and product leader at AWS who turns emerging technologies into scalable solutions for enterprise customers. She leads AWS Forward Deployed Engineering. Before that she led AWS Transform, which offers agentic AI based modernization of legacy software and infrastructure, and the transformation of AWS’s go-to-market tools and processes. In more than a decade at AWS, she has led seven 0-to-1 launches, built and scaled global teams, and led businesses through significant growth. Before AWS, Asa founded and ran two venture-backed startups, and she began her career at Bell Labs Research. She holds a Ph.D. in EECS from the University of California, Berkeley, and more than 40 patents.

正文保留英文原文,标题与摘要由机器翻译。

本文由采集器自动同步

原文发布于 AWS Machine Learning(Getting AI out of the demo and into businesses globally),标题、摘要与正文由机器翻译,术语按行业惯例统一,未经人工逐句校订。

阅读原文 ↗