The Rise of Intelligence Ownership

Part I第一部分

Everyone asked the same question每个人都在问同一个问题

Since ChatGPT launched in 2022, business leaders have been asking the same question: what can AI do for us? The answer began with low-risk tasks: summarizing documents, drafting emails, producing first drafts that a human would edit. 自 2022 年 ChatGPT 发布以来，企业领导者们一直在问同一个问题：AI 能为我们做什么？最初的答案集中在低风险任务上：总结文档、起草电子邮件、生成由人工润色的初稿。

It quickly moved into higher-value cognitive work, such as software development and content generation, and grew into more ambitious projects, like attempts to build an AI company brain, a system connected to internal knowledge, data, and tools that could coordinate work and eventually operate parts of the business autonomously. AI 很快进入了更高价值的认知工作领域，例如软件开发和内容生成，并发展出更宏大的项目，比如尝试构建“企业 AI 大脑”——一个连接内部知识、数据和工具的系统，旨在协调工作并最终自主运营部分业务。

While a lot of time, energy and tokens have been invested in AI adoption, measurable outcomes have barely been achieved at scale. However, some companies embraced being AI-first and saw enormous gains in productivity, revenue, and cost, while others lagged behind or failed to change their organizations enough to reach high ROI. 尽管在 AI 采用上投入了大量的时间、精力和 Token，但几乎没有取得可规模化的显著成效。然而，一些公司积极拥抱“AI 优先”战略，在生产力、营收和成本方面获得了巨大收益；而另一些公司则落后了，或者未能进行足够的组织变革以实现高投资回报率（ROI）。

Recent data from corporate expense management platform Ramp reveals a stark contrast in performance: the top quartile of companies investing in AI saw their revenue more than double between November 2022 and December 2025, while businesses with zero AI expenditure experienced a mere 15% increase. 企业费用管理平台 Ramp 的最新数据显示了绩效上的鲜明对比：在 2022 年 11 月至 2025 年 12 月期间，AI 支出排名前四分之一的公司营收翻了一番以上，而零 AI 支出的企业仅增长了 15%。

There are many reasons why AI has done wonders for some companies while others have struggled to see the return on their investment, but research primarily points in five directions. AI 为一些公司创造了奇迹，而另一些公司却难以看到投资回报，原因有很多，但研究主要指向五个方向。

In this article we give a detailed overview of the deployment technique the winning group keeps converging on: fine-tuning open-source models with reinforcement learning. We cover how it addresses the last three challenges above, and how it turns knowledge only your organization has (namely data, tools, and processes) into a model no vendor API can match at a fraction of the cost. 在本文中，我们将详细介绍获胜群体趋向采用的部署技术：通过强化学习对开源模型进行微调。我们将涵盖它如何解决上述后三个挑战，以及如何将仅贵司拥有的知识（即数据、工具和流程）转化为任何供应商 API 都无法比拟的模型，同时只需极低的成本。

Part II第二部分

What the winners do differently获胜者有何不同之处

Most of the companies pulling ahead in the AI race made the same discovery: owning your intelligence wins on both performance and cost. A model trained to complete your specific workflows in your specific environment is very likely to outperform a general-purpose model that has never seen inside your company. Additionally, since you do not need to pack as many general-purpose capabilities into a model that is meant to operate in a specific environment, you can often get away with a smaller model that is orders of magnitude cheaper to run. 大多数在 AI 竞赛中领先的公司都有一个相同的发现：拥有自己的智能在性能和成本上都能获胜。一个经过训练以完成特定环境下的特定工作流的模型，极有可能胜过从未深入了解过贵公司的通用模型。此外，由于不需要在旨在特定环境下运行的模型中塞入过多的通用能力，你通常可以使用一个规模更小、运行成本低几个数量级的模型。

Owning your intelligence does not mean cancelling the ChatGPT or Claude subscription. Most workflow automation still starts with frontier models, and that is the right first move: it establishes a baseline for what is technically possible, and every call generates the data (inputs, decisions, corrections) that a specialist model later trains on. Once the automation leaves the prototyping stage, the priority flips to cost and performance at volume, and that is where fine-tuning open-source models with reinforcement learning comes in. In addition to that, your Fable 5 or ChatGPT model can call the specialist model to handle the parts of the workflow that require your internal knowledge, and the specialist model can call the frontier model for tasks that require high general ability. 拥有自己的智能并不意味着取消 ChatGPT 或 Claude 的订阅。大多数工作流自动化仍然始于前沿模型，这是正确的初步举措：它为技术可行性建立了基准，并且每次调用都会生成数据（输入、决策、修正），供后续的专业模型训练使用。一旦自动化脱离原型阶段，优先级就会转向规模化后的成本和性能，这时就是通过强化学习微调开源模型发挥作用的时候了。此外，你的 Fable 5 或 ChatGPT 模型可以调用专业模型来处理需要内部知识的流程环节，而专业模型也可以在需要高通用能力的任务时调用前沿模型。

Over the past two years this has hardened into a playbook: an open-source model, proprietary task data, and a reinforcement-learning stage against a scored version of the workflow. Below, we discuss three scenarios where this approach has been applied to real-world tasks. 在过去两年中，这已演变成一套成熟策略：开源模型、专有任务数据，以及针对工作流评分版本的强化学习阶段。下面，我们讨论三个将此方法应用于实际任务的场景。

Bridgewater Associates is one of the largest hedge funds in the world. Its analysts sift a constant stream of articles, filings, and emails, judging which documents are relevant to the firm's investment thesis and where boilerplate content begins. The catch is that relevant means relevant by Bridgewater's internal judgment, and no amount of prompting got frontier models to absorb that judgment reliably. So, the company decided to train an open-source model on labels from its own expert investors. The trained model makes roughly 30% fewer mistakes than the best frontier model, at a fraction of the inference cost. Bridgewater Associates 是全球最大的对冲基金之一。其分析师需要筛选源源不断的文章、文件和电子邮件，判断哪些文档与公司的投资论点相关，以及模板化内容从何处开始。难点在于“相关”是指符合 Bridgewater 内部的判断标准，而无论如何编写提示词，前沿模型都无法可靠地吸收这种判断。因此，该公司决定利用其专家投资者的标签来训练一个开源模型。该训练模型比最好的前沿模型减少了约 30% 的错误，且推理成本仅为其一小部分。

Harvey builds AI agents for law firms. Its hardest workloads are long-horizon: transaction due diligence and legal memo drafting, where the agent navigates large document sets, errors compound across steps, and even the best frontier models at maximum reasoning effort kept falling short of the quality bar. Harvey ran reinforcement learning on an open-weight model and got a legal agent that outperforms both GPT-5.5 and Claude Opus 4.8 on its rubrics. Harvey 为律师事务所构建 AI 代理。其最艰巨的工作负载是长期任务：交易尽职调查和法律备忘录起草。在这种任务中，代理需要处理庞大的文档集，错误会在步骤间累积，即使是最好的前沿模型在最大推理努力下也无法达到质量标准。Harvey 对一个开源权重模型进行了强化学习，最终获得了一个在自有指标上优于 GPT-5.5 和 Claude Opus 4.8 的法律代理。

Intercom is a customer-service platform whose AI agent, Fin, resolves almost two million customer issues a week. At that volume the problem is unit economics: frontier per-call pricing adds up fast, and every point of resolution rate matters. So Intercom's AI group post-trained its own vertical support model, Fin Apex, on billions of customer-service interactions. Intercom reports that it resolves more issues than the best frontier models while being cheaper to run. Intercom 是一家客户服务平台，其 AI 代理 Fin 每周解决近 200 万个客户问题。在如此大的规模下，问题在于单位经济效益：前沿模型的单次调用定价累积极快，且每一个解决率的百分点都至关重要。因此，Intercom 的 AI 团队在数十亿次客户服务交互的基础上，对其垂直支持模型 Fin Apex 进行了后训练。Intercom 报告称，该模型解决的问题比最好的前沿模型更多，且运行成本更低。

The same shape repeats well beyond these three. The appendix collects eight more deployments, with what each model was trained to do and what changed once it shipped. 同样的模式在上述三家公司之外也反复出现。附录收集了另外八个部署案例，展示了每个模型被训练执行的任务以及上线后的变化。

One deployment pattern repeats across these cases: reinforcement learning pushes an open-source model past the frontier on a specific set of workflows, at a fixed and dramatically lower cost per call. The deployment typically starts by using prompt and context engineered frontier models to establish the strongest default baseline. Those frontier traces and learnings are then reused to pack the focused capability into a compact model the company owns. 这些案例中重复出现一种部署模式：强化学习使开源模型在特定工作流集上超越了前沿模型，且单次调用成本固定且大幅降低。部署通常始于使用提示词和上下文工程化的前沿模型来建立最强的默认基准。随后，这些前沿模型的轨迹和经验被复用，将核心能力封装进公司自有的紧凑模型中。

Part III第三部分

Case Study: Catalogue Integrity Agent案例研究：目录完整性代理

One of the central challenges of running an e-commerce platform is keeping the product catalog trustworthy. Every listing must land in the correct category of the platform's taxonomy, and the attributes that power search, filters, recommendations, and downstream operations must be accurately extracted from its images and description. 运营电子商务平台的核心挑战之一是保持产品目录的可信度。每个商品条目必须落在平台分类体系的正确类别中，且驱动搜索、筛选、推荐和下游运营的属性必须从其图像和描述中准确提取。

Platforms staff this work with teams of catalog-integrity analysts, which grow together with the catalog: more listings mean more categories to know, more attributes to check, and more ambiguous edge cases to judge. Inconsistent decisions propagate: products become harder to find, recommendations deteriorate, and policy violations slip through. Miss a counterfeit and you expose customers and brands to fraud; over-flag and you build an expensive review queue that frustrates legitimate sellers. 平台通常配备目录完整性分析师团队来处理这项工作，团队规模随目录增长：更多的条目意味着需要了解更多的类别、检查更多的属性，以及判断更多模糊的边缘案例。不一致的决策会导致恶性循环：产品变得难以搜索，推荐质量下降，政策违规行为漏网。漏掉一个假冒商品会使客户和品牌面临欺诈风险；过度标记则会建立昂贵的审查队列，令合规卖家感到沮丧。

Let's start with the scale. eBay alone carries about 2.5 billion live listings, and Shopify's catalog absorbs more than 10 million product updates a day. Walmart has said that doing its AI-assisted catalog work with people alone would have taken roughly 100× the headcount. 让我们从规模谈起。仅 eBay 就拥有约 25 亿个在售条目，Shopify 的目录每天吸收超过 1,000 万次产品更新。沃尔玛曾表示，如果仅靠人工进行 AI 辅助的目录工作，大约需要 100 倍的人力。

Getting it wrong costs revenue: 71% of shoppers say they have returned a product because it didn't match its listing. But the reviewing itself is also expensive. A mid-size marketplace generates on the order of 10 million listing creates and edits a day; reviewed with a frontier model, that workload costs roughly half a billion dollars a year, while a fine-tuned specialist does the same job for around $10M: 出错会损失营收：71% 的购物者表示曾因产品与描述不符而退货。但审查本身也很昂贵。中型市场每天产生约 1,000 万次条目创建和编辑；如果使用前沿模型审查，该工作负载每年成本约为 5 亿美元，而微调后的专业模型完成同样的工作仅需约 1,000 万美元：

The companies actually running this workflow confirm the math. Shopify classifies products with fine-tuned open models at roughly 40 million inferences a day, a volume it says commercial APIs can't economically serve. Inspired by the industry leaders, we ran the playbook ourselves: an agent that examines a listing, searches the product taxonomy, checks the brand, retrieves the attribute schema, and commits a structured decision, escalating to human review when the evidence is thin or the risk too high. 实际运行此工作流的公司证实了这一计算。Shopify 使用微调后的开源模型进行分类，每天进行约 4,000 万次推理，其表示商业 API 无法经济地提供此规模的服务。受行业领袖启发，我们亲自执行了该策略：一个检查条目、搜索产品分类、核对品牌、检索属性模式并提交结构化决策的代理，当证据不足或风险过高时，会自动升级至人工审查。

Part IV第四部分

A digital twin the model can practice in模型可以练习的数字孪生

A model can only practice a workflow it can actually perform, fail at, and retry. So we rebuilt the catalog workflow as a digital twin of an e-commerce platform: a simulated environment with the same listings, the same tools, and the same stakes as the real thing. The raw material is the Amazon Berkeley Objects dataset of real product images and listing data, which we turned into 177,767 review episodes: one listing each, with image, title, description, claimed brand, and region. Into that stream we planted controlled policy cases, mismatched images, conflicting brand claims, and deliberately legitimate claims as hard negatives, so every episode has a known correct answer to score against. 模型只能练习它能够实际执行、失败并重试的工作流。因此，我们重建了目录工作流作为电子商务平台的数字孪生：一个模拟环境，拥有与真实平台相同的条目、工具和风险。原始素材是 Amazon Berkeley Objects 数据集的真实产品图像和条目数据，我们将其转化为 177,767 个审查片段：每个片段包含一个条目，带有图像、标题、描述、所称品牌和地区。我们在该流中植入了受控的政策案例、不匹配的图像、冲突的品牌声明以及故意伪装成合法声明的“硬负样本”，以便每个片段都有已知的正确答案用于评分。

Inside the twin, the model works the way an analyst would. It can search a taxonomy of roughly 13,000 categories, check whether a brand is registered and protected, and retrieve the required attributes for a chosen category; then it must commit a category, the attributes, and a policy decision. A scorer grades every episode, rewarding correct outputs and penalizing missed violations, unsupported attributes, invalid categories, and wasted tool calls. The penalties encode business priorities directly: a missed violation costs 7× more than a false alarm. 在孪生系统中，模型的工作方式与分析师一样。它可以搜索约 13,000 个类别的分类体系，检查品牌是否已注册并受保护，并检索所选类别所需的属性；然后它必须提交类别、属性和政策决策。评分器对每个片段进行评分，奖励正确的输出，并惩罚漏掉的违规行为、不支持的属性、无效类别以及浪费的工具调用。惩罚直接编码了业务优先级：漏掉违规的代价是误报的 7 倍。

Benchmarking the frontier models前沿模型基准测试

Before any training run, we measured how far the frontier could get on its own. We benchmarked five frontier models (GPT-5.5, GPT-5.6-sol, Gemini 3.1 Pro, Claude Opus 4.8, and Claude Fable 5) on 200 stratified validation episodes with identical tools, images, scorer, and turn budget, both with a plain prompt and with optimized prompt instructions: 2,800 characters of extraction conventions, lookup procedure, and worked examples, tuned the way a prompt engineer would tune a production system. The chart below shows both configurations, with our final GRPO-trained model added for comparison. 在任何训练运行之前，我们都测量了前沿模型在自身能力下能达到什么水平。我们在 200 个分层验证片段上对五款前沿模型（GPT-5.5、GPT-5.6-sol、Gemini 3.1 Pro、Claude Opus 4.8 和 Claude Fable 5）进行了基准测试，使用相同的工具、图像、评分器和调用预算，分别在普通提示词和优化后的提示词指令下进行测试：包括 2,800 个字符的提取惯例、查询程序和工作示例，其调优方式与提示词工程师调优生产系统的方式一致。下图显示了两种配置，并添加了我们最终 GRPO 训练的模型进行对比。

The best frontier configuration reached 76.9% of the achievable score; the trained 9B reached 87.3%. The gap is not intelligence. A frontier model starts every episode from zero: it has never seen this store's taxonomy, doesn't know its inventory conventions, which attribute values count as supported, or how the platform wants corner cases resolved, and it has to reconstruct all of that on the fly, from whatever fits in the prompt, on every single call. Optimized instructions can compress some of that knowledge, but the corner cases that decide the score are exactly the ones no instruction prompt can enumerate. 最好的前沿配置达到了可实现分数的 76.9%；训练后的 9B 模型达到了 87.3%。差距不在于智能。前沿模型在每个片段中都是从零开始：它从未见过该商店的分类体系，不知道其库存惯例、哪些属性值被视为支持，或者平台希望如何解决边缘案例，它必须在每次调用时从提示词中容纳的内容里实时重建所有这些信息。优化后的指令可以压缩部分知识，但决定分数的边缘案例恰恰是任何指令提示词都无法穷举的。

Part V第五部分

Training details训练细节

The training setup was modest. We rented two RTX PRO 6000 GPUs, one generating rollouts and one applying gradient updates, with the open-source prime-rl framework running the infrastructure. The full run was 1,000 optimizer steps, took about three and a half days, and cost roughly $500 in GPU time. 训练设置非常简易。我们租用了两块 RTX PRO 6000 GPU，一块用于生成 rollout，一块用于应用梯度更新，并使用开源的 prime-rl 框架运行基础设施。整个运行过程为 1,000 个优化器步长，耗时约三天半，GPU 时间成本约为 500 美元。

Most of that was not needed to catch the frontier. The model crossed the frontier band after roughly 250 steps, about a day of training; the remaining steps were spent squeezing out maximum performance. The final model scores 0.626 on the benchmark, 87.3% of the achievable ceiling and about ten points above the best frontier configuration. 大部分时间并不需要用来追赶前沿模型。模型在约 250 步（约一天的训练）后就跨越了前沿模型区间；剩余的步数用于榨取最大性能。最终模型在基准测试中得分为 0.626，达到了可实现上限的 87.3%，比最好的前沿配置高出约 10 个百分点。

The gain comes from teaching the model the semantics of its environment: over thousands of scored episodes it learns how the store's taxonomy, tools, and policies relate to the reward, and converges on the optimal distribution of actions over the environment's action space. A general model spreads its capacity across everything; the specialist spends all of it on the specific task at the expense of general ability. 收益来自于教授模型其环境的语义：通过数千个评分片段，它学习了商店的分类体系、工具和政策如何与奖励相关联，并在环境的动作空间中收敛于最优的动作分布。通用模型将其能力分散在所有事物上；而专业模型则将其全部投入到特定任务中，以牺牲通用能力为代价。

The result also isn't a dead end. When a more capable open-source base model ships, the recipe transfers: as a rule of thumb, a stronger base yields a stronger post-trained specialist. And the deployed model generates its own training data as it works: logged decisions can be distilled into supervised fine-tuning sets, making each re-training cheaper and better-informed than the last. 结果也不是死胡同。当性能更强的开源基础模型发布时，配方是可以迁移的：经验法则是，更强的基础模型会产生更强的后训练专业模型。而且部署的模型在工作时会生成自己的训练数据：记录的决策可以提炼成监督微调集，使得每次重新训练都比上一次更便宜、更明智。

Part VI第六部分

What can owning AI do for you?拥有 AI 能为你做什么？

Here is the whole experiment in three sentences. We took one high-volume e-commerce workflow, reviewing product listings, and rebuilt it as a digital twin. We let a small open-source model practice inside it for a few days and roughly $500 of GPU time. The trained specialist ended above every frontier configuration we measured, at a fraction of their price per decision. 这里用三句话总结整个实验。我们选取了一个高容量的电子商务工作流——审查产品条目，并将其重建为数字孪生。我们让一个小型的开源模型在其中练习了几天，耗费约 500 美元的 GPU 时间。训练后的专业模型最终超过了我们测量过的所有前沿配置，且单次决策价格仅为其一小部分。

The recipe transfers to any work your business repeats at volume. Look for the places where people or API calls turn information into decisions all day: routing tickets, extracting fields from documents, checking submissions against policy, classifying products, approving or flagging transactions. What unites them is that every decision can be checked: there is a rule, a schema, or an expert who can say whether it was right. That check is the whole trick. If a decision can be scored, a model can practice it; if it can only be debated, it cannot. 该配方可迁移到贵司任何重复进行的高容量工作中。寻找那些人员或 API 调用整天将信息转化为决策的地方：路由工单、从文档中提取字段、根据政策检查提交内容、分类产品、批准或标记交易。它们的共同点是每个决策都可以被检查：存在规则、模式或专家可以判断其是否正确。这种检查就是全部诀窍。如果决策可以评分，模型就可以练习它；如果决策只能争论，那就无法练习。

It is just as important to know where this is the wrong tool. Two questions settle most cases: how frequently the task runs, and whether its outcome is verifiable. 了解在何处使用错误的工具同样重要。两个问题可以解决大多数情况：任务运行的频率以及结果是否可验证。

If that describes one of your workflows, the question is no longer whether it's the right fit for you, but which of your decisions you should stop renting. Owning your intelligence is the kind of advantage that compounds: the model, the evaluation, and the data stay yours and improve together, on your terms. If you would like to see what that looks like for your business, we would be glad to work through it with you: 如果这描述了你的一个工作流，问题就不再是它是否适合你，而是你应该停止租用哪些决策。拥有自己的智能是一种会复合的优势：模型、评估和数据保持为你所有，并根据你的条款共同改进。如果你想看看这对你的业务意味着什么，我们很乐意与你一起探讨：