Turning contracts into searchable data at OpenAI

将合同转化为OpenAI中的可搜索数据

https://openai.com/index/openai-contract-data-agent/

---

视频

https://openai.com/index/openai-contract-data-agent/

---

This is part of our series sharing internal examples of how OpenAI is using its own technology and APIs. These tools are being used internally, only at OpenAI, and are shared here as illustrative examples of how frontier AI is supporting use cases across our teams. We’re also sharing the internal tool names for a clearer look at how frontier AI helps our teams get work done.

这是我们系列分享的一部分，展示OpenAI如何利用自身技术和API的内部案例。这些工具仅在OpenAI内部使用，此处分享作为前沿AI如何支持我们各团队工作场景的示例说明。我们还公开了内部工具名称，以便更清晰地了解前沿AI如何协助团队完成工作。

When contracts became the bottleneck

Every enterprise deal comes with a signed contract. Each one has start dates, billing terms, renewal clauses.

At first, the process was manageable: read line by line, retype into a spreadsheet, move on. But when volume doubled and doubled again, this manual approach broke.

“In less than six months, the team went from reviewing hundreds of contracts each month to more than a thousand. And yet we’d only hired one new person. It was obvious that the process wasn’t going to scale,” says Wei An Lee, AI Engineer.

当合同成为瓶颈时 每笔企业交易都伴随着一份签署的合同。每份合同都包含起始日期、计费条款和续约条款。

最初这个流程尚可应付：逐行阅读、重新输入电子表格、继续处理。但当业务量翻倍再翻倍时，这种人工方式就崩溃了。

"不到半年时间，团队每月审核的合同从几百份激增到上千份。而我们只新增了一名员工。很明显，这个流程根本无法扩展，"人工智能工程师李维安表示。

Building a smarter workflow

Instead of throwing more people at the problem, our finance and engineering teams built a contract data agent. The design principle was simple: take the repetition out of contract review, keep experts firmly in control.

构建更智能的工作流程
我们的财务和工程团队没有投入更多人力解决问题，而是开发了一个合同数据代理工具。设计原则很简单：消除合同审查中的重复劳动，同时确保专家始终掌握控制权。

The Agent works in three steps:

• Ingest data: PDFs, scanned copies, even phone photos marked up with handwritten edits. What used to be dozens of inconsistent files now flow into one pipeline.
• Inference with prompting: Using retrieval-augmented prompting, the system parses contracts into structured data. It doesn’t dump a thousand pages into context; it pulls only what’s relevant, reasons against it, and shows its work.
• Review: Finance experts review the structured output, complete with annotations and references for any non-standard terms. The agent highlights what’s unusual; humans are then looped in to review.

该代理的工作分为三个步骤：

数据摄取：处理PDF文件、扫描副本，甚至是带有手写批注的手机照片。过去数十份格式不一致的文件如今统一汇入一条处理流程。
推理提示：系统通过检索增强提示技术，将合同解析为结构化数据。它不会将上千页内容直接塞入上下文，而是精准提取相关信息，进行逻辑推理并展示推导过程。
审核环节：财务专家会审查结构化输出结果，对任何非标准条款添加完整注释和参考依据。代理程序会标出异常内容，随后由人工介入复核。

（注：根据技术文档翻译规范，保留了"retrieval-augmented prompting"等专业术语的直译，同时将"shows its work"意译为"展示推导过程"以符合中文技术文档表述习惯。"loop in"采用"介入"的译法准确传达了人机协作的工作机制）

“We’re not just parsing, we’re reasoning—showing why a term is considered non-standard, citing the reference material, and letting the reviewer confirm the ASC 606 classification.”

Siddharth Jain, AI Engineer

Confident contract reviews

The output is a dataset that’s immediately useful across finance workflows. What once took hours arrives overnight, annotated and ready for validation. Experts remain in the loop, but their role shifts from manual entry to judgment.

自信的合同审查
输出结果是一个数据集，可立即应用于各类财务工作流程。以往需要数小时完成的任务，如今一夜之间就能完成，并附有注释，随时可供验证。专家仍参与其中，但其角色已从手动录入转变为判断决策。

“The amazing thing is that the heavy lifting happens with AI—and then our teams wake up in the morning to data that’s ready for them to review.”

Wei An Lee, AI Engineer

This design ensures confidence: professionals get structured, reasoned data at scale, but their expertise drives the outcome.

The results:

• Faster turnaround. Reviews cut in half, ready overnight.
• Higher capacity. Thousands of contracts processed without expanding headcount in lockstep.
• Smarter context. Non-standard terms flagged with reasoning and references.
• Queryable results. Tabular output in the data warehouse allows for easier data analysis.

该设计确保可信度：专业人士能获得规模化、结构化且有依据的数据，但其专业能力决定最终成果。

成效如下：

• 周转更快：审核时间减半，隔夜即可完成
• 处理量更大：无需同比增加人力即可处理数千份合同
• 语境更智能：非标准条款自动标注原因并提供依据
• 结果可查询：数据仓库中的表格化输出便于分析

Each cycle of human feedback sharpens the Agent, making every review faster and more accurate.

“The only way we can scale as OpenAI scales is through this,” Wei An said. “Without it, you’d have to grow your team linearly in lockstep with contract volume. This lets us keep the team lean while handling hypergrowth.”

人类反馈的每一轮循环都让智能体更加敏锐，使每次审核都更快更准确。

"随着OpenAI规模扩大，我们实现规模化的唯一途径就是通过这种方式，"魏安说。"没有它，你就必须让团队规模与合同数量同步线性增长。这让我们能在处理高速增长的同时保持团队精简。"

Beyond contracts

This architecture now supports procurement, compliance, even month-end close. The same principle applies: automate the rote work, keep humans in charge of judgment. 

Engineers describe it as “manual work already done,” not decisions replaced. Finance teams still write the story of the numbers; the Agent ensures they don’t spend their day doing painstaking work.

超越合同
这一架构现已支持采购、合规，甚至月末结算。同样的原则适用：自动化重复性工作，由人类掌控判断。

工程师将其描述为“已完成的手动工作”，而非替代决策。财务团队依然负责解读数字背后的故事；智能代理则确保他们无需整日埋头于繁琐工作。

A new operating model for finance

What started as a fix for contracts has become a new way of working in finance. Data parsing runs overnight. Professionals focus on analysis and strategy. Leaders scale confidently with growth, without growing teams in lockstep.

The contract data Agent is a blueprint for how AI can responsibly transform regulated, high-stakes work. It shows what becomes possible when experts partner with intelligent systems: more leverage, more confidence, and more time spent on what matters most.

最初作为合同修复方案的技术，如今已成为金融行业的新型工作模式。数据解析实现夜间自动化处理，专业人士得以专注分析与战略制定。企业领导者能够伴随业务增长从容扩展规模，无需按比例扩充团队规模。

合同数据智能体为人工智能如何负责任地改造受监管的高风险工作提供了范本。它展现了专家与智能系统协作时实现的可能：更强的杠杆效应、更高的决策信心，以及更多投入在核心事务上的时间。

---

Turning contracts into searchable data at OpenAI