商业智能泔水：企业采用 AI 是如何滋生职场平庸的

背景与摘要： 当企业在真正发现问题之前就购买了昂贵的企业级软件时，员工往往会被迫使用该工具，仅仅是为了在财务报表上证明这笔开销的合理性。在生成式 AI 时代，这种动态关系发生了恶性转变。通过强制性培训、流于表面的研讨会以及展现“生产力”的压力，职场人士越来越多地生成出未经核实、包含 AI 幻觉的文档、行动计划和商业方案——这种现象被贴切地称为“商业智能泔水（BI Slop）”。正如工程师们已经学会审查那些难以卒读、由 AI 生成的代码一样，各个组织也必须以同样严谨的态度对待 AI 生成的商业智能，而不是把甄别的负担转嫁给同事。

Summary

When companies purchase expensive enterprise software before identifying a real problem, employees are forced to use the tool simply to justify its cost on the balance sheet. In the era of generative AI, this dynamic has taken a toxic turn. Through mandatory training, superficial workshops, and the pressure to show "productivity," professionals are increasingly generating unverified, AI-hallucinated documents, action items, and business plans—a phenomenon aptly dubbed "BI Slop" (Business Intelligence Slop). Just as engineers have learned to scrutinize unreadable, AI-generated code, organizations must apply the same rigor to AI-produced business intelligence before passing the burden onto colleagues.

强制使用工具的代价

你必须使用这个工具。我们必须证明它物有所值。

The Cost of Forced Tooling

You must use the tool. We must justify it.

公司并不会直接这么说，但归根结底就是这么回事。当你还没清楚地界定问题就购买了一个昂贵的解决方案时，你就必须使用这个工具来证明电子表格里那一项支出的合理性。在我工作过的大多数公司里，总会有某些应用程序我只用过一两次。当我想第三次使用它时，它往往已经不见了。如果一个工具不受欢迎，财务团队会毫不犹豫地将其砍掉。

These aren't the words companies use, but that's what it comes down to. When you buy an expensive solution before you've clearly identified the problem, you have to use the tool to justify its line item in the spreadsheet. At most companies I've worked in, there's always some application I've used once or twice. The third time I want to use it, it's gone. If a tool isn't popular, the finance team has no trouble dropping it.

但不知为何，面对 AI，情况却有所不同。如果没有人使用它，公司就会想当然地认为员工的生产力还不够高。

But for some reason, it's different with AI. If nobody's using it, the assumption is that employees aren't being productive enough.

强制培训的悖论

于是，我们被迫报名参加了强制性培训。名额有限，但我眼疾手快抢到了一个位子。自从 ChatGPT 问世以来我就一直在用它，所以我希望能通过这次培训让我从一个普通用户进阶为高手。

The Mandatory Training Paradox

So we got signed up for mandatory training. There was limited space, but I moved fast and got a spot. I'd been using ChatGPT since it came out, so I was hoping the training would turn me from a casual user into a pro.

然而，培训教我们的却是如何将数据从 Excel 等数据源复制粘贴到聊天界面中。我们使用的是示例数据，课堂上总有人会说“我的怎么没反应”。在场的开发人员询问了关于 Codex 的问题，而那位获得 OpenAI 认证的讲师回答说她不是开发人员。当有人问起既然 Copilot 已经集成在 Excel 里，我们能不能直接用它时，场面一度十分尴尬。

Instead, we were shown how to copy and paste data from Excel and other sources into the chat interface. We worked with sample data, and there was always someone in class who'd say "mine didn't work." The developers in the room asked about Codex. The OpenAI-certified instructor replied that she wasn't a developer. There was an awkward moment when someone asked if we could just use Copilot, since it's already integrated into Excel.

我拿到了我的证书。但我并不认为我学到了任何自己花点业余时间用免费账号学不到的东西。

I got my certificate. But I don't think I learned anything I couldn't have picked up with a free account on my own time.

拿锤子找钉子：生产力剧场的陷阱

然而，伴随证书而来的是必须使用该工具的义务。我手里拿着一把锤子，看什么都像钉子。

From Hammer to Nail: The Trap of Productivity Theater

What comes with a certificate, though, is the obligation to use the tool. I had a hammer, and everything looked like a nail.

每当我遇到冗长繁杂的信息，我就会复制、粘贴，然后让它提取核心观点。妙就妙在，生成的结果看起来总是条理清晰。我一开始就没时间去读那些密集的信息，所以自然也不会去核实它是否准确地引用了源材料。况且，大家都是这么干的。我讨厌别人用 AI 来写 Slack 消息，但在眼下，这似乎已经成了无可避免的事。

Whenever I ran into dense information, I'd copy, paste, and extract insight. The great thing is that the result always looks neatly presented. I didn't have time to read the dense information in the first place, so of course I wasn't going to verify whether it was accurately referencing the source material. Besides, everyone was doing the same thing. I hate it when people use AI to write a Slack message, but at this point, it's inevitable.

幻觉流水线

有一次，开完会后，我提取了会议记录，并让 ChatGPT 将其整理成行动项（action items），然后根据内容生成一个问答。它给了我一份非常详尽的文档。或者说，看起来非常详尽。

The Hallucination Pipeline

One time, after a meeting, I pulled the transcript and asked ChatGPT to organize it into action items and generate a Q&A based on the content. It gave me something very elaborate. Or at least it looked elaborate.

我粗略浏览了一下，然后就把它转交给了下一个人，他实际上需要使用这份文档来为一个项目进行调研。我没有注意到的是，文档中很大一部分内容是 AI 凭空捏造（幻觉）出来的。

I skimmed it and passed it along to the next person, who actually had to use it to research a project. What I failed to notice was that a large part of the document was hallucinated.

其中一些行动项听起来相当专业，甚至把我给骗了。但是对于真正要去执行这些任务的人来说，它们完全不合逻辑。我对这件事感到内疚，而这已经成为一种常态。就像我们作为软件工程师用大型语言模型写代码一样，如今这些代码背后的规划和架构也越来越多地由 LLM 来代笔了。

Some of the action items sounded technical enough to fool me. But to someone who actually had to work on them, they made no sense at all. I'm guilty of this, and it's become the norm. The same way we use large language models to write code as software engineers, the planning and architecting behind that code is increasingly written by LLMs too.

登场：商业智能泔水 (BI Slop)

在我们家，绝对禁止看那些粗制滥造的 AI 生成视频 (AI slop videos)，以至于我的孩子们还会监督那些看这类东西的朋友。所以，当我的儿子看到我电脑上开着 ChatGPT 时，他问我是在制造商业智能泔水 (BI Slop) 吗？

Enter: BI Slop

In my household, AI slop videos are strictly banned, to the point that my kids police their friends who watch that stuff. So when my son saw ChatGPT open on my computer, he asked if I was making BI Slop.

“BI Slop？那是什么？”

“就是商业智能泔水！”

"BI Slop? What's that?"

"It's Business Intelligence Slop!"

我为他能准确识别出我在做什么感到无比自豪。同时我也对自己感到无比失望，因为我正在做的，确确实实就是这个。

I was so proud of him for recognizing exactly what I was doing. And I was so disappointed in myself, because that's exactly what I was doing.

正如我们会仔细审查代码、拒绝合并完全由 AI 生成的 PR（拉取请求）一样，我们也需要关注我们使用这些工具的其他方式。由 LLM 产生的商业决策和计划在被接受之前，同样需要经过合理性的初步检验（smell test）。

The same way we scrutinize code and refuse to merge PRs (Pull Requests) that are entirely AI-generated, we need to pay attention to the other ways we use these tools. Business decisions and plans produced by LLMs need to pass the smell test before they're accepted.

效率的假象

问题本质上是一样的。我们之所以会合并那些庞大的 PR，是因为我们没有时间去仔细阅读它们，然后我们在日后往往会以意想不到的方式为此付出代价。

The Illusion of Efficiency

The issue is fundamentally the same. We merge large PRs because we don't have time to read them, then suffer the consequences later in unexpected ways.

在商业智能方面，我们因为内容过于庞杂而不去阅读输出的结果，又因为它表面上看起来不错就批准了它。结果，我们再次在日后以意想不到的方式承受了苦果。

With business intelligence, we don't read the output because of its sheer volume, and we approve it because it looks good on the surface. Then, again, we suffer the consequences later in unexpected ways.

这个问题的产生仅仅是因为我们被塞给了一个强制性的解决方案：用 AI 来证明电子表格里那项支出的合理性。 它所谓的“节省时间”，仅仅是指你跳过了阅读输出内容的环节，而你之所以敢这么做，是因为它在表面上看起来十分完美。

This is a problem created only because we were handed a mandatory solution: use AI to justify the line item in the spreadsheet. It only saves you time in the sense that you skip reading the output, and you do that because it looks good on the surface.