思想领袖
AI 成功通过
在 Google 上的首选来源中添加 Unite.AI
大约三年前，我买了我第一次付的AI订阅费，并开始测试它能为我做什么。 逐渐地，开始的个人实验变成了一个小型的内部项目，然后是全公司的商业订阅，最后是几乎遍及每一个部门的AI推出——发展、金融、法律、营销和支助。
第一阶段基本上奏效了。 我们的内部发展速度在大约六个月内增加了大约五倍。 与此同时，另一个数字上升了：我们的AI支出增加了大约15倍。
当时我明白，成功采用大赦国际本身就会产生问题。 一旦技术被广泛使用，我们开始的规则就不再起作用。 由于AI正在获取更多的公司数据和承担更多的任务，成本可能失控，而许多工具本身仍然由外部提供者控制。
所以，当AI的收养工作有效时，首席执行官应该做什么，但是公司已经不能完全控制它是如何被使用的了？让我们试着回答这个问题。
停止将主动使用AI与商业成果等同起来
我立刻注意到的是，在第一次推出期间，收养指标，例如活跃用户人数或使用频率，是有道理的。 作为一个领导者，你想知道员工是否在使用这些工具，尤其是哪些工具，以及AI是否真的可以节省他们的时间。 然而，一旦整个公司的使用范围扩大，这些数字就变得不那么有用。
我们自己也看到了一旦人们能够接触到强大的模型，他们开始将模型用于几乎一切。 一度，员工开始问AI在纽约几点。 当然，这个例子是荒谬的，但显示如何快速地增长使用，而不为公司创造任何实质性价值。
Shopify在集成AI之后经历了一个类似的问题，以至于其内部的agent,River，现在起草多达其生产码拉请求的一半。 然而CEO托比·吕特克(Tobi Lütke)最近警告他所谓的“slop grew”：人工智能生成的工作是雇员在没有适当检查的情况下传递的，让别人来审查、修复甚至总结产出。
换句话说，更多的AI活动可以很容易地产生更多的工作而不是更多的价值。 公司可能拥有令人印象深刻的使用量和巨大的产出量，而部分产出只是为别人创造了另一个任务。
这是CEO需要改变的提示、登录和信使消费都显示工作流程内有大量活动。 不过他们并没有显示AI是否真的自费。
速度创建新瓶装
有一句谚语：仓促浪费。 我想这也适用于AI，正如我们得知的，让人们在AI身上“变得更快”并不一定意味着你突然没有多少工作要做。
在开发中，AI现在允许我们的团队在同样时间内生产出更多的代码，但最主要的是，所有代码在启用之前仍然需要检查。 毫无疑问，自动审查有帮助。
却与我盲目相信的东西相去甚远。 所以，虽然开发者自己花费的时间较少，他们现在有更多的产出来审查，测试和决定。
换句话说，它只是显示在另一个伪装。
这使得我们发展速度的五倍增长听起来略低一些，因为需要人类关注的代码数量也随之上升。 人工智能工作流程的实际成本还包括检查人工智能生产内容所花费的时间，最重要的是，决定你是否能够首先相信该产出。
第二次推出是必然的
从我的立场来看，公司需要所谓的“第二次”AI推出。
第一次推出主要是关于领养：给人机会，寻找使用案例，看技术实际帮助哪些地方和不帮助哪些地方。 但经过几个月的实际使用，你已经有了比假设更有价值的东西——你自己的证据，说明成本在哪里增长，哪些团队能产生更好的效果，以及人们花太多时间检查AI输出。
因此，我认为在这里，重新设计围绕AI的规则，而不是简单地推动更广泛的使用，是有意义的。 在实践中，我着重谈三件事：
- 把模型和任务匹配起来常规请求可以被引导到更小，更便宜的模型，而更复杂的工作则转到那些证明需要额外成本的模型。 最终目标是避免为低价值任务支付前沿模型价格。
- 确定人类控制是强制性的。 对于生产代码，法律文件或其他影响大的产出，界定在工作流程进一步推进之前，人类必须批准哪些结果。 另外，为敏感数据设定访问规则，让员工知道哪些模型可以接收，哪些不能。
建立备用计划，如果关键的工作流程依赖于外部提供者，则记录提示、集成、数据流动和具体模型的依赖性，使其足以在紧急情况下将工作量转移到另一提供者。
这是第一次AI推出与第二次AI推出的区别。 第一个问题：我们在哪里可以使用AI？第二个问题：我们应该在哪里使用，在什么规则下使用，以什么代价使用？
审计再次扩大前的推出
一旦人工智能收养成功，在实际使用大约90至180天后
这意味着衡量公认的业务成果，使模型及其成本与任务的实际价值相匹配，确定哪些方面的人文审查仍然是强制性的，哪些方面的工作流程可以在某种程度上更接近于完全自动化。
首次推出证明AI可以让公司更快。 第二个因素是如何防止这种速度变成不必要的开支、不受控制的准入或企业从未自觉同意的供应商依赖。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Thought Leaders
Successful AI Adoption Creates a Second Rollout Companies Rarely Plan For
Add Unite.AI to your preferred sources on Google
About three years ago, I bought my first paid AI subscription and kicked off testing what it could actually do for me. Gradually, what began as personal experimentation turned into a small internal project, then company-wide business subscriptions and, eventually, an AI rollout spanning almost every department — development, finance, legal, marketing and support.
The first phase, basically, worked. Our internal development speed increased roughly fivefold within about six months. At the same time, another number rose: our AI spending increased about 15 times.
That was when I understood that successful AI adoption could create problems of its own. Once the technology was widely used, the rules we started with no longer worked. Costs could get out of hand as AI was getting access to more company data and absorbing more tasks, while many of the tools themselves were still controlled by outside providers.
So, what should a CEO do when AI adoption works, but the company is not fully in control of how it is being used anymore? Let’s try to answer that.
Stop Equating Active AI Usage With Business Results
What I noticed right away was that during the first rollout, adoption metrics, such as the number of active users or frequency of use, made sense. As a leader, you want to know if employees are using the tools, which ones in particular, and whether AI actually saves them time. Yet, once usage spreads across the whole company, these numbers become much less useful.
We saw this ourselves. Once people had access to powerful models, they started using them for almost everything. At one point, employees began asking AI what time it was in New York. Of course, the example is ridiculous, but it shows how quickly usage can grow without creating any substantial value for the company.
Shopify experienced a similar problem after integrating AI so deeply that its internal agent, River, now drafts up to half of its production-code pull requests. Yet CEO, Tobi Lütke, recently warned about what he calls “slop grenades”: AI-generated work that employees pass on without properly checking it, leaving someone else to review, fix or even summarize the output.
In other words, more AI activity can easily produce more work instead of more value. A company may have impressive usage numbers and huge amounts of output, while part of that output simply creates another task for somebody else.
This is what CEOs need to change. Prompts, logins and token consumption all show plenty of activity inside the workflow. Still, they don’t show if AI is actually paying for itself.
Speed Creates a New Bottleneck
There’s a saying: haste makes waste. I think it applies to AI as well, as we learned that making people “faster” with AI doesn’t necessarily mean you suddenly have less work to do.
In development, AI now allows our team to produce considerably more code in the same amount of time, but the main thing is that all that code still needs to be checked before it goes live. Without a doubt, automated review helps.
Yet, it’s still far from something I’d trust blindly. So, while developers spend less time writing everything themselves, they now have much more output to review, test and make decisions about.
In other words, the work doesn’t disappear. It just shows up in another guise.
This makes our fivefold increase in development speed sound a little less overwhelming, given the amount of code requiring human attention rises alongside it. The real cost of an AI-enabled workflow also includes the time spent checking what AI produces and, above all, deciding whether you can trust that output in the first place.
The Second Rollout Is Imperative
From where I stand, companies need what I’d call a “second” AI rollout.
The first rollout is primarily about adoption: giving people access, finding use cases, seeing where the technology actually helps and where it doesn’t. But after several months of real usage, you already have something much more valuable than assumptions — your own evidence on where costs are growing, which teams produce better results and where people spend too much time checking AI outputs.
So, this is where I think it makes sense to redesign the rules around AI instead of simply pushing for broader usage. In practice, I would focus on three things:
- Match the model to the task. Routine requests can be routed to smaller, cheaper models, while more complex work goes to the models that justify the extra cost. The ultimate goal is to avoid paying frontier-model prices for low-value tasks.
- Identify where human control is obligatory. For production code, legal documents, or other high-impact outputs, define where a human has to approve the result before it moves further in the workflow. Separately, set access rules for sensitive data so employees know which models can receive it and which can’t.
- Build a backup plan even if you think you don’t need one. If a critical workflow depends on an external provider, document the prompts, integrations, data flows and model-specific dependencies well enough that the workload can be migrated to another provider in case of emergency.
That’s the difference between the first and second AI rollout. The first asks: where can we use AI? The second asks: where should we use it, under what rules and at what cost?
Audit the Rollout Before Scaling It Again
Once AI adoption works and the first rollout seems to be successful, I wouldn’t keep scaling it in the same way for another year. After approximately 90 to 180 days of real usage, I’d stop and audit what the company has actually built.
This means measuring accepted business outcomes, matching the model and its cost to the actual value of the task, identifying where human review is still mandatory and where a workflow can be somehow closer to full automation.
The first rollout proves that AI can make the company faster. The second one is what keeps that speed from turning into unnecessary spending, uncontrolled access, or vendor dependencies the business never consciously agreed to.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.