IBM沃森在AI商业战略中做错了什么

What did IBM Watsondo wrong in BusinessAI Strategy

Last month, IBM pared down its Waston drug discovery effort, about 10 months after it scaled back its Watson hospital business.

在缩减Watson医院业务十个月之后，IBM上个月又缩减了其在新药研发领域的业务。

It is inevitable for some to ask the question, “is AI overhyped? is the long-touted next AI winter coming?” Well, predicting the future is really hard, especially regarding things undergoing heavy evolution. But the short answer is “no, no”.

一些人忍不住要问这个问题：“人工智能是否被夸大了？长期以来受到吹捧的人工智能是不是要迎来另一个寒冬了？”我不得不承认，预测未来真得很难，对于那些我们正在经历的重大事件而言更是如此。但简而言之，我的答案是“不”。

Actually, IBM still has some of the best researchers and engineers, and it still works with some of the top academic institutions.

实际上，IBM仍然拥有一批最优秀的研究人员和工程师，而且它还一直保持着和顶级学术机构的合作关系。

Where IBM failed, was what I call “lackluster business AI strategy at the executive level”. I’m not blaming IBM top executives for their shortcomings. Truth be told, IBM started its Watson health care 4 years before I preached the business AI strategy principal — “High value, low stake”. So, it’s not like that I knew the future better, and earlier than IBM.

我将IBM的失败称之为“AI商业战略在执行层面上的僵死”。我并不是在责备IBM高层管理人员的缺点。说实话，在我提出所谓的“高价值，低风险”的AI战略原则之前四年，IBM就推出了Watson医疗服务。所以坦白地说，我并不比IBM更了解未来。

It does not mean there is nothing we can learn from the Watson misfortune though  — especially in new Business Strategy in the AI era.

但这并不意味着，在这个AI时代，我们没办法从Watson的失败中学到任何东西，尤其是在战略层面上。

What caused Watson health’s trouble? It is actually great technology, AI was able to recognize pictures of cancer at a higher rate than doctors. But it’s great in the lab. The business executives had minimum reference and experience in introducing AI as a business, and mainly let engineers show them what can be accomplished, and went to market without fully integrated Business-AI strategy.

所以到底是什么给Watson健康带来了麻烦？AI实际上是一项伟大的技术；和医生相比，它能够以更高的准确率识别出癌症图片。至少，在实验室的时候它是这样的。业务主管在将AI作为一项业务引进时并没有太多的参考和经验，他们主要让工程师向他们展示了目前可以实现的目标，于是AI和业务在并没有完全集成的情况下就被推向了市场。

The problem of the health care industry is exactly what my Business AI strategy principal shuns — it is a High Stake topic. We are talking about people’s lives. While on paper (or more harshly speaking, in numbers) AI could have increased accuracy of diagnosis, mistakes, no matter how small the percentage is, could cost lives — and for those who suffer from those “unlikely” outcomes, that’s 100% for the individual and the family.

医疗保健行业所面临的问题正是我的人工智能商业战略原则所回避的——它面临的风险太高。我们在这里所讨论的是人们的生命。虽然在论文中（或者更严格的说，在数字上），AI可以提高诊断的准确性；但无论百分比有多低，AI犯下的错误都是以生命为代价的。对于那些遭受了“不太可能的”结果的个人和家庭来说，这个数字就是100%。

A doctor, while could make mistakes, is a trained physician who can react as soon as the error is found, who can continuously make judgments based on feedback, react to treatment (or the lack of), and correct the mistake. On the other hand, AI can’t, it gives its opinion based on “engineered” data as input and is unable to treat the patients, so whatever happens to the patients is disconnected from the AI’s opinion. While on paper, it reduced misdiagnosis, it does so in a single round comparison. By disconnecting diagnosis and treatment, it increases the risk of those who were misdiagnosed. Also, because human doctors hardly understand why AI makes certain recommendations, it makes it harder for human doctors to identify and correct those mistakes.

尽管医生也会犯错，但是作为一名训练有素的医生，一旦发现错误他就能迅速做出回应。他可以持续地根据反馈做出判断，对治疗（或者缺少治疗）进行调整，并纠正错误。但另一方面，AI却做不到这一点，它没办法对患者进行治疗。AI所给出的诊断全部来自于作为输入的“工程数据”。因此无论在治疗过程中病人出现了什么问题，都和AI的诊断相脱节（无法根据治疗反馈对AI的诊断进行修正）。只看字面数据，在一轮的人机比较中，它确实减少了误诊的可能性。但是由于AI的使用导致了诊断和治疗的脱节，被误诊的患者的风险反而增加了。此外，由于人类医生很难理解为什么AI会给出某些建议（可解释性差），因此想要识别并修正这些错误也会非常困难。

For drug research, it is similar, while AI can have a higher chance of finding a drug cocktail that might work, when it comes to the clinical trial, we are talking about human lives, and the lack of understanding (or agreement)of why the AI makes certain recommendations is not helping to convince me to take a trial drug.

就药物研究而言，问题可能是类似的。尽管AI可能有更高的机会找到可用的鸡尾酒疗法，但在临床试验中，我们谈论的是人的生命。由于对AI提出的某些建议缺乏理解（或者并不赞同），我并不会被说服去服用这些试验药物。

So, what are “high value, low stake” business process? Generally, these are high volume (lots of repeats), and high fault tolerance processes. For example, highway toll road license plate recognition, it is a repeating job (boring), and if the AI misses charging a vehicle, that’s not the end of the world. (but if AI sent out a ticket automatically, it can become a high stake use case, so we would not recommend sending out ticket automatically).

那么，什么是“高价值，低风险”的业务流程呢？通常来说，这类过程是高容量（大量重复性工作）和高容错率的。例如，高速公路收费站对车牌进行拍照识别就是一件重复且无聊的事情。如果AI忘了对某辆汽车进行收费，世界末日也不会到来。（但如果是使用AI自动发放通行凭证，事情的风险可能会变得比较高，所以我并不建议你这样干。）

Another example is a restaurant menu recommendation system — many people go to a restaurant do not know what they want to eat, asking the server for recommendations only goes that far, a voice-driven system can actually listen to the diner’s question, analyze their tone even accent and give a best-guess. If it’s a hit, great, if not, likely no harm was done “other than” increased customer engagement. (yeah, talking about Jeopardy win, right?)

另外一个例子就是餐馆的菜单推荐系统。许多人去餐馆吃饭时不知应该点些什么，于是他们会向服务员寻求建议（到目前为止都是这样）。一个语音驱动的系统（包含有语音识别，语义分析等功能）可以倾听顾客的问题，分析他们的语气甚至口音，并给出相应的建议。如果系统推荐的菜单恰好符合顾客的口味，那很好；如果没有，除了顾客的“积极性”会受到打击之外，也没其他什么损失。

How can businesses avoid what happened to IBM Watson?

企业应该如何避免发生在IBM Watson身上的事情呢？

It’s a tall order — IBM is one of the longest thriving tech company, being better than IBM is not something easy.

这是一项艰难的任务。毕竟IBM是这个世界上最繁荣的科技公司之一，想要做的比它更好并不容易。

That being said, there is hope. Seeing IBM testing out some business strategies gives the rest of us more insights into what to avoid, what to embark on.

话虽如此，但是仍有希望。在IBM试水了某些商业策略之后，我们这些人有机会更深入地了解应该做什么，避免什么。

Number one, it’s a business strategy, not a technology strategy, unless you are a research institute. What does it mean is the business leaders need to understand what AI is good at, (and not so good at) for their own business. Executives need to learn enough about AI technology to make strategic decisions, they can’t delegate AI to techie people only, otherwise, they might very well end up with a “technically feasible” but financially not viable, or business-wise undesirable AI case.

第一点，这是一个商业战略问题，而不是一个技术战略问题，除非说你拥有的是一家研究机构。这就意味着，企业的领导者需要了解在自己公司的业务上，AI擅长做什么，又不太擅长做什么。管理人员在制定战略决策之前需要充分了解AI技术，而是不将AI丢给技术人员来处理；否则，他们最终很可能得到“技术上可行”但经济上不划算或者业务层面不合理的AI应用。

Number two, AI does not simply replace human. (This one is really hard for executives who do not learn about AI, see #1 above). Lots of people managers are used to “motivate” people for doing things, and implicitly delegate decision making to their people. They use “soft rules” expecting human employees to fill the gaps between the lines. AI is different, leaders need to have a very clear vision of what are the rules of engagements, and what processes should entail. Fault tolerance needs to be designed into the process, instead of relying on the workforce’s common sense. Ambiguity with AI will be amplified and the result would be unpredictable, or uninterpretable.

第二点，AI并不能简单地取代人。（对于那些不了解AI的高管来说，理解这一点很难。）许多经理习惯于“激励”员工工作，并在暗中将决策权交给员工。他们使用“软规则”进行管理，并希望员工能够自动补齐“要求和期望”之间的差距。但涉及到AI时却并不能这样做。领导者需要对参与规则以及应该采取哪些流程有清晰的认识。容错要求需要添加到流程设计之中，而不能依赖员工的常识。AI相关的歧义会被放大，且结果将无法预测或无法解释。

Number three, AI can do things human could not (or very costly to do). For example, finding a face from thousands of faces, or monitor millions of occurrences of signal and identify abnormally. So designing a new business process that previously never existed could be the most valuable, yet most challenging responsibility for business leaders. For example, most people know NetFlix’s recommendation really helps us finding what we want to watch next. However, before Machine Learning, movie recommendations were not individualized — some “experts” wrote general critiques and we read then decide if we would risk a couple of hours of our lives to watch a title, AI allowed Netfix to know what I like to watch better than myself, it does not rely on critics articles, nor my willingness to read and trust them. This is a new business model Blockbuster did not realize and suffered the consequence.

第三点，AI可以完成那些人类无法完成或者实现成本非常高昂的事情。例如说，从数以千计的人脸中寻找特定的一张人脸，或者对上百万次的信号进行监控，并从中识别出异常信息。因此对于领导者们而言，设计一个以前从未存在过的、全新的业务流程具有非常高的价值，但它同时也是一项具有挑战性的责任。例如说，大部分人都知道Netflix（网飞）的推荐帮助我们找到了我们想看的电影。但在机器学习出现之前，电影推荐都是非个性化的。一些“专家”为电影撰写影评，我们阅读这些内容然后决定是否要冒着浪费数小时生命的风险观看电影。通过AI，Netflix能够比我自己还要了解我喜欢看什么；它不需要评论文章，也不需要我去阅读并相信这些玩意。然而，Blockbuster（百视达，影音娱乐公司）并没有意识到这一全新的商业模型，并为此付出了沉重的代价（在和Netflix的竞争中失败并破产）。

So is there a surefire shortcut for AI business strategy? Sorry to bare the tough news, I have not run into one successful AI business that attributed their success to some surefire rules. But there are some key commonalities of those who are successful.

那么在构建AI商业战略的过程中有什么可靠的捷径吗？我很抱歉的表示，尽管成功者在某些特性上是相似的，但我还没有遇到任何一家成功的AI公司将他们的成功归功于一些万无一失的规则。

These businesses tend to have Data Scientists (or ML experts) within their core executive team. The knowledge and understanding of AI can’t be a “skill” hanging low on the totem pole. Cross-discipline skills covering “what AI can do for my business” is a common competitive advantage that separated them from the rest (see Number one above).

这些企业往往在其核心执行团队中拥有几名数据科学家（或者机器学习专家）。对于人工智能的了解和理解不能成为图腾柱上的“技能”。涵盖了“人工智能可以如何帮助我的业务”等内容的跨学科技能才是（成功者们共同拥有的）竞争优势，它可以帮助企业甩开竞争对手。（见上文第一点）

They also find a problem that has not yet been solved that is big enough — sometime the problem might not be screaming at you, and need to be uncovered. For example, StitchFix found that average John and Joe can use stylist as well, but if you do user group survey, most likely people won’t tell you this was a need — I never thought about using a personal stylist, cause that sounds cheesy and “just not my type of people’s” way of living. It turned out that was a problem, that I just kept avoiding it. And StitchFix uncovered it. Identifying the problem is half the solution, in case of AI, it is probably 2/3 of the solution.

他们（成功的AI公司）还能发现那些尚未得到解决的大问题。尽管这些问题可能不会对你惊声尖叫，但是它们同样需要被关注，被解决。例如，StitchFix发现，如果你进行大规模的用户调查，你会发现大部分人似乎并不需要个人造型师。人们会告诉你，我从来没想过个人造型师的事情，因为这听起来就很俗气，而且“我”也不是需要这种生活方式的人。但事实是我不承认只是因为我在回避。不管是张三还是李四，他们都可以拥有自己的个人造型师。识别出一个问题就解决了问题的一半；在涉及到AI时，这一点可能解决了问题的三分之二。

Then there are the knowledge, tools, and ability to solve the problem. Remember no single AI algorithm can solve all big problems. To deliver functionality, AI is a (key) component of a connected system. There are thing Machine Learning achieves, but other tasks of the process chain are best suited for good old programming logic. There might be processes need to be changed so that AI can achieve the same goals (or better) as human used to do, but using different approaches. So a functioning decision-making unit that covers good old software development, integration, and process engineering, organizational change management on top of state-of-the-art AI understanding are needed.

然后就是解决问题的知识、工具和能力。请记住，没有任何一种AI算法可以解决所有的问题。AI只是为了提供某些功能而对系统进行连接的关键组件。有一些任务适合使用机器学习进行处理，但流程中的其他任务最好仍然使用良好的、旧式的编程逻辑进行处理。任务流程可能需要进行更改，以便AI能够通过不同的方法实现与人类相同（或者更好的）的目标。因此，我们需要一个功能强大的决策单元，它需要涵盖良好的、旧式的，软件开发、集成和过程管理等内容，它还需要基于对AI的理解对组织变更进行管理。

Last but not least, AI evolves really really fast. So an AI business strategy needs to understand this is not a competition of highest accuracy number to the 4 digits after the decimal point. The AI business strategy needs to be shaped with the future in mind: what if your core algorithm is no longer the best of the class, is your architecture flexible enough to change some components without completely collapse, is there sufficient barrier of entry other than your AI algorithms, will newer AI capabilities undermine your initiative in a surprising way…

最后同样重要的一点，人工智能的发展是非常迅速的。因此，在制定AI商业战略前你需要理解，AI业务的竞争不是小数点后最高精度的竞争。AI商业战略需要考虑到未来的发展与可能性：如果你的核心算法不再是同类产品中最好的，那么你的架构是否足够灵活，可以在不完全崩溃的情况下对某些组件进行更改；在AI算法之外，你是否还有其他的障碍可以阻碍竞争对手；当更新更强的AI算法出现之后，你的商业计划是否会被破坏。

In short, looking at what is available on the market to buy is not a winning AI strategy for most businesses. The future will likely see more consolidation, and the few lucky early adopters might “take it all”. So the stakes are pretty high for most businesses.

简而言之，对大多数企业而言，成功的AI商业战略并不是去调查市场上可供收购的产品有哪些。我们在未来可能会看到更多的整合，少数早期的幸运儿可能会“赢者通吃”。所以对大多数企业来说，风险都是非常高的。