一个活儿能不能交给 AI?先想清楚三个问题Can This Task Be Handed to AI? Three Questions to Think Through First
常有人问:哪些活儿可以交给 AI?
我认为,这个问题的答案,不该是一份「适合 AI 完成的任务」清单。
同样是「写合同」,起草一份合作意向书,和修改一份已经谈好价格的正式合同,能交给 AI 的程度差别很大。所以判断之前,有三个问题要想清楚。
第一个问题:如果错了,能不能马上发现?
AI 写的邮件、整理的会议纪要、做的长文摘要,你读一遍就知道对不对,错了当场改掉,代价是几分钟。这类活儿可以放心交。
麻烦的是错了不容易发现的。比如让它解读一份合同条款,再讲给客户听。它可能说得很流畅,读的人未必看得出哪里理解偏了,但是等对方找到了问题,就已经晚了。这类活儿要么不交,要么在发出去之前要有人核一遍。
第二个问题:错了的话,代价有多大?
能发现,不等于代价小。凡是涉及钱、权限、对外承诺的,比如给客户报价、批一笔退款、改系统里的权限,AI 可以起草、可以给建议,但是最后要由人来拍板,不能让它直接生效。
第三个问题:需不需要精确计算?
金额是多少、日期过没过期、有没有超标准,这类事要精确计算,对的结果只有一个。模型大多数时候也能算对,但没法保证每次都对,这类错又正好直接对应钱。所以算的事交给规则和程序,模型做前面那一步:听懂问的是什么,找到该用的数据或条款。
举个例子,我们做物业智能客服的时候就是这么做的。业主问「这个月费用怎么多了」,模型负责听懂他问的是哪一笔,再去调账单系统的接口查明细;查回来的数由系统给,模型只负责把它讲清楚。涉及减免、退费这类要动钱的,应该生成工单,交给人去批。
一个活儿,通常要拆开交
想清楚这三个问题,你会发现,真正的应用场景中,一个活儿很少整个交给 AI,也很少完全不能交。更常见的是拆开:理解和起草交给 AI,计算和校验交给规则,拍板留给人。
上面三个问题回答的是能不能交。值不值得交,还要再问一个问题:这个活儿多久做一次?一年只做一次的事,就算把工具搭好了,也省不下多少时间。所以刚开始,先从最常做的事里挑一件或几件,每天都在做的最好。流程跑顺了,再扩到其他任务上。
People often ask: which tasks can be handed to AI?
In my view, the answer to that question should not be a list of "tasks suitable for AI".
Take "writing a contract". Drafting a letter of intent and revising a formal contract whose price has already been agreed are very different in how much of the work can go to AI. So before deciding, there are three questions to think through.
The first question: if it goes wrong, can you spot it straight away?
An email written by AI, meeting minutes it has tidied up, a summary of a long document: you read it once and you know whether it is right. If it is wrong, you fix it on the spot, and it costs you a few minutes. Work like this can be handed over with confidence.
The trouble is work where mistakes are hard to spot. Say you ask it to interpret a contract clause and then explain it to a client. It may explain it very fluently, and the reader may not see where the understanding has gone wrong, but by the time the other side finds the problem, it is already too late. Work like this should either not be handed over at all, or someone has to check it before it goes out.
The second question: if it goes wrong, how costly is that?
Being able to spot a mistake does not mean the cost is small. For anything that involves money, permissions or commitments to outside parties, such as quoting a price to a client, approving a refund or changing permissions in a system, AI can draft and can make suggestions, but the final decision has to be made by a person. AI must not be allowed to put it into effect directly.
The third question: does it need exact calculation?
How much an amount is, whether a date has passed, whether something exceeds a set standard: these need exact calculation, and there is only one correct result. Models get them right most of the time, but they cannot guarantee being right every time, and these are exactly the errors that turn directly into money. So the calculating goes to rules and programs, and the model does the step before it: understanding what is being asked, and finding the right data or clause.
For example, that is how we did it when we built an AI customer service system for property management. When a homeowner asks "Why are my charges higher this month?", the model works out which charge they mean and then calls the billing system's interface to look up the details. The figures come from the system; the model only explains them clearly. Anything that moves money, such as a reduction or a refund, should generate a work order for a person to approve.
A task usually has to be split up
Once you have thought through these three questions, you will find that in real applications a task is rarely handed to AI whole, and rarely cannot be handed over at all. More often it is split up: understanding and drafting go to AI, calculation and checking go to rules, and the final decision stays with people.
The three questions above answer whether a task can be handed over. Whether it is worth handing over takes one more question: how often is the task done? Something done only once a year will not save much time, even once the tools are built. So at the start, pick one or a few of the tasks done most often, ideally ones done every day. Once that process runs smoothly, extend it to other tasks.