王峰 AI × 经营 约一次诊断

总目 / 文章 / WF-0436

文章 · 存档

财务人学AI先别追星,先学会评判

财务人学AI最容易走偏的一步就是先去追模型。今天听说这个模型会推理,明天看到那个模型会做表,后天又来一个新工具,演示视频里它像天神下凡,三分钟读懂财报,五分钟做完分析,十分钟还能给老板写经营建议。看得人热血上头,仿佛只要把这些名字背熟,自己就已经半只脚踏进未来了。可一回到办公室,账还是那本账,表还是

问题就出在这里。财务场景和图片、视频、营销文案那类场景,根本不是一个物种。图片糊一点,也许还能叫氛围;视频假一点,也许还能叫风格;广告夸一点,转化高就行。财务不吃这套。财务要的是口径清楚、逻辑闭环、数字能对、结论能落地。差一位小数,可能预算判断就偏了;少看一个口径差异,管理动作就会失焦;把一个异常波动讲成行业共性,老板就可能在错误方向上越走越快。财务场景对精确性的要求,决定了AI在这里不能只靠“像”,更不能只靠“听起来很有道理”。

偏偏AI最擅长制造的第一印象,就是“它很有道理”。第一次拿一份报表给它分析,确实容易惊艳。收入增长、毛利下滑、费用上升、现金承压,它都能说,甚至说得很顺。词也漂亮,结构也完整,语言比不少职场汇报还利索。可这类惊艳往往撑不过几轮。看多了就知道,它的分析高度依赖概率,依赖常见套路,依赖过往语料里那些“像分析”的表达。它会抓得到表面变化,抓不太住关键变化;它会讲共性,抓不太准个性;它会把数字排成一段像模像样的话,抓不太住那个真正决定经营走向的小钉子。

这个“小钉子”恰恰是财务工作里最值钱的地方。收入涨了,到底是销量涨了、单价涨了、渠道压货了,还是确认时点变了?利润掉了,到底是原材料吃掉了、人工效率掉了、费用前置了,还是某个大客户结构变了?应收账款上升,到底是业务扩张带来的正常结果,还是客户质量已经开始变坏?库存增加,到底是在为旺季备货,还是车间已经在悄悄积压?这些东西,AI很容易讲一圈大路货,真正能把针插进去的那一下,经常还得靠人。因为关键经营变化往往不长在报表表面,它躲在业务动作、会计口径、组织行为和管理习惯的交界处。那里没有那么多标准答案,只有具体企业的具体脾气。

所以财务人学AI,先学评判比先学崇拜重要得多。先别急着问“哪个模型最强”,先问“它这次到底对不对”。别急着被一段漂亮分析打动,先拆开看它有没有偷换概念,有没有把同比环比混着讲,有没有把结果当原因,有没有把行业常识硬套进企业个案里。AI现在特别会一件事,就是把七十分的话讲出九十分的气势。对没经验的人,这很致命,因为你很容易把流畅感误认成准确性,把结构感误认成洞察力,把会说话误认成会判断。财务人一旦在这里失守,后面学得越热闹,离真正的能力反而越远。

真正靠谱的学习路径,反而没那么花哨。把需求描述清楚,让AI帮你生成代码、生成公式、生成清洗逻辑、生成自动化流程,再让这些代码和逻辑在固定场景里稳定运行。这条路为什么更可靠?因为它把AI从“神谕机”拉回了“工具机”。你不再求它直接给你答案,而是让它参与构建一个可验证、可重复、可纠错的处理过程。今天让它写一个凭证生成脚本,明天让它做一个科目映射逻辑,后天让它搭一个预算对比的异常提醒规则。你每跑一次,就知道哪里对了,哪里错了,哪里需要补条件,哪里需要加限制。这个过程很朴素,却特别适合财务,因为财务真正需要的,从来都不是一句聪明话,而是一套能稳定跑下去的机制。

这也解释了一个常见误区:不少财务人以为学AI,就是学提示词,学怎么问,学怎么让它“一次出神图”。这种学法当然能带来一点小快乐,可快乐很短。财务场景真正拉开差距的,不是谁更会哄模型,而是谁更会定义问题。你要说清楚取数范围,说清楚口径,说清楚分类规则,说清楚异常处理,说清楚输出格式,说清楚哪些地方绝对不能猜,哪些地方可以给备选方案。描述能力越强,AI越有用;边界给得越清楚,结果越稳。到最后你会意识到,自己学的表面上是AI,骨子里练的还是财务基本功,是逻辑、口径、流程、控制点,是把模糊任务压成明确步骤的能力。

还有一点,财务人学AI,一定得养成“多用、多比、多判断”的习惯。一个模型给的答案,别急着信,换一个再看;同一个问题,改几个提示再看;分析结论,回到原始数据核一下;自动化脚本,拿历史样本跑几轮;生成的经营结论,放到真实业务里问一句站不站得住。这个“比”的动作非常重要,因为它能迅速把迷信打碎。你一比较,就知道模型擅长什么、偷懒在哪里、胡编喜欢怎么编、哪些表达只是气氛组,哪些输出能进入工作流。没有比较,人很容易崇拜;有了比较,工具才会落回工具的位置。

财务人最不该学会的,就是把AI当偶像。今天追这个,明天追那个,像在模型排行榜里追星,看谁红就扑向谁。追到最后,收藏夹越来越满,工作台还是老样子。真正高效的人,往往没那么激动。他们会盯着自己的高频场景:数据清洗能不能更快,报表生成能不能更稳,核对逻辑能不能自动跑,分析初稿能不能先出一版,异常波动能不能先预警,重复劳动能不能再砍掉一截。他们关心的是生产力,不是明星榜;关心的是结果能不能复用,不是模型名字好不好听。

说到底,财务人和AI之间最健康的关系就是你来做裁判,它来做工具。你负责定规则、查漏洞、做判断、扛结果,它负责提速、试算、生成、执行。这个位置一旦摆正,AI就很好用;位置一旦摆歪,AI就会把人带进一种很滑稽的状态:明明工作离不开精确,自己却天天沉迷于概率性的惊艳表演。

所以财务人学AI,先别急着追星。先把评判力练出来,先把问题定义清楚,先把比较习惯养出来,先让代码和逻辑在真实场景里跑起来。谁先做到这一点,谁就不会被模型牵着鼻子走。工具更新再快,也只是外面的风。判断力一旦长在自己身上,手里才真正有东西。

If want to learn AI, Learn how to judge first.

If finance people want to learn AI, do not chase stars first. Learn how to judge first.

The easiest wrong step for finance people when they learn AI is this: they first go chase models. Today they hear this model can think. Tomorrow they see that model can make tables. The day after that, a new tool comes again. In demo videos, it looks like a god coming down from the sky. It can read a finance report in three minutes, finish an analysis in five minutes, and write business advice for the boss in ten minutes. People get excited when they see this. It feels like if they know these names well, then one foot is already in the future. But when they go back to the office, the books are still the same books, the tables are still the same tables, the data is still the same dirty data, and the analysis due at month-end is still stuck in their hands. They watched a lot of excitement, but the work did not get much lighter.

The problem is here. Finance work is not the same kind of thing as pictures, videos, or ad writing. If a picture is a little unclear, maybe people can still call it a mood. If a video is a little fake, maybe people can still call it a style. If an ad says too much, that is fine as long as it gets results. Finance does not work like that. Finance needs clear rules, full logic, numbers that match, and a result that can really be used. If one small number is off, the budget judgment may go the wrong way. If one rule difference is missed, the management move may lose focus. If one strange jump is explained like it is a normal thing in the whole field, the boss may move faster and faster in the wrong direction. Finance needs accuracy. Because of that, AI here cannot live only by “it looks right.” It also cannot live only by “it sounds very smart.”

But the first thing AI is best at making is exactly this feeling: “it sounds very smart.” The first time you give it a report and ask for analysis, it really can amaze you. It can talk about sales going up, gross profit going down, cost going up, and cash getting tight. It can even say these things very smoothly. The words look nice. The shape is full. The language is even smoother than many work reports from real people. But this kind of surprise often does not last long. If you look more, you will see that its analysis depends a lot on chance, on common patterns, and on old language in past data that “looks like analysis.” It can catch the changes on the surface, but it often misses the key change. It can talk about common things, but it often misses the special thing. It can line up numbers into a piece of text that looks good, but it often misses the small nail that really decides where the business will go.

And this “small nail” is exactly the most valuable part of finance work. If sales go up, is it because more units were sold, because the price went up, because more goods were pushed into the channel, or because the time of booking the sale changed? If profit falls, is it because raw materials ate it up, because worker speed fell, because some cost came earlier, or because the customer mix changed? If money to collect goes up, is it a normal result of business growth, or is customer quality already getting worse? If stock goes up, is it because the company is getting ready for the busy season, or because goods are already quietly piling up in the shop? For things like these, AI can easily talk in a big circle of common words. But that one move that really hits the point still often has to come from a person. That is because the key business change often does not sit on the face of the report. It hides where business action, accounting rules, team behavior, and management habits meet. There are not many ready-made right answers there. There are only the real habits of one real company.

So when finance people learn AI, learning how to judge is much more important than learning how to admire. Do not rush to ask, “Which model is the strongest?” Ask first, “Was it right this time or not?” Do not rush to be moved by one pretty piece of analysis. Break it open first. Did it quietly change the meaning? Did it mix year-on-year and month-on-month? Did it treat the result like the cause? Did it force general industry ideas into one company case? AI is now very good at one thing: it can make seventy-point words sound like ninety-point words. For people with less experience, this is dangerous. It is easy to mistake smooth words for accuracy, good shape for insight, and being good at talking for being good at judging. Once finance people lose here, then the more excitedly they learn later, the farther they may move from real skill.

The truly good learning path is not that fancy. First, describe the need clearly. Then let AI help write code, write formulas, write cleaning logic, and build auto work steps. Then let this code and this logic run in one fixed scene in a steady way. Why is this road more reliable? Because it pulls AI down from a “magic answer box” back to a “tool box.” You stop asking it to directly give the answer. Instead, you let it help build a process that can be checked, repeated, and fixed. Today you let it write a script for making vouchers. Tomorrow you let it build a logic for account mapping. The day after that, you let it build a warning rule for strange budget differences. Each time you run it, you see where it was right, where it was wrong, where you need more rules, and where you need more limits. This process is very plain, but it fits finance very well, because what finance truly needs was never one smart sentence. It needs a system that can run in a steady way.

This also explains one common wrong idea. Many finance people think learning AI means learning prompt words, learning how to ask, and learning how to make it “give a magic picture in one try.” Of course this can bring a little short joy. But that joy does not last long. In finance work, the real gap does not come from who is better at pleasing the model. It comes from who is better at defining the problem. You need to clearly say the data range, clearly say the rule, clearly say the class rules, clearly say how to handle strange cases, clearly say the output form, clearly say where it must never guess, and clearly say where it may give other choices. The stronger your way of describing is, the more useful AI becomes. The clearer the limits you give, the more steady the result becomes. In the end, you will see that what you learn on the outside is AI, but deep inside you are still training the basic skill of finance. You are training logic, rules, process, control points, and the power to press a fuzzy task into clear steps.

There is one more thing. If finance people want to learn AI, they must build the habit of “use more, compare more, judge more.” When one model gives an answer, do not rush to trust it. Change to another one and look again. Ask the same question in a few different ways and look again. Check the analysis result against the raw data. Run the auto script on old samples a few times. Take the business answer it gives and ask one real question in the real business: does this really stand up or not? This act of “compare” is very important, because it quickly breaks blind faith. Once you compare, you will know what the model is good at, where it is lazy, how it likes to make things up, which words are only there to sound good, and which results can really enter your work flow. Without comparing, people easily admire. With comparing, the tool goes back to being a tool.

What finance people should least learn is to treat AI like an idol. Today they chase this one. Tomorrow they chase that one. It is like chasing stars in a model ranking list. Whoever is hot, they run to that one. In the end, their saved links grow more and more, but their work desk still looks the same. Truly efficient people are often not that excited. They look at their own high-use scenes: can data cleaning get faster, can report making get steadier, can checking logic run by itself, can the first draft of the analysis come out first, can strange changes be warned first, and can one more part of repeated work be cut away? What they care about is work power, not the star list. What they care about is whether the result can be used again, not whether the model name sounds good.

In the end, the healthiest relationship between finance people and AI is this: you are the judge, and it is the tool. You make the rules, check the holes, make the judgment, and carry the result. It speeds things up, tests, makes drafts, and does the work. Once this place is set right, AI is very useful. Once this place is set wrong, AI will pull people into a very funny state: their work cannot live without accuracy, but every day they are lost inside surprise shows based on chance.

So if finance people want to learn AI, do not rush to chase stars first. First build your power to judge. First define the problem clearly. First build the habit of comparing. First let code and logic run in real work scenes. Whoever does this first will not be led around by the nose by models. Tools may change very fast, but that is only the wind outside. Once judgment grows inside you, then you truly hold something in your hand.