文章 · 存档
别做单一模型信徒,财务人真正要练的是驾驭AI的能力
财务圈有一种很微妙的气氛。一边,大家都知道AI已经不能再当热闹看了,谁都明白它正在进入工作现场;另一边,很多人又很容易陷进另一种焦虑:到底该押注哪个模型?
今天有人说这个最强,明天有人说那个更准,后天又冒出一个新王炸。于是很多财务人刚准备开始学,就先被选模型这件事困住了。选错了怕浪费时间,选晚了怕掉队,选多了又怕自己显得不专业。可这种纠结本身就已经把人带偏了。真正成熟的财务思维,本来就不该把自己交给单一工具,更不该把工作方式建立在对某一个模型的摩拜上。(用惯DeepSeek的我可没说你啊)
其实,连各家模型厂商自己都从来没有把一个模型通吃所有任务当成现实。
OpenAI在官方文档里明确把推理模型和通用GPT模型区分开来,强调它们行为方式不同;
Anthropic在模型选择指南里也写得很直白,选模型本来就是在能力、速度和成本之间做权衡;
Google在Vertex AI的模型说明里同样把高能力推理、低延迟、高吞吐这些能力拆开来定义,Pro、Flash、Lite各有各的任务边界。
厂商自己都在告诉你没有谁天生适合一切场景,模型差异本身就是产品设计的一部分。财务人如果还执着于“只信一个”,其实等于在主动缩小自己的工具箱。
这种固化思维在财务岗位上尤其容易吃亏。因为财务工作天然就不是单线程任务。你今天可能先要读政策、查口径、做解释,下午又要处理Excel里的复杂公式、跨表逻辑和数据清洗,晚上还得把一段分析写成老板看得懂的话。这里面有的任务看重速度,有的任务看重长文本理解,有的任务看重推理稳定性,有的任务又对数据边界和权限控制更敏感。你拿一个模型去硬扛所有场景,结果往往很像让一个会计既去做总账、又去谈融资、又去写税务争议回复、又去跑系统开发需求,表面上省事,最后每个环节都拧巴。
很多财务人真正栽过的坑都不是抽象的。比如做复杂公式时,一个模型前两步解释得头头是道,到了嵌套IF、跨表引用、日期口径切换和汇率折算同时出现的时候,开始一本正经地错;再比如清洗经营数据时,有些模型对文字总结很漂亮,可一碰到列名映射、结构还原、字段规则判断,稳定性就开始往下掉。你甚至会遇到这样的情况:同一个问题,A模型解释清楚但公式错,B模型公式大体对但没把业务口径讲明白,C模型反应很快,却只适合做第一轮草稿。问题从来不在于谁绝对最好,问题在于你是不是误以为只要找到一个最强模型,以后所有任务就都能交代出去。
说穿了,这篇文章真正想处理的就是这种行业里的执念:直击追新模型、只信一个工具的固化思维。
我越来越觉得多模型切换会成为财务人的一项新基本功。
它听起来像技术动作,骨子里却很像财务早就熟悉的工作逻辑。财务从来不会因为一个指标好看就放弃交叉验证,也不会因为一张报表做出来了,就不再看现金流、利润表、业务明细和异常波动。那为什么到了AI这里,反而愿意把全部信任压给一个模型?同一任务至少问三个大模型,这件事本质上很像对数字做复核。第一个模型给你初步思路,第二个模型帮你找逻辑漏洞,第三个模型帮你确认表达、结构或者替代解法。很多原本让人卡住的地方一旦放进多模型比对里,答案不一定立刻唯一,但错误会更容易暴露,盲点会更快浮出来,最终可用的结果反而更稳。
所以,真正值得练的,从来都不是忠诚于哪个模型,而是形成一套自己的调用秩序。查概念、做初稿、列提纲,可以先用快模型;碰到复杂公式、规则设计、逻辑推演,再切到更擅长推理的模型;涉及内部数据、合同、预算、经营口径时,就尽量放进权限和数据治理更清晰的企业环境里跑。你甚至可以让一个模型负责拆题,一个模型负责求解,一个模型负责审稿。这样做一点都不混乱,反而非常像成熟财务人的日常工作:分工、复核、对照、落地。
走到最后你会发现,财务人与AI之间最危险的关系,就是把自己变成某一个模型的附庸。今天它领先,你跟着它走;明天它掉队,你整套方法也跟着动摇。这样的工作方式看上去省脑子,实际上最没有安全感。真正稳的路是让自己掌握切换能力、比较能力、验证能力和组合能力。模型更新再快,工具迭代再猛,真正留在你身上的,永远是你调度工具的能力。
也正因为这样,我更愿意把一本真正有用的新书,写成一本开放的工具书,而不是某个模型的宣传册。技术会变,版本会变,排行榜会变,今天被捧上天的,过几个月也可能被新的替代。能陪财务人走得更久的,不该是迷信谁,而该是怎么判断、怎么切换、怎么组合、怎么落地。
我的这本新书想做的,正是把这件事讲透:它不绑架任何模型,也不要求读者向某个工具宣誓效忠,它更像把方向盘交回到财务人自己手里。到那时,AI就不再是谁牵着你走的风口,它会慢慢变成你手里那套安静、顺手、可替换、可升级的生产力系统。
Don't be a believer in only one model.
Don't be a believer in only one model. What finance people really need to learn is how to control AI
There is a very subtle feeling in the finance world. On one side, everyone knows AI is no longer just something fun to watch, and everyone sees it is coming into real work; on the other side, many people fall into another kind of worry: which model should I choose?
Today someone says this one is the strongest, tomorrow someone says that one is more accurate, and the next day a new “best one” shows up. So many finance people get stuck at the very beginning just because of choosing a model. If they choose wrong, they are afraid of wasting time. If they choose too late, they are afraid of falling behind. If they choose too many, they worry they look unprofessional. But this kind of struggle already leads people in the wrong direction. Real mature finance thinking should not give itself to one single tool, and should not build the whole way of working on believing in just one model. (If you always use DeepSeek, I'm not talking about you.)
In fact, even the model companies themselves never think that one model can do everything.
OpenAI clearly separates reasoning models and general GPT models in its official documents, and says they work in different ways;
Anthropic also says it very clearly in its guide that choosing a model is always a trade-off between ability, speed, and cost;
Google, in its Vertex AI model description, also separates high ability reasoning, low delay, and high volume, and says Pro, Flash, and Lite each have their own task limits.
Even the companies tell you that no model is naturally good for every situation, and the differences between models are part of the design. If finance people still insist on “trusting only one,” they are actually making their own toolbox smaller.
This fixed way of thinking is even more risky in finance jobs. Because finance work is never just one simple task.
Today you may read policies, check rules, and give explanations, in the afternoon you may deal with complex Excel formulas, cross-table logic, and data cleaning, and at night you may need to write analysis in a way the boss can understand.
Some tasks need speed, some need strong reading of long text, some need stable reasoning, and some need strict data control and permissions.
If you use one model for all these situations, it is like asking one accountant to do general accounting, talk about financing, write tax dispute replies, and also manage system development. It looks simple, but in the end everything feels awkward.
The real problems many finance people meet are not abstract.
For example, when doing complex formulas, one model explains the first steps very well, but when nested IF, cross-table references, date rules, and exchange rates all appear together, it starts making serious mistakes;
Or when cleaning business data, some models are very good at writing summaries, but when they face column mapping, structure rebuilding, and rule checking, they become less stable.
You may even see this: for the same question, model A explains clearly but gives a wrong formula, model B gives a mostly correct formula but does not explain the business rule well, and model C is fast but only good for a first draft.
The problem is never who is the absolute best. The problem is whether you think that once you find the strongest model, all tasks can be given to it.
To say it simply, what this article really wants to deal with is this fixed thinking in the industry: chasing new models and trusting only one tool.
I more and more feel that switching between multiple models will become a basic skill for finance people.
It sounds like a technical action, but inside it is very similar to finance logic. Finance people never trust just one number without checking others, and never stop after one report is done without looking at cash flow, profit, business details, and unusual changes.
So why, when it comes to AI, do people want to trust only one model?
For the same task, asking at least three models is like checking numbers again.
The first model gives you an initial idea, the second helps you find logic problems, and the third helps you confirm the wording, structure, or other solutions.
Many problems that used to block you become clearer when you compare multiple models. The answer may not be one right away, but mistakes show up faster, blind spots appear sooner, and the final result becomes more reliable.
So what is really worth learning is not being loyal to one model, but building your own way of using them.
When checking concepts, making first drafts, or writing outlines, you can use a fast model first;
When facing complex formulas, rule design, or logic thinking, switch to a model that is better at reasoning;
When dealing with internal data, contracts, budgets, and business rules, try to use an enterprise environment with clear permissions and data control.
You can even let one model break down the problem, one solve it, and one review it.
This is not messy at all. It is very much like how experienced finance people work: divide tasks, check again, compare, and then apply.
In the end, you will see that the most dangerous relationship between finance people and AI is becoming dependent on one model.
Today it is leading, so you follow it. Tomorrow it falls behind, and your whole way of working also becomes unstable.
This way of working looks easy, but actually gives the least sense of security.
The stable path is to learn how to switch, compare, check, and combine.
No matter how fast models change or tools improve, what stays with you is always your ability to use and control tools.
Because of this, I would rather write a truly useful new book as an open toolbox, not as a promotion for one model.
Technology will change, versions will change, rankings will change, and what is popular today may be replaced in a few months.
What can stay with finance people for a long time is not trusting one model, but knowing how to judge, switch, combine, and apply.
What my new book wants to do is to explain this clearly: it does not tie you to any model, and it does not ask you to follow any tool. It is more like giving the steering wheel back to finance people.
At that point, AI will no longer be something that leads you, but will slowly become a quiet, easy-to-use, replaceable, and upgradeable system that helps your work every day.