2027 年分析从业者如何避免被 AI 取代

我仍然记得第一次使用 Claude Code 时有多么兴奋。
只需一个提示词,它就把我多年前设计的一个算法适配到了我当前的项目,并部署到我的服务器上,随时可以连接客户的 Google Sheets。
这是一次颠覆性的改变,让我能以半自动的方式,把我花数月开发的任何优化算法产品化并部署。
当时,我为它带来的生产力提升感到兴奋。
而 18 个月后的今天,我每天醒来都会问自己:“我的价值在哪里,我怎样才能不被 AI 取代?”
如今,AI 智能体承担了分析价值链中越来越大的一部分,因此我不得不弄清楚,作为一名供应链与分析专家,我在哪些地方仍然具有价值。
幸运的是,我在制造业和物流运营中落地 AI 的经验告诉我,这类项目很少能达到人们对自动化的预期。
我见过一位计划员放弃了他用于 Excel 的 AI 助手,因为它每天对同一个问题给出不同的答案;也见过一家公司里的各个团队用 36 种不同的、没有文档记录的方式衡量同一个 KPI。
没有任何智能体能独自解决这些问题,而这正是我们(作为分析专业人士)保持自身价值的机会。
我 2027 年的目标:确保自己不会变成我们所说的“肉身代理”(meat proxy)——一个唯一工作就是把请求转交给 AI、再把答案转回来的人。
在本文中,我将分享我在中型和大型企业里管理智能体项目的经验,以说明 AI 在企业环境中的局限。
这将帮助我们理解 AI 无法复制的那些根本技能:
了解你的数据所经过的整条链路:数据在企业系统中如何生成、转换和使用
把需求转化为团队真正需要的工具
最后,我会给出个人建议,说明如何利用我在这个博客上分享的免费资源来培养这些技能。
落地 AI 的过程让我看清了它的局限
先交代一下背景:我是一名供应链从业者,学习数据科学是为了帮助制造和物流运营优化其流程,以提升盈利能力和可持续性。
自 2020 年以来,我一直在这个博客中分享真实用例,展示分析技术如何帮助解决某个运营或业务问题。
我的第一篇文章介绍了我构建的一套算法,用于缩短仓库操作员的行走距离(图片由 Samir Saci 提供)
当时,我的想法是让供应链从业者了解如何用数据科学解决他们在公司中遇到的问题。
反响非常热烈:数据科学提供了一个用高等数学为企业创造价值的机会。
但自 2025 年以来,焦点已转向用 AI 构建工作流,在这些工作流中,智能体理应自主处理极其复杂的业务和运营流程。
遗憾的是,在我所看到的项目中,这些方案并未兑现其承诺。
因为支撑一家公司运转的系统(ERP、CRM、WMS……)——对分析和 AI 专家来说不幸的是——并非即插即用的方案,无法直接适配那些需要统一、清洗后数据的模型。
障碍一:AI 无法仅凭查看数据库就推断出系统是如何被使用的
企业部署 AI 解决方案时面临的第一个障碍是流程映射,也就是我们现在所说的本体(ontology)。
AI 与它本应运行的系统之间的壁垒(图片来源:Samir Saci)
在一家典型的消费品公司里,管理制造、物流、财务和人力资源的系统并不是一个拥有集中式文档化数据来源的铁板一块。
其中大多数是遗留系统,多年来为了满足各团队按自己方式生成和使用数据的需求,做过许多专门开发。
四个系统,四种生成和使用数据的方式(图片来源:Samir Saci)
让我用一家消费品公司项目中的例子来说明。
我的职责是设计一款工具,用来找出门店送货延迟的根本原因。
这一切始于一个简单的任务:定义每次交付是准时还是延迟。
从理论上看,公式很简单:取企业资源规划(ERP)系统中名为 Delivery Time 的字段,与 Requested Delivery Time 字段进行比较,生成一个布尔值(准时为 1,否则为 0)。
按时按量交付:OTIF -(图片由 Samir Saci 提供)
在与物流团队开了 12 周的研讨会后,我发现他们用了 36 种不同的方式来定义这个布尔值。
法国团队采用发货时间(Shipping Time),因为门店会派出自己的卡车
巴西团队只有在货物于要求交付时间之后 10 天到达时,才会将其标记为延迟,因为他们使用海运,而海运的交付周期波动较大
同一次交付,三种延迟定义(图片由 Samir Saci 提供)
这些边缘情况在任何地方都没有文档记录。
它们需要一个人——也就是我——去询问其他人,并将其回答转化为对系统内外所执行的数据转换的建模。
AI 智能体仅靠查看数据库中的事务数据,无法感知这种复杂性。
花了 12 周的研讨会才做出这张表(图片由 Samir Saci 提供)
这需要人类软技能,而智能体目前尚不具备:
在公司内部建立人脉,以便接触到掌握你所需信息的合适人选
主导并组织与运营和业务团队的研讨会,以收集正确的信息(例如,你如何定义门店延迟交付?)
行业经验往往不会记录在网上(例如,根据我过去的项目经验,我知道 ERP 中的同一个字段可能被两个团队以不同方式使用)
如果你从团队中提取出这些知识,你就为落地任何与公司系统相连的分析或智能体产品打下了基础。
遗憾的是,仅仅部署解决方案并不总能推动用户采用。
障碍2:AI 模型具有随机性,当用户需要稳健可靠的解决方案时,这可能成为问题。
在数据科学领域,我们很快就认识到,由于其统计特性,大语言模型(LLM)对同一输入可能产生不同的输出。
但这种行为很少适合那些以可预测性和可复现性为关键的环境。
一家制造企业为其计划团队配备了 AI 助手以支持日常工作,之后该企业联系了我。
其供应链总监希望让手下的计划人员能够借助 Claude Cowork 等工具创建自动化工作流,从 Excel 文件中提取数据,并回答战略性和运营性问题:
我们下周应生产多少件产品才能满足客户需求?
我们需要多少个生产批次才能将库存减少 15%?
他的想法是借助智能体实现「分析即服务」,让智能体按需回答规划人员的问题。
这个想法是:用智能体提供分析即服务(图片由 Samir Saci 提供)
供应链总监:「这让运营陷入了混乱。」
一位规划人员坦言,她又换回了自己的 Excel 规划表格,因为「同一个问题,AI 今天和明天给出的结果可能不一样」。
我们一起查看了该智能体生成的各种分析结果,发现对于同一个问题和同一个输入 Excel 文件,它生成了两个不同的 Python 脚本。
随机模型无法可靠地生成稳健且可复现的解决方案——(图片由 Samir Saci 提供)
我决定组织研讨会来了解这些规划人员的需求,我问的是:
他们需要哪些类型的分析?
涉及哪些输入和输出数据?
哪些任务需要完全自动化,为什么?
结论很明确:他们不需要 AI。
用确定性代码编写的分析和数据科学工具能够恰好满足他们的需求:
用 Python 编写的库存管理规则,如本文所述,用于计算满足生产线需求所需的最低原材料安全库存
像 Wagner-Whitin 这样的动态算法可以告诉他们每批生产多少单位,以交付客户订单,同时避免积累过多库存
这些解决方案是确定性的(运行 100 次,你总会得到相同的输出),并且由一个人(我自己)设计,他花时间了解了运营背景及其内部和外部约束。
供应链总监:“我们需要有人来质疑我们的需求,提出一套结构化流程,从数据中生成洞察,并设计出我们可以依赖的稳健工具。”
为了说明他们对 AI 工作原理的理解有多么匮乏,当我解释说这个助手同时也在生成代码来回答他规划团队的问题时,这位总监大为惊讶。
另一个问题在于,智能体只掌握规划者在提示词中共享的那点上下文,而我这边则花了几个小时向他团队提出数百个问题。
这两个例子有一个共同点:一名技能高超的工程师能够弥补 AI 的局限,并通过把运营与业务需求转化为分析解决方案来打开局面。
工程师可以发挥作用的环节 ——(图片由 Samir Saci 提供)
随着企业将智能体式工作流引入越来越多的流程,正是这些技能能让你保持自身价值。
现在的问题是:你该如何培养这些技能?
原文较长,译文为节选;译文由英文原文机器翻译,术语按行业惯例统一,可在文末展开原文核对。
查看英文原文(How to Avoid AI Replacement as an Analytics Professional in 2027)
I still remember how excited I was the first time I used Claude Code.
From a single prompt, it adapted an algorithm I designed years ago to my current project and deployed it on my server, ready to connect to the customer’s Google Sheets.
It was a game changer, giving me a semi-automated way to productise and deploy any optimisation algorithm I took months to develop.
At the time, I was excited by the productivity gains this brought.
Now, 18 months later, I wake up every day with the question: “Where is my value, and how do I avoid being replaced by AI?”
AI agents now handle a growing part of the analytics value chain, so I had to work out where I remain relevant as a supply chain and analytics expert.
Fortunately, my experience implementing AI for manufacturing and logistics operations taught me that these projects rarely meet automation expectations.
I have seen a planner drop his AI assistant for Excel because it answered the same question differently each day, and teams in one company measure the same KPI in 36 different undocumented ways.
No agent could have solved these on its own, and this is our chance (as analytics professionals) to stay relevant.
My goal for 2027: make sure that I don't become what we call a "meat proxy", a human whose only job is to pass requests to an AI and its answers back.
In this article, I will share my experience managing agentic projects in mid-size and large companies to show the limits of AI in a corporate environment.
This will help us to understand the fundamental skills AI cannot replicate:
Learning the whole chain your data passes through: how data is generated, transformed and used in corporate systems
Turning a request into the tool the team actually needs
I will conclude with my personal advice on how to build these skills using the free resources I share on this blog.
Implementing AI taught me what its limits are
For context, I am a supply chain professional who learned data science to help manufacturing and logistics operations optimise their processes for profitability and sustainability.
Since 2020, I have shared real use cases in this blog where analytics helps solve an operational or business problem.

My first article presented a set of algorithms I built to reduce the walking distance of warehouse operators (Image by Samir Saci)
At the time, the idea was to educate supply chain professionals on using data science to solve the problems they face in their companies.
Interest was huge: data science was an opportunity to use advanced mathematics to create value for companies.
But since 2025, the focus has shifted to using AI to build workflows where agents are supposed to autonomously handle very complex business and operational processes.
Unfortunately, in the projects I have seen, these solutions fail to deliver on their promise.
Because the systems (ERP, CRM, WMS, ...) that run a company are, unfortunately for analytics and AI experts, not plug-and-play solutions for models that require harmonised, cleaned data.
Obstacle 1: AI cannot guess how systems are used by only looking at the database
The first obstacle companies deploying AI solutions face is the mapping of the processes, what we now call ontology.
The wall between AI and the systems it is supposed to run (Image by Samir Saci)
In a typical consumer goods company, systems managing manufacturing, logistics, finance and HR do not represent a monolithic bloc with a central source of documented data.
Most are legacy systems, with many specific developments over the years to meet the needs of teams that generate and use data their own way.
Four systems, four ways to generate and use data (Image by Samir Saci)
Let me illustrate with an example from a project at a consumer goods company.
My role was to design a tool to find the root cause of late store deliveries.
This started with a simple task: defining whether each delivery was on time or late.
On paper, the formula is simple: take the field named Delivery Time in the Enterprise Resource Planning (ERP) and compare it with the field Requested Delivery Time to generate a boolean (1 if on time, 0 otherwise).
On Time In Full: OTIF - (Image by Samir Saci)
After 12 weeks of workshops with the logistics teams, I discovered that they used 36 different ways to define this boolean.
France teams take the Shipping Time because stores send their own trucks
Brazil team flags a delivery as late only if the shipment arrives 10 days after the Requested Delivery Time because they use sea freight that has variable lead times
One delivery, three definitions of late (Image by Samir Saci)
These edge cases aren't documented anywhere.
They require a human, myself, to ask other humans and translate their answers into a model of the data transformations performed inside and outside the systems.
An AI agent can't sense this complexity by looking only at transactional data in the database.
It took 12 weeks of workshops to come up with this table (Image by Samir Saci)
This requires human soft skills that agents do not have yet:
Build connections inside a company to have access to the right person who has the information you need
Lead and structure workshops with operational and business teams to collect the right information (e.g., how do you define a late store delivery?)
Industry experience that is often not documented online (e.g., I knew the same field in the ERP can be used differently by two teams based on my past projects)
If you extract this knowledge from the team, you have built the foundation to implement any analytics or agentic products connected to the company's systems.
Unfortunately, deploying solutions isn't always enough to drive user adoption.
Obstacle 2: AI models are stochastic, and this can be a problem when users need robust solutions
In the data science world, we quickly learned that large language models (LLMs) can produce different outputs for the same input because of their statistical nature.
But this behaviour rarely fits environments where predictability and reproducibility are key.
A manufacturing company contacted me after equipping its planning teams with AI assistants to support their daily tasks.
Their supply chain director wanted to enable his planners to create automated workflows with tools like Claude Cowork to pull data from an Excel file and answer strategic and operational questions:
How many units should we produce next week to meet customer demand?
How many production batches do we need to reduce the inventory by 15%?
His idea was to implement analytics-as-a-service with agents that would answer planners’ questions on demand.
The idea: analytics-as-a-service with agents (Image by Samir Saci)
Supply Chain Director: “It became an operational chaos.”
A planner admitted she switched back to her Excel planning spreadsheet because “AI can give me different outputs for the same question from one day to another.”
We looked together at the different analyses generated by the agent and discovered that, for the same question and the same input Excel file, it generated two different Python scripts.
Stochastic models are not reliable for generating robust and reproducible solutions - (Image by Samir Saci)
I decided to organise workshops to understand the needs of these planners, asking:
What type of analyses do they need?
What input and output data are involved?
Which tasks require full automation and why?
The conclusion was clear: they did not need AI.
Analytics and data science tools written with deterministic code can exactly meet their needs:
Inventory management rules written in Python, explained in this article, to find the minimum safety stock of raw materials required to meet production lines’ needs
Dynamic algorithms like Wagner-Whitin can tell them how many units to produce in each batch to deliver customer orders while avoiding accumulating too much inventory
These solutions are deterministic (run them 100 times and you will always get the same output) and designed by a human (myself) who took the time to understand the operational context and their internal and external constraints.
Supply Chain Director: “We needed someone to challenge our requests, propose a structured process to generate insights from data and design robust tools we can rely on.”
To illustrate how little they understood about how AI works, the director was surprised when I explained that the assistant was also generating code to answer his planners’ questions.
The other problem was that agents only had the context the planner shared in the prompt, while on my side I spent hours asking his team hundreds of questions.
These two examples had one thing in common: a highly skilled engineer could compensate for AI's limits and unlock situations by translating operational and business requests into analytics solutions.
Where the engineer can help — (Image by Samir Saci)
These are the skills that will keep you relevant as companies introduce agentic workflows into more and more processes.
The question now is: How do you develop these skills?
In the next sections, I share my vision, based on dozens of analytics and AI projects.
Skill 1: Learn the whole chain your data passes through
In the first example, we learned that we can’t understand how a company is run by looking at transactional records in a database.
The first skill I developed comes from my years as a continuous improvement engineer: understanding a process end to end.
Each team in front of me knew its own step, and the tool had to hold all of them — (Image by Samir Saci)
When I was working in the logistics industry, I often understood the customer’s supply chain better than their own teams.
I found out by asking them how their processes worked, and they could not answer.
Because my role was to develop a holistic view: understand how warehouse and transportation operations work as one chain.
Each person in front of me thought about their own step.
As an analytics expert, you need to follow data along its life cycle:
Input: how data is entered into the systems
Processing: how information is stored and updated in the systems
Output: how teams read the data and what they do with it
This framework will help you structure your workshops, ask the right questions of the teams, and set expectations before you start building the solution.
You don't need ten years of supply chain experience to ask the right questions.
Follow the data along its life cycle (Image by Samir Saci)
Let’s go back to our FMCG company with its 36 formulas to measure if a store has been delivered late:
Input: the requested delivery date is included in store orders and may be modified by customer service admins if the customer asks for a new date
Processing: French teams have implemented a script that modifies the requested delivery date if the delay is below the service level agreement (SLA)
Output: American teams do not look at the delivery date but at the actual ship date when flagging late deliveries because they are not responsible for transportation
These three points summarise 12 weeks of intense workshops with the teams to extract all this knowledge.
One on-time delivery measure, followed from input to output (Image by Samir Saci)
The supply chain director was surprised by the complexity of his distribution chain and admitted that no one in the company expected measuring on-time delivery to be so complex.
My question to you: “Do you think that plugging Claude Code into the SQL database of the ERP would help to get these insights?”
No, in most of the AI projects I've seen, we still needed someone who knows what each team takes in, changes, and hands to the next.
In most organisations I've worked with, this person doesn't exist, which makes this a management problem more than a technical one.
Therefore, companies will always need someone with analytics skills who can collect process knowledge from business and operational teams, structure it, and translate it into data that agentic workflows can use.
Skill 2: Turn the request into the tool the team needs
I am convinced companies don’t need stochastic parrots designed to say yes to any user request.
Therefore, the second skill you need is translating (and challenging) a business request into the tool the team needs.
Let me go back to the second example: the planners’ AI assistants answered their requests based on a limited context.
They didn't ask for additional information or run a workshop to clarify the needs; they generated the most probable output from the prompt text.
The assistant works from the prompt, not from the context around it (Image by Samir Saci)
And they didn't challenge the request: the planners didn't need ad hoc analyses generated by code rewritten each time.
An AI assistant answers the request; it does not ask if it is the right one.
Let me illustrate this with a project I did with a mechanical parts manufacturer in Switzerland.
The supply chain director found me through my article about production planning optimisation and asked for a similar solution to reduce their costs.
At that time, he could have asked Claude Code to pull the Python implementation of the tool I shared on my GitHub and run it with their data.
They needed a unified process, not a magic algorithm.
But when I looked at how planners worked, the algorithm was not the first problem.
Customer orders arrived by email, directly from customers or from the commercial teams, in one mailbox per planner group.
One mailbox, one group of planners, one Excel file per lane (Image by Samir Saci)
Planners then created production lots by hand in Excel, often with too much inventory or too many batches, because orders for SKUs that could share batches weren't grouped.
Instead of a standalone algorithm, we designed a procedure in three steps, each validated with the planners before the next:
Data: all customer orders collected in one place, Anaplan, instead of one mailbox per group of planners
Optimisation: a Wagner-Whitin lot-sizing algorithm grouping the orders of SKUs, or families of SKUs, that can share batches
Adoption: the optimisation runs inside the planners’ daily workflow, with a dashboard and an email workflow, instead of being a separate tool
They asked for an algorithm; I designed a procedure (Image by Samir Saci)
This answered their needs: production costs are estimated to drop by 30%, mainly thanks to lower inventory.
Because the optimisation is deterministic code packaged as a service, other teams reused it for their own simulations, with a central logging system to track all the analyses.
A coding assistant would write the algorithm they asked for and would not question the process.
Before you build anything, ask what decision the tool will change and who will make it.
So far, the agents don’t do this.
That is why I believe this skill matters if we want to be antifragile in this new AI-first era.
But how do we develop these skills?
My skills come from years of experience designing and optimising supply chain solutions across 60+ international supply chain operations in Asia Pacific and Europe.
Learn the processes and operations behind each model
You can learn from my experience through this blog and my YouTube channel, where I share real-life operational projects with the data science community.
For analytics professionals who want to learn operational skills, I created a playlist of 5-minute videosin which I introduce 25+ analytics and data science concepts with:
An introduction to a real-world operational and business use case
A statement of the problem using operational terms
An approach to deploy the solution ensuring operational impact
The source code and the anonymised dataset
These short tutorials are meant to help you improve your operational and business culture while strengthening the skills needed to avoid being replaced by a stochastic parrot that generates code.
If you follow these guidelines, you can start 2027 confident you won't be replaced by an agentic workflow.
Let’s connect on LinkedIn and Twitter. I am a Supply Chain Engineer who uses data analytics to improve logistics operations and reduce costs.
For consulting or advice on analytics and sustainable supply chain transformation, feel free to contact me via Logigreen.
If you are interested in Data Analytics and Supply Chain, look at my personal blog and YouTube Channel for more tutorials.
Learn how to build your first PyTorch model, by using the "magical" Linear layer!
December 12, 20238 min read
Open-Source Tools for Industrial Automation
July 27, 202114 min read
A full guide on how to prepare for a bootcamp, successfully complete the course, and act afterward
September 2, 202314 min read
Data Foundations: From Architecture Design To Machine Learning
Four tell-tale signs that you should look for another job
December 17, 20246 min read
Deep dive into multiple linear regression with examples in Python
April 25, 202314 min read
SQL, Python, and other technical skills are critical, but they are only half the battle
April 20, 20239 min read
Hello there. Today I will present a guided tutorial for applying Kemp & Tenembaum's brilliant "form discovery" algorithm to a wine dataset...
January 27, 202117 min read
How to train a neural network to predict precipitation from satellite images.
November 19, 20219 min read
原文为英文,正文由自动抓取并保留原文。
原文发布于 Towards Data Science(How to Avoid AI Replacement as an Analytics Professional in 2027),标题、摘要与正文由机器翻译,术语按行业惯例统一,未经人工逐句校订。