Editor's note: This post is adapted from a talk I prepared for the 22nd session of "AI Afternoon Tea," an event series by the Henan Artificial Intelligence Association. To make it suitable for public reading, I've cut the stage directions and live-demo segments, and tightened up project-status descriptions and efficiency numbers from the original script that could easily be misread. The cases in this piece reflect the practice and thinking I had at the time — they don't mean every workflow described has been running long-term or at scale.
I'm Zhang Yukui (张玉魁), a clinical medicine major, not someone with a technical background.
I can't write code without AI's help, and I don't understand complex database engineering. But through repeatedly using AI to handle content, information, and business needs, I've gradually realized: what really determines whether AI implementation works usually isn't how smart the model is, but whether it's plugged into a concrete workflow.
There are plenty of AI systems that can chat. Far fewer can actually take something from input to output.
That's my starting point for understanding a "digital employee."
Why AI implementation often doesn't work
Many companies are already using AI: writing copy, looking things up, summarizing, generating slide decks. But if you push a little further and ask — which specific step of repetitive work did it actually cut out? — a lot of people struggle to answer.
The problem usually comes down to three things:
- AI only gives answers, without connecting to any follow-up action;
- data, spreadsheets, messages, and business systems are scattered across different places, so someone has to manually move things between them;
- AI doesn't automatically know when to start working, or who the results should go to.
So a digital employee isn't a chat box that feels more human — it's a workflow that can be triggered, that calls tools, processes information by rules, and delivers the result to a specific destination.
This doesn't mean AI can fully replace employees. A more accurate way to put it: it's suited to taking over a portion of the repetitive, clearly-defined execution work that's easy to check.
Scenario one: cutting down manual data handoffs
There's a common problem in companies: data is scattered across ad platforms, business spreadsheets, and different systems. When the boss wants to see a result, someone usually has to export it, pull it together, clean it up, and then explain it.
In the original script, I called this an "AI data analyst." Looking back now, that label needs to be more careful.
What I can actually do isn't build a complex data platform or complete database development on my own — it's help sort out this chain:
- what the boss ultimately needs to see;
- where each piece of data comes from;
- which steps can run on a schedule;
- which anomalies need an alert;
- which conclusions must be confirmed by a person.
AI can help generate summary logic, analysis frameworks, and draft reports, and it can work alongside existing tools to cut down on moving spreadsheets, copying, and repetitive tidying. But metric definitions, permissions, security, and the final business judgment still need to be handled by qualified people.
The most important thing here isn't "can AI analyze this" — it's first getting clear on the inputs, outputs, and where responsibility lies.
Scenario two: turning content production from a blank page into an editable draft
I've run a WeChat Official Account called "爱玩 AI 的医学生" (A Med Student Who Loves Playing With AI). Content production has a lot of repetitive steps: gathering material, structuring, drafting, refining the wording, formatting.
AI can clearly cut the cost of these execution tasks, but only if a person still owns the viewpoint and judgment.
I've gradually settled into a workflow that suits me better:
- use articles I've written before as a style reference;
- spell out the audience, purpose, viewpoint, and phrasing that shouldn't appear, up front;
- let AI handle organizing material and producing a first draft;
- check every case, conclusion, and fact one by one;
- keep revising with specific feedback, instead of just saying "this is bad, rewrite it."
AI doesn't automatically pick up your style. What feels "like something I wrote" actually comes from continuously feeding it samples, standards, and feedback.
So this role is more like a content assistant, not an author speaking on my behalf. It handles preparing material and drafts; I decide what's worth saying, whether the facts hold up, and whether it ultimately gets published.
Scenario three: breaking a content matrix into an observable workflow
I also tried a Xiaohongshu (RedNote) content-matrix project, hoping to link together content collection, topic analysis, copy generation, Feishu (Lark) review, and publishing.
This project never became the fully stable, "fully automated closed loop" the early script made it sound like. It got partial chains and features validated, but plenty of problems still need work — content quality, account security, platform rules, error recovery, and long-term maintenance.
What this experience really taught me wasn't "one person can automate running lots of accounts" — it was how to break a complex piece of work apart:
- what information to collect;
- what criteria to filter by;
- how different accounts keep distinct voices;
- which step must go through human review;
- how to log and handle a failed publish;
- which platform risks shouldn't be left to automation to gamble with.
AI can help generate code and iterate on it, but what I'm responsible for is describing requirements, breaking down the workflow, testing, giving feedback, and making trade-offs. Without AI, I couldn't build an equivalent system on my own; having AI doesn't mean the system is inherently reliable either.
The limits of a digital employee
The "digital employee" metaphor is appealing, but that's no reason to ignore its limits.
AI is suited to tasks that are repetitive, clearly ruled, high-volume, and checkable. When it comes to face-to-face communication, building trust, final decisions, medical judgment, complex business judgment, and high-risk operations, a person has to be there.
To judge whether a task is suitable for AI, I now start by asking:
- Does this happen repeatedly?
- Are the inputs and outputs clear enough?
- Can a person quickly check the result?
- Is the cost of a mistake controllable?
- Who's responsible for the final outcome?
If these questions don't have clear answers, it's not time to rush into automating.
What's actually scarce is finding the right scenario
Tools update fast — a product that's popular today might get replaced tomorrow. But "finding a work node worth reshaping" is a much more durable skill.
It doesn't require everyone to know how to code, but it does require you to seriously observe your own work:
Which task keeps eating up your time, and shouldn't really require a person to do it by hand forever?
Find it, then build a rough but testable version first. Once it's running, keep improving it based on errors and feedback.
To me, AI implementation isn't some grand slogan about "digital transformation" — it's turning one specific pain point into a workflow that's clearer, less effortful, and more controllable.