Expensive AI courses still fail to address the 'unknown unknowns' problem

I just saw that course on 2026 workplace AI automation workflows for 6,900 yuan. It's a bundle of two classes, teaching both managers and individuals, with the slogan "Build Your First Dedicated Bot." The logic makes sense, but the price reminded me of what I did last week. I spent zero yuan and welded together an automated workflow for myself in WorkBuddy.

My situation is this: I'm a freelance translator. Clients are scattered, and for every job, I have to manually confirm the original text format, glossary, delivery time, and payment method with the other party. This kind of info is scattered across emails, WeChat, and Excel, and every client has different habits.

Old clients are annoying, new clients are annoying too. Two weeks ago, I was organizing my order-taking process and thought, can I semi-automate this part using WorkBuddy?

Yes. But it's not like the "throw it in, spit it out" you might imagine.

Admitting First: Version 1 Was a Failure

I initially tried to create a universal workflow for "all clients." I recorded a long natural language description, roughly meaning "Read various files I receive, extract project info, generate a confirmation email." As a result, WorkBuddy recognized things haltingly, and the email templates were hit-or-miss.

Later, I changed my approach: Serve one real scenario at a time. I chose the client with the most complex contracts who most frequently chases me asking "When will you reply?" Their requirement sheets are never a unified template—sometimes PDF, sometimes scanned images, sometimes just three lines written directly in the email body.

I created a new Skill in WorkBuddy. The name sounds fancy, but it's essentially a configured workflow. You give it some inputs, it executes fixed actions, and produces fixed results. I set the inputs to three fields: Project Number, Client Name, Original File. The output was set to a standardized requirement sheet.

Here is exactly how I did it:

Open WorkBuddy → Click "New Skill" in the left workspace → Enter name "Client A Order Pre-processing" → Add the above three fields in "Input Configuration" → Write clearly in "Processing Logic": Read Original File, regardless of PDF or image, uniformly run OCR text recognition (it has a built-in iFlytek OCR connector that can be called directly) → Extract dates, language direction, word count, payment terms from the file → Output extracted results according to the table template

The easiest thing to mess up here is "letting the model decide what the template looks like." So I pinned down the column names and formats in advance in the Skill's output settings. Column names used the terminology from Client A's historical contracts, such as Delivery Date, Unit Price, Remarks. The model is only responsible for filling values, not redesigning the structure.

Permissions: This Is What I Want to Talk About Most

Tools like this default to open permissions. I got burned by this initially. When creating a Skill in WorkBuddy, you can set the execution scope—whoever owns it gets the rights. I set two layers:

• Who can view: Only me.
• Who can edit: Only me.

Translation agencies occasionally ask for progress updates. I send them a read-only summary link. They can see the progress but cannot change my configuration. This avoids a lot of unnecessary information tug-of-war.

As for auto-generating daily/weekly reports, I still gave company colleagues read-only access. They have an entry point for editing needs, but the entry point for actually running processes and touching logic is mine alone. The reason is that this stuff is highly personalized. If others randomly change your rules, you'll go crazy.

Daily Maintenance Is More Important Than You Think

A Skill isn't done once created. My experience is: Run through five fixed checkpoints monthly.

• First, check the last 20 run logs for instances where "fields were recognized as blank." If found, add them to the prompt, e.g., "If date is empty, default to calculating from the email receipt date."
• Second, check if connected email authorizations have expired. WorkBuddy requests long-term tokens when connecting external mailboxes; these tokens occasionally fail silently.
• Third, verify a benchmark test: Run 3 past real contracts through it and see if the results differ significantly from the manually processed tables.
• Fourth, cross-check hardcoded contact emails and payment info in the Skill against reality to prevent issues if the client changes personnel.
• Fifth, clean up temporary folders generated from repeated testing while you're at it.

This routine saves me about 20 to 30 minutes per job. Currently, all new inquiries and heavy jobs I receive are fed through this process first. Only those that fail get handled manually. The success rate is about 70%; the remaining 30% are particularly irregular scans that still need manual fallback.

As for that 6,900 yuan course, I won't say it's useless. After all, it forces you to clarify "what you want." But what truly blocks people is whether you're willing to spend an afternoon writing down your own rules clearly. Tools are magnifying glasses; they can only magnify the order you already have, not your chaos.

My current process isn't flashy yet, but at least next time a new client comes, I know the first step is clicking "New Skill" somewhere specific, rather than searching everywhere for templates.