Turn Client Chats into Mini-Tools Without Writing Code

Bottom line: If you're an ordinary person wanting to get the first slice of the AI pie, don't start by training models. Start with a small workflow. I work on AI solutions at Huawei Cloud, and when demoing for clients, I've found that salespeople are most willing to pay not for "how smart the large model is," but for "whether customer issues can be automatically categorized." I ran through some sales chat logs, organized "what customers ask, what they fear, and how to respond" into tags and scripts, which made money faster than learning algorithms. Can this capability be packaged as a product? Will users have a strong willingness to pay? It depends on whether salespeople are willing to pay for saving two hours a week.

Two Options, I Ran Them Before Choosing

Option A: Local Large Model. A large model is a computer program that reads and generates text; some smaller models can run on regular PCs or domestic computing power. I used this for about a month. Pros: Data doesn't leave locally, making it easier to explain sensitive client info. Cons are hard: Environment setup, VRAM requirements, and compatibility with domestic computing power—newbies often freeze up right after turning on the PC.

Option B: Local File Rules + Cloud or Low-Code Workflow. Low-code means writing less code, relying on drag-and-drop and templates. I used WorkBuddy for file organization; it sorts scanned documents and chat exports into different folders based on rules. I used this for 4 weeks, and it definitely saved manual sorting time. Then I let the model handle only the "pre-sales inquiry" pile. I ran 20 WeChat customer records: Option A took about 8 minutes to generate tags, 17 were directly usable, 3 misclassified "competitor comparison" as "price objection"; Option B took about 40 minutes, from raw records to a FAQ draft ready for clients (FAQ = Frequently Asked Questions list), tags were more stable, but data masking had to be done well.

Someone who has done supply chain for ten years doesn't necessarily need to compete in algorithms from scratch.

So the tutorial goes with Option B. The reason is simple: Clients don't care which model you use, they only care if you can categorize issues today so salespeople get scolded less tomorrow.

Step-by-Step: Running Through from 0 to 1

1. Pick a small problem first. Don't start with a "company intelligent assistant." Just do one thing: Categorize customer chat records into 5 types. For example: Price objection, Feature mismatch, Competitor comparison, Trial application, After-sales fault. Tags are just classifying issues for later statistics.

2. Prepare 20 to 50 real chats. Export as txt or csv. Csv is a table file, each row a record. Mask data—in plain terms, delete info that identifies people: names, phone numbers, WeChat IDs, company names. Create a 01_Raw_Records folder on your PC and drag files in. Expectation: Folder contains only text, no phone numbers.

3. Create three folders: 01_Raw_Records, 02_Rule_Templates, 03_Output. In 02_Rule_Templates, create Tag_Table.csv with columns: Customer Question, Tag, What Customer Fears, Standard Reply. This step is for humans to read, not for the model.

4. Write judgment criteria. One line per tag, don't exceed 50 characters. Example: "Price objection: Customer mentions expensive, budget, discount, think again, and hasn't said specific features are lacking." The model needs clear boundaries, otherwise it might classify "I think it's too expensive" as feature mismatch.

5. Start running. In WorkBuddy or similar tools, click New Project, select 01_Raw_Records, then choose output to 03_Output. Paste the prompt (instructions for the model) into the input box:

Judge only based on text, do not fabricate. If unable to judge, output 'Pending Manual Confirmation'. Output per line: Original summary, Tag, What Customer Fears, Standard Reply Draft.

If the interface supports template selection, choose Tag_Table.csv. See the progress bar, wait for generation. Expectation: A table appears in 03_Output, each row has a tag column.

6. Manual review. Focus on "Pending Manual Confirmation" and conflicting tags. I had 3 out of 20 needing changes. When editing, only change tags and replies, don't touch original text. Expectation: A clean table.

7. Package as a product. Convert the table to FAQs, categorized by industry, e.g., "Manufacturing Sales Objection Pack," "E-commerce CS Return/Exchange Pack." When demoing to clients, don't talk about models, talk about results: Import chats today, salespeople ask colleagues fewer questions tomorrow. Can this capability be packaged as a product? Do users have strong willingness to pay? I usually ask: How many hours a week does your team currently spend on this classification?

Pitfalls and Next Steps

The easiest mistake is dirty data. Chat records often have hundreds of words in one paragraph, causing the model to mix multiple issues. Solution: Keep paragraphs under 10 sentences; split if longer. Second pitfall: Incomplete data masking. Don't find it troublesome; use Find & Replace to swap phone/WeChat IDs with "Masked" first. Third pitfall: Too many tags. Newbies starting with 20 tags can't even judge them themselves, making the model messier. Start with 5, add more once it works.

Another realistic pitfall. I recently started using Subpool less than a week ago; it's a self-hosted ledger tool (self-hosted means software runs on your own PC/server), suitable for recording costs per AI call. I've used domestic computing power and Ascend for 3 weeks; Ascend is Huawei's proprietary AI computing platform, compatibility needs separate testing. If clients ask if data can be privatized, don't overpromise; run a small sample first.

After learning this, next steps include applying the same workflow to customer complaints, e-commerce CS, and real estate agent inquiries. Make 10 tags first, run 50 records, then discuss payment. Looking ahead, small AI products aren't competing on who has the bigger model, but on who organizes business experience into deliverable templates.