Summary: A furniture trading company with RMB 30M annual revenue deployed a multilingual customer service Agent and email Agent on KaiheAiBox F1, covering English, Spanish, French, and Arabic for pre-sales inquiries while automating 48-hour email responses and follow-ups. This article walks through the selection logic, deployment configuration, and 90-day operational data — no fluff, just numbers.

Mr. Sun has been running a furniture trading business in Ningbo for thirteen years. His company has distributors in North America, Europe, and the Middle East. His team's daily routine was consumed by three things: replying to emails, responding to inquiries, and dealing with time zone differences.

"Our sales team of 8 people handles over 50 emails and 30+ online inquiries per person per day," Mr. Sun said. Before deploying the KaiheAiBox F1, he did a time audit: "Pure email and chat response time accounted for more than 70% of our working day."

This wasn't an efficiency problem. It was a structural problem: customers are spread across global time zones, but employees can only sleep in one.

Why Not SaaS Subscriptions?

Mr. Sun's company had tried three approaches before, each with its own issues.

Approach 1: SaaS customer service platforms (Zendesk, Freshdesk, etc.)

Hundreds of dollars per month, billed per seat. The problem was that email and live chat ran on two separate systems — email context and history from distributors was completely disconnected from online conversations. Plus, all data lived in the cloud, which always raised compliance and privacy concerns.

Approach 2: OpenAI API + custom integration

The company's technical co-founder spent two months writing a custom integration between OpenAI's API and their email + web chat system. It worked well initially, but the monthly API bill skyrocketed from $200 in week one to $800 by week four — more customers, more questions, more costs, with no cap.

Approach 3: So-called "AI trading tools" on the market

They tried several AI-powered foreign trade tools. The problem was universal: insufficient domain knowledge. Feed it a Spanish customer inquiry email, and it would translate to Chinese just fine. But the reply would be hollow — lacking understanding of furniture-specific product parameters, certification requirements, packaging standards, or other professional knowledge.

"These tools are more like translation plugins than industry-specific Agents," Mr. Sun concluded.

The KaiheAiBox F1 approach was fundamentally different — a one-time hardware investment, local operation of open-source LLMs, and zero data leaving the local drive. The three key selection criteria:

1. Private deployment: Customer data, pricing strategies, and supplier information all stay on the local server, never touching any third-party API
2. One-time cost, unlimited usage: No "per-token" pricing. A thousand Agent replies cost the same as one
3. Customizable Agents: Enterprise product knowledge bases and historical email corpora can be layered on top of local models, making the Agent truly sector-aware

Deployment: One KaiheAiBox F1 Running Three Agents

Mr. Sun chose the KaiheAiBox F1 (high-spec edition), deploying two open-source LLMs as the Agent engines:

• Qwen2.5-72B-Instruct: The primary model for the customer service Agent, handling real-time multilingual online conversations. Its multilingual benchmark performance on smaller languages approaches GPT-4 level, but runs entirely locally.
• Qwen2.5-32B-Instruct: The primary model for the email Agent, handling email composition, summarization, and classification. The 32B parameter count is sufficient for long-text formatting and tone control, with inference speed 40%+ faster than the 72B.

On top of these two models, three Agent instances were deployed:

Deployment Timeline

Mr. Sun's IT team (two people) worked remotely with KaiheAiBox support:

• Day 1: Hardware setup, power on, system initialization
• Days 2-3: Deploy Qwen2.5-72B and 32B models, configure Agent runtime environment
• Days 4-5: Import 2 years of historical emails (~8,000) and product data (SKU database, quote templates, certifications as PDFs), build RAG knowledge base
• Days 6-7: Configure email server integration and live chat tool integration
• Day 8: Launch online CS Agent in gray-scale mode (English inquiries only)
• Day 14: Full launch of email Agent
• Day 21: Full multilingual CS rollout (Spanish, French, Arabic)

"From hardware arrival to all three Agents fully online — exactly 21 days," said Mr. Sun. "The most time-consuming part wasn't the technical deployment. It was organizing the historical emails and product data — all scattered across individual sales reps' laptops. Building the knowledge base took a full week."

Three Months of Operational Data

As of the interview, the KaiheAiBox F1 had been running stably for 90 days. Here are the core metrics.

Multilingual CS Agent

The most striking metric was the conversion rate increase. Mr. Sun analyzed: "Middle Eastern customers prefer live chat for price inquiries, and their local working hours are Beijing afternoon to early morning. Before, these messages sat unanswered until the next day, by which time the customer had already found another supplier. After the 24/7 Agent rollout, our Middle East online inquiry conversion rate doubled."

Email Agent

The email Agent strategy wasn't auto-send — it was "generate draft + human review." But operational data showed 88% of drafts were ready to send. Reviewing a draft is far less work than writing from scratch. "Our salespeople now have at least two extra hours per day for proactive customer outreach instead of email grunt work," Mr. Sun calculated.

An Unexpected Metric: Complaint Reduction

Customer complaints dropped 62% in the first three months post-deployment. Investigation traced most pre-deployment complaints back to "information gaps" — emails gone unanswered, follow-ups ignored, or misunderstandings born from slow response. With 24/7 Agent replies, customers no longer experienced that "can't reach anyone" feeling.

Three Hard-Earned Lessons

Mr. Sun's team was willing to share their mistakes so others can avoid them.

Lesson 1: Don't use 72B for emails — it writes too much

During initial testing, they connected the 72B model directly to email. The Agent replied to every message like a business report — structurally perfect but way too long. "A client just wants a price and delivery time. Writing three pages confuses them." They switched to the 32B model with a concise prompt, capping replies at 200 words. Customer experience improved.

Lesson 2: Arabic text direction needs preprocessing

Arabic is right-to-left, and Agent outputs occasionally produced formatting issues when mixing English and Arabic in a single email. The fix was adding a text direction preprocessing module in the Agent pipeline, forcing RTL formatting for all Arabic output.

Lesson 3: The "cold start" problem for live chat

When a new customer started their first conversation, the Agent had no historical context and could only guess needs from the knowledge base. The opening was often awkward. The solution: a "first-conversation template." When the Agent detects insufficient user-provided info after three rounds, it proactively asks about order volume, destination port, material preferences — turning "guess the need" into "fill the form."

Advice for Trading Companies

Drawing from Mr. Sun's experience, if you're considering a private multilingual Agent for your trading business:

1. Fix your knowledge base first, then tune the model

Mr. Sun's team spent a week organizing their knowledge base — the longest phase of the entire project. If your product data, historical emails, and quote templates are scattered, even the best Agent will be a "smart stranger who doesn't know you."

2. Start email Agent in "draft mode"

Letting the Agent send emails directly is too risky — one wrong tone can lose a long-term client. Run a "Agent draft → human review" mode for one to two weeks to build trust data. Consider auto-send only when the revision rate drops below 15%.

3. Don't skimp on model size for smaller languages

If your customer base includes non-English markets, don't choose a smaller model to save money. 7B/8B models suffer from severe "translation-ese" on smaller languages — grammatically correct but stiff in tone. Customers can immediately tell they're talking to a bot. 72B is the baseline.

4. Don't roll out all languages at once

Mr. Sun's company started with English CS gray-scale testing for two weeks. Only after confirming stability did they open Spanish, French, and Arabic one by one. "For every new language, we watched the first three days like hawks — smaller language issues are only detectable by native speakers."

On Cost

The KaiheAiBox F1 is a one-time hardware investment. No monthly per-seat fees. No per-API-call charges. In 90 days, the three Agents handled over 57,000 conversations and 34,000 emails — with zero incremental cost. Compare that to the SaaS and API model: not "the more you use, the more you pay," but "the more you use, the better the value."

This isn't a story about how cool AI is. It's a story about how AI helps a traditional trading company make more money and worry less. When your customers across five time zones can reach your "virtual sales team" late at night — the orders start coming naturally.

KaiAIBox | Agentaibox that lets AI work for you 24/7 · User Case