8% to 23% — I watched this number happen
Bottom line first: 4 people, 3 months, a handful of AI tools, and inquiry conversion went from 8% to 23%.
Sounds impressive? It's honestly not that magical.
The company does building materials export. 4-person team, roughly $2M in annual sales. The pain point was simple: lots of inquiries, slow replies, lost deals. Alibaba International throws 10-15 inquiries daily, and 4 people can't keep up. Sometimes a client sends an RFQ and waits two days for a reply — by then, they've already found someone else.
Here's the brutal truth in foreign trade: reply speed matters more than content quality. No matter how good your words are, two days late means they're worthless.
What did they use? Nothing exotic
The tool stack is modest:
- ChatGPT — outbound emails, inquiry replies, English polishing
- Claude — complex inquiry analysis, breaking down client needs
- Midjourney — product image rendering, replacing some real photography
- HeyGen — multilingual product intro videos
Nothing fancy. All commercially available. The key is how they combined them.
Outbound emails: ChatGPT + Claude doubles team
Before, 4 people wrote maybe 20 outbound emails daily. Styles varied wildly — some too formal, some too casual.
After the process change:
- ChatGPT generates drafts in batches, organized by industry/country/product type
- Claude does second-pass polishing, adjusts tone, injects industry terminology
- Human review at the end, adds personalized touches
Output: 60-80 emails daily, with more consistent quality than before.
One trick: never let AI write from scratch to finish. The more specific the input material, the less the output feels AI-generated. They fed their actual product specs, the client's pain points, even feedback from previous deals — things AI can't fabricate.
Inquiry replies: speed is everything
Old workflow: sales rep translates → thinks about reply → writes English → colleague reviews → sends. Average time: 48 hours.
New workflow:
- AI translates inquiry content (30 seconds)
- Sales rep writes core reply points in Chinese (5 minutes)
- Claude translates to professional English + polishes (1 minute)
- Human checks key specs, then sends (2 minutes)
Average time: 2 hours.
Don't underestimate this speed gap. In foreign trade, replying within 24 hours triples the closing probability compared to 48 hours.
Product images — saving money beats saving time
Building materials export demands product images. Previously they spent about $8,000/year on photography + post-production, enough for one product catalog lasting half a year.
New process:
- Core products (5-10 SKUs) still get real photography
- Other SKUs rendered by Midjourney, referencing the real shots for style and lighting
- Complex application scenes (e.g., products in architectural contexts) fully AI-generated
Over six months, photography costs dropped from $8,000 to $2,000. Saved $6,000. No client ever complained about image quality.
The trick: AI images aren't generated from nothing. You feed a good real photo to Midjourney as a style reference, and what it produces is nearly indistinguishable from the real thing.
The video move — they got this right
The last piece was HeyGen for multilingual product intro videos.
They'd previously spent $3,000 on a single English product video. It worked fine but only in English. German, French, Japanese clients kept asking "Do you have a version in our language?" — nope.
Now with HeyGen:
- One English script, auto-generated into German, French, Japanese, Spanish versions
- Each video: about 1 minute
- Cost: roughly $50/month (HeyGen subscription)
Results? Video click-through in inquiries went from 12% to 28%. German and Spanish versions especially boosted trust from European and Latin American clients.
3 big mistakes they made — you need to know these
It wasn't smooth sailing. Several missteps:
Mistake 1: AI emails were too "perfect" — clients distrusted them
At first, ChatGPT wrote everything. Perfect grammar, perfect logic, perfect phrasing — reply rates actually dropped 15%.
Why? Too perfect means not real. A building materials buyer receiving a flawlessly written email instantly thinks "this is a template, not for me."
Solution: AI drafts, human adds "imperfections" — a casual opening line, a specific question about the client's project. Imperfection equals authenticity.
Mistake 2: Product image consistency broke once
During a rush project, 3 SKUs got AI renders in different styles. Client received the quote and asked: "Your product images look inconsistent — are you the same supplier?"
Awkward.
Solution: Unified Midjourney style parameter template (style reference + color palette), shared across all SKUs.
Mistake 3: AI video translation glitched
HeyGen's Japanese translation had one obvious error — it translated "pressure rating" as "emotional resilience rating." Industry terminology is not AI's strong suit, especially in Japanese.
Solution: All AI-generated video scripts must go through manual terminology review. Better to be 10 minutes slower than lose professional credibility.
The ledger: what did they save, what did they earn?
3-month data:
- Inquiry reply time: 48 hours → 2 hours (96% faster)
- Outbound email volume: 20/day → 60-80/day (3-4x)
- Inquiry conversion rate: 8% → 23%
- Photography costs: $8,000/6mo → $2,000/6mo
- Video production: $3,000/project → $50/month
- Total AI tool costs: ~$150/month (ChatGPT Plus $20 + Claude Pro $20 + Midjourney $10 + HeyGen $50 + translation API ~$50)
Saved $6,000 photography + $2,950 video, spent $450 on tools. Net savings: $8,500.
But the bigger story is conversion rate gains — at their average deal size of $5,000, roughly 15 additional orders in 3 months, adding $75,000 in revenue.
3 takeaways you can actually use
1. Add flaws to AI-written content — too perfect feels fake. Add a conversational touch, a specific detail, something that signals "a real person wrote this."
2. Speed beats quality, but don't chase speed alone — replying within 24 hours is the baseline, but key specs and terminology must be human-checked.
3. Anchor all AI images to one real reference — feed one great real photo as a style reference, then lock all AI outputs to it. Prevents style chaos.
Look, what these 4 people did wasn't "AI replacing humans" — it was "AI giving each person leverage." One sales rep used to handle 5 clients; now they handle 15. Reply speed, coverage, language capability — all amplified.
Clients never noticed AI was helping — which means they did it right.