Inside an AI That Thinks Before It Sells
Most chatbots follow scripts. Ours reasons in Bangla, handles objections like a seasoned closer, and turned 373 Facebook Messenger conversations into a 9.7% close rate among engaged customers.
We gave our AI agent a name (Himu), a personality (direct, warm, zero desperation), and a sales playbook. Then we let 373 real customers talk to it on Facebook Messenger. This is what happened — starting with one conversation that shows everything.
How an AI Chatbot Sells Without Discounting
Himu doesn't follow a flowchart. Himu follows a playbook — the same way a trained salesperson does.
Here are the four tactics that showed up across 373 conversations:
1. Per-Unit Reframing: The Alex Hormozi Pricing Tactic
When a customer says "কম করেন" (reduce the price), Himu doesn't check with a manager or offer 10% off. It reframes:
"দিনে ৫-৭ টাকা পড়ছে, এক কাপ চায়ের দামও না।"
Per-unit reframing. The product isn't ৳350. It's ৳7 a day. Less than tea. This is the same tactic Alex Hormozi describes in his sales frameworks — make the price feel irrelevant by reframing the unit of measurement.
The agent checks checkPromotions first. If a real discount exists, it offers it. If not, it reframes instead of fabricating one. Zero panic selling across 373 conversations.
2. Assumptive Close: Don't Ask, Collect
Himu never says "আপনি কি অর্ডার করতে চান?" (Do you want to order?). When the customer signals intent — giving their name, asking about delivery to their area, requesting payment options — Himu starts collecting order fields.
The customer's action IS the answer. Asking "do you want to order?" after someone just gave you their address is a trust-breaker. It signals uncertainty. Himu skips it.
3. Why the Best AI Sales Agents Don't Chase Customers
When a customer hesitates, goes silent, or says "পরে নেব" (I'll take it later), Himu doesn't chase:
"ঠিক আছে, ভেবে দেখুন। দরকার হলে বলবেন।"
Translation: "No problem, think about it. Let me know if you need anything." Then silence. No follow-up. No "limited time offer!" No desperation.
This is counterintuitive for a sales agent. But in Bangladeshi commerce culture, desperation kills trust. A relaxed seller is a confident seller. Himu is relaxed.
4. How AI Reasoning Traces Improve Sales Conversations
This is what makes Himu fundamentally different from rule-based chatbots.
Before every response, the agent produces a thinking trace — in Bangla. It reasons about:
- What the customer actually wants (not just what they typed)
- Which tool to call (or not call)
- What sales stage they're in
- How to phrase the response
98.9% of these thinking traces are in Bangla. The agent doesn't just output Bangla — it reasons in Bangla. When a customer types "থানা কা ডে ক্রিম" (a phonetic misspelling), the thinking trace shows: "এটা 'থানাকা ফেস ক্রিম' এর বিকৃত সংস্করণ" — it recognizes the misspelling and maps it to the correct product before responding.
This isn't pattern matching. It's comprehension.
AI Chatbot Conversion Rate: Real Data From 373 Facebook Messenger Conversations
After the conversation, the data. Here's what happens when you put an AI salesperson on a real Facebook page with real traffic.
Facebook Messenger Sales Funnel: Bounce Rate to Close Rate
39% of visitors bounce after a single message. They ask "দাম কত?", get prices in under 10 seconds, and leave. They're window shopping. The agent's job for them: be fast, be accurate, leave a good impression.
9.7% of engaged conversations close. When someone stays for 2+ messages — asks follow-up questions, negotiates, provides an address — nearly 1 in 10 buys. The agent's job for them: sell.
What Do Customers Ask a Facebook Page Chatbot?
First Message | % |
Price query ("দাম কত?", "price", "praic koto") | 65% |
Product query ("থানাকা", "cream", "ফেস পেক") | 21% |
Facebook auto-icebreaker ("Can you check the price?") | 6% |
Greeting ("hi", "হলো") | 4% |
Image (product photo, ad screenshot) | 4% |
65% open with "দাম কত?" and nothing else. Price-first traffic. The agent gives prices immediately — no gatekeeping, no "আপনি কোন প্রোডাক্ট চান?" (which product do you want?). That single behavior — answering the question that was actually asked — is what keeps people in the conversation long enough to buy.
24/7 AI Customer Service: Response Time and Availability
- Response time: <10 seconds, every message, including 3 AM
- Zero missed messages across 373 conversations
- 24/7 availability — no shifts, no breaks, no sick days
- 2 human handoffs in the entire deployment (0.5%) — both for complex complaints, not sales
Can an AI Chatbot Handle Objections in Local Languages?
No self-respecting case study hides the failures. Here's what we're still fixing:
1 markdown violation in 1,100+ messages. Himu used **bold** formatting once — triggered by a triple Facebook icebreaker question. Messenger doesn't render markdown. It showed literal asterisks. Ugly.
1 Devanagari numeral slip. Hindi numerals (२, ३) instead of Bengali (২, ৩) in one message. The scripts look similar but are linguistically wrong. Like writing "3" with a Cyrillic character.
Product data gaps. When রিমা asked "কত মিলি" (how many ml), the inventory had no size data. The agent admitted it honestly instead of guessing — but a complete inventory would have closed faster.
Can't send images. Multiple customers asked to see product photos. The agent redirects to the Facebook page — functional but not ideal.
Every mistake is traceable to a specific trigger. The fixes are prompt changes, not engineering overhauls. Compare that to human staff where quality issues are invisible until a customer complains.
AI Sales Agent vs Human Sales Rep: Side-by-Side Comparison
Himu (AI Agent) | Human Messenger Staff | |
Response time | <10 seconds | 3-12 minutes (when available) |
Availability | 24/7/365 | 8-10 hours/day |
Close rate (engaged) | 9.7% | Unknown (no tracking) |
Price accuracy | 100% (from database) | Variable (memory-dependent) |
Objection handling | Consistent playbook | Depends on training + mood |
Language register | 100% formal আপনি | Varies |
Mistakes per 100 messages | ~1 | Unknown |
Missed messages | 0 | Common during off-hours |
Learns from every conversation | Yes (thinking traces logged) | No systematic feedback loop |
The question isn't whether AI can sell. The question is whether you can afford not to use it when 39% of your Messenger traffic arrives while your staff is asleep.
How to Automate Facebook Messenger Sales Without Losing the Personal Touch
Based on 373 conversations of production data, three priorities:
- Product image sending. রিমা asked to see the Face Cream. The agent couldn't show her. Facebook's Messenger API supports image attachments — the agent should pull photos from inventory and send them inline.
- Re-engagement for bounced conversations. 39% of customers send one message and leave. A well-timed follow-up — "গতকাল উপটান নিয়ে জানতে চেয়েছিলেন — কোনো প্রশ্ন থাকলে বলুন" — could recover a portion of these.
- Multi-vendor expansion. The playbook, personality engine, and tool architecture are vendor-agnostic. The next deployment targets a different product vertical to validate that the system — not just the product — drives results.
This is the second in a series of operational case studies from Karigor AI Labs' Facebook Commerce agent deployment. The covers the initial 24-hour results.
Sources:
- First 24-Hour Case Study
- Gorgias: 2026 State of Conversational Commerce Report
- Envive: 50 E-Commerce Conversion Rate Statistics for 2026
- Ringly: Ecommerce Customer Service — Hiring vs AI in 2026
- Internal production data: 373 conversations analyzed via deepInspect and conversionFunnel queries, March 29, 2026
