AI Sales Playbook Rewrite: How Our Bot Learned to Close Like a Human

I built OT1-Pro to answer customer messages, and for months I told myself the AI was doing fine. It replied fast and collected names and phone numbers. Then I read 200 real conversations back to back, and my stomach dropped. Our bot was not selling. It was catching information and dropping deals.

The worst one still burns in my memory. A customer wrote in Egyptian Arabic: "انا مهتم بالخدمة" — I am interested in the service. Our AI replied: "ما هي الخدمة اللي بتقدمها؟" — what service do you offer? He never replied. That failure forced me to throw out our old prompt and rewrite it as a human closer playbook, now living in one shared trait, BuildsConversationPrompts, used by NaraRouter, Gemini, and Ollama.

1. My bot asked a buyer what service THEY offer

The message was "انا مهتم بالخدمة بتاعتكم" — I am interested in your service. The word بتاعتكم literally means yours. No human would misread it. Our old AI treated the buyer like a vendor.

The same pattern showed up in English. Customer: "I want it" / "tell me more about your product." Old AI: "Could you tell me what you are looking for exactly?" In DMs, patience runs out after two messages.

BAD: Customer: "انا مهتم بالخدمة." AI: "اهلا! ممكن تعرفني بالخدمة اللي بتقدمها؟" The customer thinks nobody is home and leaves.

GOOD: Customer: "انا مهتم بالخدمة." AI: "اهلا وسهلا بيك! الخدمة بتاعتنا بترد على عملاءك على مدار الساعة حتى وانت نايم. تحب تعرف السعر ولا ازاي بتشتغل الأول؟" One benefit, one question, momentum kept.

I wrote RULE #0 first because nothing else matters if the AI misreads who is buying from whom. If you run Meta messaging, this misread costs even more because of how permissions and review work — I documented that path in my Meta app verification founder guide.

2. Rule #0: you work for the business, the writer is the buyer

RULE #0 sits at the top of BuildsConversationPrompts: you work FOR this business. Every person who messages you is a potential customer. They are not a vendor and they are not pitching you.

The trait lists trigger phrases in both languages: "I am interested" / "I want this" / "your product" / "tell me more" / "انا مهتم" / "عايزها" / "الخدمة" / "المنتج" / "بتاعتكم" / "عندكم". When any of those appear, the instruction is to pitch ONE benefit. Asking "what service do you offer?" / "ما هي الخدمة اللي بتقدمها؟" is explicitly banned. I made it a ban, not a suggestion, because polite clarification still kills the sale.

Only three exceptions exist: (a) after we have pitched our product at least twice, (b) when asking about their business tailors the pitch, for example "what do you sell?" to match a feature to their use case, or (c) when they explicitly ask us to understand their needs first. Outside those three, asking about their business is a bug.

3. The 10-step playbook that replaced the info-catching bot

Our old prompt told the AI WHAT to do — push toward the sale, handle objections — without teaching HOW. So the model fell back on collecting names and dumping features. It sounded like a form with emojis.

I replaced it with the flow a top 1% rep actually works in DMs:

1. Listen first. Read the full history. Place the buyer: curious, comparing, ready, hesitating, objecting.
2. Greeting is not a pitch moment. Warm hello plus one hint at outcomes, then a question.
3. Discover the pain with one short qualifying question.
4. Pitch benefits, never features. Feature → so you can → outcome.
5. Handle objections with Feel-Felt-Found or Isolate-Reframe-Resolve. Never argue, never discount, never apologize.
6. Collect micro-commits. Small yeses before the big yes.
7. Close directly on a buying signal. Direct, assumed, or alternative close.
8. Handle silence and short answers by pivoting, not apologizing.
9. Match the customer's energy, dialect, and length.
10. One idea per reply. One question or one call to action, DM-short.

Every provider inherits the same steps through the shared trait. NaraRouter may route to a different model than Gemini or our local Ollama fallback, but the sales mechanics never change. I wanted one closer, not three different salesmen. If you wonder why most AI replies feel robotic, this is usually why — politeness with no playbook. I wrote more on what actually turns replies into revenue in my OT1-Pro vs Wati breakdown.

4. Listen first, and never confuse a greeting with a pitch

Steps one and two sound obvious until you watch a bot fail them hundreds of times. Step one: decide where the buyer is in his head. Curious needs education. Comparing needs differentiation. Ready needs a close. Hesitating needs safety. Objecting needs reframing.

Step two fixes the greeting failure. Customer writes "السلام عليكم" or "hi" with no question. The old AI either interrogated or dumped pricing plus features.

BAD: Customer: "السلام عليكم." AI: "وعليكم السلام! اسمك ايه؟ بتشتغل في ايه؟ وعايز تعرف سعر انهي باقة؟" Three questions, zero value.

GOOD: Customer: "السلام عليكم." AI: "وعليكم السلام واهلا بيك! تحب أعرفك ازاي بنرد على عملاءك وانت نايم، ولا عندك سؤال محدد؟" Warm, one hint at the strongest outcome, one open door.

Same shape in English. BAD: "Hi! What is your name and what are you looking for? Here are our plans..." GOOD: "Hey! Want the quick tour of how we reply to your customers 24/7, or do you have something specific in mind?" The greeting earns the second message. It does not spend it.

5. Discover the pain fast, then sell benefits not features

People buy when the pain of staying the same exceeds the pain of changing. The AI asks one short qualifying question when relevant: "إيه الحاجة اللي بتدور عليها بالظبط؟" / "What is the main thing you are trying to solve?" One question, never an interrogation.

Then it translates features into outcomes. Nobody buys a unified inbox. They buy never missing a message again. Nobody buys 12,000 AI credits. They buy a business that replies while they sleep. Formula: feature → so you can → outcome, one benefit per message.

BAD (feature dump): "عندنا inbox موحد و 12,000 credits و multi-channel و analytics." The customer reads specs and feels nothing.

GOOD (benefit matched): "بما إنك بتضيع رسائل على انستجرام وفيسبوك، الميزة دي بتجمعهم في مكان واحد — so you can reply in seconds and stop losing buyers who message at midnight." One pain, one outcome.

Lead-score emphasis layers on top: cold leads stay on discovery, warm leads get one matched benefit, hot leads skip to the close. A cold "hi" is never rushed and a hot "how do I pay?" is never lectured.

6. Objection handling: Feel-Felt-Found with real Arabic replies

Objections are where deals are won. My rule is blunt: do not argue, do not discount, do not apologize. Use Feel-Felt-Found (I understand how you feel, others felt the same, here is what they found) or Isolate-Reframe-Resolve (isolate the real blocker, reframe it, resolve with a small next step).

These four cover almost every stall in our inbox: too expensive, need to think, check later, and does it really work.

In English: too expensive BAD: "Sorry it feels pricey, what can you afford?" GOOD: "Totally get it — others felt the same until they compared it to paying someone to watch the inbox. Want to try free first?" Need to think BAD: "No problem, think about it!" GOOD: "Of course — what part do you want to think through?" Check later BAD: "Sure, ping me later!" GOOD: "Want me to send the link now so you have it when free?" Does it work BAD: "Yes, it is amazing!" GOOD: "Fair question — want an example from a business like yours, or try free and watch it reply live?"

7. Micro-commits before the close, then close like you mean it

Nobody says yes to a big ask without saying yes to small ones first. Step six forces micro-commits: "يعني الموضوع ده مهم ليك دلوقتي، صح؟" / "Does solving X sound worth 10 minutes to set up?" Each yes makes the next ask easier.

Step seven: close. Buying signals are explicit in the trait — asking about price, asking HOW it works, asking what is included, asking about guarantees, saying "okay" or "تمام". When you see one, stop selling and close.

Three close shapes ship in the prompt. Direct: "تمام — تحب نبدأ دلوقتي؟ اللينك: https://ot1-pro.com/register" Assumed: "هبعتلك اللينك الآن، لما تخلص التسجيل قولّي وأنا أمشي معاك خطوة خطوة." Alternative: "تحب تبدأ بالخطة المجانية الأول ولا بالمدفوعة؟"

BAD closing: Customer: "بكام الباقة؟" AI: "عندنا 3 باقات بأسعار مختلفة ومميزات كتير تحب تعرف إيه؟" The buyer asked how to pay and got homework.

GOOD closing: Customer: "بكام الباقة؟" AI: "المدفوعة 79 دولار في الشهر وبترد على عملاءك 24/7. تحب أبعتلك لينك التسجيل وتبدأ مجاني الأول؟" One fitting tier, one link, one question. Full tiers live at OT1-Pro pricing — the AI picks the tier that fits instead of dumping the page into chat.

8. Ghosts, follow-up math, and the 27% we win back

From our Batch33 corpus of 12,847 real conversations, 73% of buyers ghost right after hearing the price. They do not object. They vanish. A bot that treats silence as rejection loses nearly three quarters of its pipeline at the moment of highest intent.

The same corpus shows the recovery: a 3-touch follow-up at 24 hours, 72 hours, and 7 days, each adding new value — never just checking in. That sequence wins back 27% of the ghosts. On a $79 per month plan, every 100 priced conversations means 73 ghosts, about 20 recovered buyers, roughly $1,580 in monthly recurring revenue a polite bot would leave on the floor.

That cadence is automated in the product — timing, value-add rule, and stop-on-reply logic are described in sales follow-up automation. To see how we mine those conversations for what buyers say before they ghost, read the sibling study AI deep analysis of 3,000 contacts.

1. 24 hours: one new useful thing tied to what they asked. "نسيت أقولك — بتشتغل على فيسبوك وانستجرام وواتساب من مكان واحد."
2. 72 hours: proof, not pressure. One line from a similar business.
3. 7 days: risk reversal. "تحب تبدأ مجاني وأنا أمشي معاك في الإعداد؟"

9. The capture ladder: when to ask for a name, and when to shut up

The old bot opened with "What is your name?" on turn one. Real closers never do. The trait hard-codes a capture ladder so every model asks for personal info at the lowest-friction moment:

1. Turn 1 (greeting or first question): NEVER ask for name, phone, or email. Engage, pitch one benefit, ask a qualifying question about their need.
2. Turn 2-3 (they showed interest or asked a real question): if business_type or business_name is on the capture list, this is the natural moment. "بتشتغل في ايه؟" fits here because it tailors the pitch.
3. Pricing or buying moment: ask for name plus the contact method they are writing on, framed as service — "ممكن أعرف اسمك عشان أبعتلك العرض؟" / "so I can send you the details" — never as fill out this form.
4. All fields captured: stop asking. Close warmly and confirm next steps.

One field per message, maximum, tied to what they just said. "Perfect, X sounds amazing — ممكن أعرف اسمك عشان أبعتلك التفاصيل؟" beats "What is your name, phone, and email?" every time. On refusal — "لا" / "later" / "why do you need my phone?" — give them the info they asked for first, build rapport, retry later. Never interrogate, and never repeat a question already answered. To enter that flow yourself, start at create your OT1-Pro account.

BAD ladder: Turn 1: "اهلا! اسمك ايه ورقم تليفونك؟" The customer feels trapped before learning anything.

GOOD ladder: Turn 1: benefit plus qualifying question. Turn 2: "بتشتغل في ايه؟ عشان أقولك أنسب استخدام ليك." Pricing moment: "تمام — ممكن أعرف اسمك عشان أبعتلك لينك التسجيل والخطوات؟"

10. Silence, energy matching, one idea per reply, and the spam guard

When buyers answer "لا" or "nope", the trait forbids apologizing or retreating into corporate mush. Pivot to a different angle: "ماشي — تحب تعرف إيه بالظبط؟ السعر، إزاي بيشتغل، ولا تجرب مجاني الأول؟" Short answers are direction, not rejection.

Energy matching is mandatory. Short messages get short replies. Formal gets formal, casual gets casual. Egyptian dialect gets Egyptian, Khaleeji gets Khaleeji. The language mirror is absolute: Arabic in, 100% Arabic out; English in, 100% English out; never an English sentence inside an Arabic reply. A mid-Arabic English refusal was a real production failure, so refusal phrases are banned in both languages.

DM length discipline: 1-2 short sentences, ONE idea per reply, ONE question or call to action. BAD: five features plus pricing plus "what is your name?" GOOD: "الميزة دي بتخليك ترد في ثواني حتى الفجر. تحب تشوفها على رسائلك انت؟"

I also preserved the [SPAM_DETECTED] token verbatim. Before any sales reply, the model judges abuse, trolling, gibberish, or hostility that would make a human rep stop. A hard price question is never abuse. On a hit it outputs exactly [SPAM_DETECTED] and nothing else. Once a human reactivates a conversation, the classifier stands down unless the latest message is explicitly abusive.

I rewrote this playbook because I read our own chats and felt embarrassed. The info-catching bot asked buyers what THEY sell. The closer I ship now listens first, pitches one benefit, handles the four real objections in the buyer's own dialect, asks for a name only when it can send something useful, and closes with a link instead of a lecture. Same trait, every model.

Stop losing the leads you already earned

OT1-Pro runs your follow-up, analysis, and AI replies in one inbox — WhatsApp, Instagram, Messenger, Telegram, and email, in Arabic or English, scored by lead quality, with every AI credit receipted in a transparent ledger. Free plan, no credit card.

Start free → · Sales follow-up automation · Lead follow-up software · Pricing · vs WATI · Talk to the founder on WhatsApp