How to Write Prompts for AI Sales Replies: Price Floors and Handoff

An AI sales reply can read beautifully and still be unsendable. The root cause is almost never the model's writing ability — it's that the prompt covered tone and skipped business constraints. Write your price floor, banned promises, and human-handoff triggers into the prompt, and the AI stops being a fluent writer and starts being something you can actually send. In short, how to write prompts for AI-generated sales replies is about giving reps leverage, not replacing them.

Why your AI sales replies read well but you don't dare send them

Start with three crash scenes. A customer asks for a discount, and the AI replies: "No problem, let me apply another 5% for you." A customer asks about lead time, and the AI replies: "Guaranteed delivery in 3 days." A customer asks about after-sales, and the AI replies: "We offer lifetime free support." The salesperson stares at the screen and deletes the whole thing. One generation, zero usable output.

The problem isn't that the AI writes badly. It's that your instructions are missing the single most important category of information. Most people start their prompt like this: "You are a professional, enthusiastic foreign trade salesperson. Replies should be concise, polite, and sound human." That's all style instruction. It determines whether the reply sounds like someone on your team. It does not determine whether the reply can be sent. What determines that is business constraint instruction: the lowest price you're allowed to quote, the things you must never say, and the situations where the AI has to stop and ask a human. Almost nobody writes this layer.

The test is simple. Whether an AI reply can go out directly has nothing to do with how smooth it reads. It has everything to do with whether it crossed your price floor, your promise boundary, or its permission scope. Cross any one of those, and it's a dead draft no matter how polished it looks.

The three-layer prompt structure: style, facts, constraints

Split the prompt into three layers and you won't miss anything.

Style layer (how to write): tone, language, length, forms of address. For example: "Reply in English. Professional but not stiff. Keep each message under 80 words. Address the customer as Mr./Ms. plus surname." This layer decides whether it sounds like your team.

Facts layer (what to write): product specs, quoting rules, lead times, MOQ, certifications. This layer decides whether the reply has a basis. The AI doesn't know your real prices. If you don't give them, it will invent them.

Constraint layer (what must never be written): price floor, banned promises, situations requiring human handoff. This is the safety valve, and it's the layer people skip most completely.

A reusable skeleton looks like this:

1. Role definition: who you are, which customers you serve.
2. Available fact sources: answer only from the following knowledge base entries; anything not covered goes to the fallback script.
3. Hard prohibitions: no unilateral discounting, no specific delivery date commitments, no payment terms or exclusivity commitments.
4. Boundary fallback action: for discount approvals, contract terms, or compensation, always reply "Let me connect you with the colleague handling this" instead of answering directly.

Compare the two versions. A prompt with only the first three layers produces an AI that enthusiastically negotiates your price into the ground. Add the fourth layer, and it knows when to shut up.

Writing business constraints into the prompt: how to make price floors and banned promises stick

A price floor must be a concrete rule, not an adjective. Wrong: "Prices should be reasonable. Don't discount easily." Right: "The lowest quotable price is X. Below that, always reply 'I need to confirm with my manager and get back to you.' Never discount unilaterally. Never use suggestive phrasing like 'I'll try to apply for you' or 'that should be possible.'" Adjectives leave room for interpretation. Concrete rules don't.

Banned promises need to be listed one by one, each with a fallback script. Cover at least these categories:

• Lead time: you may say "usually 15–20 business days," never "guaranteed delivery by [date]."
• Stock: you may say "let me check current stock for you," never "plenty in stock, ships anytime."
• Payment terms: you may say "payment terms are something we can discuss," never "we can give you 60-day terms."
• Exclusivity: you may say "distribution policy needs commercial confirmation," never "you can have exclusivity for this region."
• Certification scope: you may say "we hold certification X; I'll send you the documentation on its exact scope," never "this certification covers all models."
• Warranty length: you may say "warranty terms are in the contract," never "lifetime free after-sales."

Write both the allowed and the forbidden phrasing for each item. Only then does the AI have a clear boundary to hold.

Close it off with "boundary means human handoff." For anything involving discount approval, contract terms, or compensation, the prompt should explicitly require the AI not to answer directly but to generate a line like "Let me connect you with the colleague handling this," returning the decision to a person. This step matters more than all the rules above, because it draws a line the AI has no authority to cross.

A common trap: the vaguer the constraint, the more the AI leans toward pleasing the customer. The AI's default goal is to keep the conversation moving. When a customer pushes, it naturally wants to offer something. Write "try not to discount," and it reads "discount if you have to." Write "below price X, hand off to a human," and it has no room to concede. A vague constraint is no constraint.

Where the facts layer comes from: don't let the AI invent product information from "general knowledge"

The AI doesn't know your real prices, real stock, or real certifications. Any fact not in the prompt, it can only guess — and invented numbers are the most dangerous kind, because they read exactly like real ones.

The fix is to build product knowledge into citable Q&A entries, and instruct the AI in the prompt to answer only from those entries, with everything else going to a fallback script. Turn "What's the MOQ?" "Which voltages does this model support?" "How long does shipping to Germany take?" into standard Q&A pairs. Then the AI quotes answers you've confirmed, not the industry average in its head.

The knowledge base has to keep updating. If a new product launch, a price change, or a policy shift doesn't get synced, the AI will answer customers with outdated information — worse than not answering, because the customer has already treated it as a commitment.

This is also why a single "universal prompt" can't carry a long-term business. The prompt governs rules; the knowledge base governs facts. You need both. The knowledge base design in Sellenca's features — which mines Q&A and scripts from real closed conversations and keeps updating — is exactly what carries the facts layer here. Writing "answer only from the knowledge base" into your prompt only works if someone is maintaining that knowledge base.

Switching prompts by sales stage: same constraints, different emphasis

The constraint layer stays identical across all stages. The style layer and the emphasis shift by stage.

Inquiry stage: the focus is quickly confirming needs and giving a compliant initial price range. The constraint layer must strictly prevent "quoting the floor price immediately" — many salespeople, trying to look sincere, give their lowest price in the first round and leave no room to negotiate. Write it into the prompt: first round quotes a range only, never the floor.

Negotiation stage: the focus is holding the price floor and trading concessions for conditions. Specify in the prompt which conditions can be traded (volume, payment terms, repeat-purchase commitments) and which cannot (unconditional unit price cuts). For example: "When the customer asks for a discount, steer toward volume tiers first rather than conceding on price."

Closing and after-sales stage: the focus is accurate lead time and commitment language. Forbid the AI from giving specific arrival dates or compensation plans on its own; route everything to "let me confirm and get back to you." Customers are most sensitive at this stage, and one "guaranteed delivery next week" can become evidence in a dispute.

Rather than maintaining one giant prompt, split it into a few by stage. Each version changes only the style layer and emphasis; the constraint layer stays as is. Maintenance cost drops and your messaging stays consistent.

How to verify after writing: three tests to confirm the AI won't cross the line

Stress test: ask in a customer's voice — "Can you do 10% cheaper?" "Guaranteed delivery next week?" "Can you give me exclusivity?" — and check whether the AI triggers the fallback script instead of agreeing. One wrong answer, and you go back and fix the constraint layer.

Spot-check real conversations: pull recent won and lost chat logs and compare AI-generated replies against what was actually sent, looking for constraint-layer bypasses. Where a salesperson edited by hand is usually where the prompt has a gap.

Build a review loop: collect cases where "the AI said something it shouldn't have" and feed them back into the prompt and knowledge base. Constraints aren't a one-time write; they get sharper as the business changes.

From prompt to platform: when constraints and facts need team-wide management

One person can keep a prompt in their head. When a team uses AI to reply to customers, price floors, banned promises, and product facts have to be consistent — otherwise every salesperson sends different words, and a customer comparing quotes spots it instantly.

Writing a good prompt is essentially distilling a veteran salesperson's judgment into rules. Do that well, and the AI goes from "can write" to "safe to send." But once conversation volume rises, maintaining constraints, facts, and scripts becomes ongoing operations work — new products to add, price changes to update, new banned items to insert. Copy-pasting prompts by hand will eventually miss something.

That's when it's worth considering a tool. Sellenca is a Chrome extension that sits on top of WhatsApp Web. Sales don't change numbers or migrate to the Business API. The AI drafts replies from the company's own knowledge base, and the salesperson confirms before sending. It puts constraints and facts into a shared admin console instead of scattering them across each salesperson's prompt drafts. If your team is growing, look at pricing to weigh the cost — $19 per seat per month, or $190 per seat annually — against the lost deals that inconsistent messaging causes.

FAQ

Is writing detailed tone and scripts enough for an AI sales reply prompt?

No, and that's the most common mistake. Tone and scripts belong to the style layer, which only decides whether the reply sounds like your team. What decides whether it can be sent directly is the constraint layer: price floors, banned promises, and human handoff. With only the style layer, the AI writes replies that read professionally but casually discount or over-promise, and the salesperson has to delete and rewrite.

I wrote a price floor into the prompt, but the AI still caves. What do I do?

First check whether the floor is written as an adjective. Phrases like "prices should be reasonable" or "try not to discount" leave room for interpretation, and the AI leans toward conceding when a customer pushes. Replace them with concrete rules: "The lowest quotable price is X. Below that, always reply 'I need to confirm with my manager and get back to you.' Never discount unilaterally." Then add a boundary fallback action that forces discount approvals to a human.

Which matters more, the prompt or the product knowledge base?

They govern different things, and you need both. The prompt governs rules — what can be said, what can't, when to hand off to a human. The knowledge base governs facts — the real answers on pricing, specs, lead times, and certifications. Writing "answer only from the knowledge base" into your prompt only works if the knowledge base has content and stays updated; and no matter how complete the knowledge base is, without a constraint layer the AI will still use real information to over-promise.

If everyone on the team writes their own prompt, will messaging drift?

Yes, and it's the most hidden risk when a team uses AI to reply to customers. Ask the same pricing question, and salesperson A's reply differs from salesperson B's. A customer comparing quotes sees it immediately. The constraint layer and facts layer should be maintained by the team; individual salespeople should adjust the style layer at most. That's why, once conversation volume rises, it's worth centralizing messaging in a tool rather than having everyone store their own prompt.

If you now see the value of the constraint layer and the facts layer and want to watch how the AI holds a price floor in a real conversation and hands off when a boundary is crossed, book a demo to see the actual flow.