How to Answer Price Questions on WhatsApp With AI: A Practical Guide

How to Answer Price Questions on WhatsApp With AI: A Practical Guide

To answer price questions on WhatsApp with AI, the agent should pull the value from an official company source (a price list, catalog or system), ask the few questions that change the price before quoting, and hand the conversation to a person when the case falls outside the rules. The price must never come from the model's memory or from an invented estimate. When there is no reliable data, the correct answer is to explain how the price is built and bring in a human.

Key takeaways

• The AI only states a price that came from an official source, with an effective date and a named owner.
• If the price depends on variables, it explains what changes the value and asks the minimum needed before quoting.
• Discounts, negotiation, volume deals and anything outside the price list go to a person, with the context already summarized.
• If the price lookup fails or returns nothing, the AI does not estimate: it says it will confirm the value and hands off.
• Every price answer should be logged with the source used, so it can be audited later.

Why price is the most delicate question on WhatsApp

Someone who writes "how much does it cost?" is usually at the start of a decision. That message determines whether the conversation continues or the customer goes to another company. The question deserves a fast answer, but speed without control creates two problems.

The first is the wrong number. A language model is good at producing plausible sentences, and a plausible number looks correct even when it exists in no price list. In the chat, the customer can take a screenshot and hold you to that figure. The second problem is the opposite: the AI becomes so cautious that it only says a consultant will be in touch, and the customer waits without getting anything useful.

The goal, then, is a useful answer that does not promise what the company will not honor. It is worth agreeing with your legal advisor on how messages with prices and terms are treated in your line of business, because what the AI writes speaks on behalf of the company. For the agent's setup, our article on how to write instructions for an AI support agent explains how to record these rules.

Where can the AI get the price from?

The core rule is simple: the AI states only what is in an official source. In practice, that means a few design decisions.

• One source for each type of price: product list, service list, discount rules and payment terms. When two different spreadsheets exist, the AI will repeat the discrepancy.
• A tool lookup, not text pasted into the instructions. A long table copied into the prompt goes stale without anyone noticing. A query to the system (ERP, CRM or a controlled spreadsheet) returns the current value.
• An owner and an effective date. Each price has a person responsible for updating it and a date until which it is valid. If the validity has expired, the AI does not answer with it.
• Terms alongside the value. The answer says what is included, how long the quote is valid and what is not included.
• A safe failure. If the lookup errors out, comes back empty or returns more than one possible result, the AI does not pick one or estimate. It says it will confirm and hands off.

When part of the rule is written text, such as shipping policy, warranty terms or installment rules, the technique of retrieving the right passage from approved documents before answering helps a great deal. It is commonly called retrieval-augmented generation. For numeric prices, though, a structured query to the system is safer than reading a document.

It is also worth planning for a customer who tries to force a value, for example with a message telling the AI to ignore the price list. Industry security guidance on applications built on language models covers this kind of abuse and the risk of giving an agent too much autonomy. The practical defense is to keep the price out of the conversation's reach: the AI reads the value but has no tool to change it.

When should the AI ask questions before quoting?

If the price is fixed and public, the AI should answer first and ask later. Making the customer go through questions before hearing the number is the fastest way to lose them. When the value depends on variables, though, giving a bare number is risky, and qualification becomes part of the answer.

The logic is to explain why you are asking. Instead of opening an interrogation, the AI says the price changes depending on a few points and asks about the first one. This approach fits what we describe in lead qualification on WhatsApp without the form.

Good qualifying questions have three traits: they genuinely change the price, the customer can answer them, and they do not require unnecessary personal data. Generic examples:

• Which product or service exactly does the customer want?
• What quantity or size?
• In which city or region will it be delivered or performed?
• What timeline do they need?
• Is it for an individual or a company?

Ask for one or two pieces of information per message. Collect only what is needed to quote, in line with the data minimization principle of Brazil's LGPD. If the AI needs a personal or company tax ID (CPF or CNPJ), that should happen at a later step and with a clear purpose, not in the first reply about price. The Brazilian data protection authority (ANPD) publishes official guidance on the subject.

Decision guide: what the AI does for each type of price question

A short guide, written before the agent goes live, prevents improvisation. Adapt it to your operation:

• Product or service with a fixed price in the list: states the value, what is included and the validity, and offers the next step.
• Price that depends on variables: explains what changes the value, asks the first question and only quotes when the official rule covers the data received.
• Officially published price range: may state the range, making clear it is a range and not a quote.
• Request for a discount or special terms: does not negotiate and hands off to a person.
• Comparison with a competitor's price: does not comment on the competitor and hands off if the customer insists.
• Price not found or lookup error: says it will confirm the value and hands off, without estimating.
• Question about installments or payment method: answers only what is registered as a rule; anything else is handed off.
• Customer who cites a figure quoted in another conversation: neither confirms nor denies, and hands off with the history.

When should it hand off to a human?

A handoff is not a failure of the AI. It is part of the design. The situations where a person should take over are usually these:

• Any negotiation: discount, special deadline, commercial terms outside the price list.
• Large-volume requests, contracts or corporate customers with their own rules.
• Custom items that do not exist in the catalog.
• An angry customer, a threat to complain, or a mention of a price charged before. To detect tone, see sentiment analysis in customer support.
• A mismatch between what the customer says and what the system shows.
• Any moment when the AI has no official source to support the answer.

The quality of the handoff matters as much as the decision to hand off. The AI should give the agent a summary of what the customer wants, what has already been answered, what data was collected and why the conversation was transferred. The customer should not have to repeat everything. It is also important to be honest about when a person will reply: if it is outside business hours, the AI says so and states when to expect a response, without promising what the team will not deliver.

Illustrative scenario: an installation quote

Imagine, as an illustrative scenario, a company that installs equipment and has an official price list by service type, with a surcharge by region. The customer writes: how much does installation cost?

The AI recognizes that the value depends on the type of equipment and the city. It replies:

The installation price changes with the type of equipment and the city. Which equipment do you want installed, and in which city?

The customer provides both. The AI queries the price list, finds the service and the regional surcharge, and replies with the value, what is included and until when the list is valid. It then asks whether the customer wants to schedule a visit.

If the customer says the installation involves a difficult wall and asks for a discount as a longtime customer, the AI does not improvise. It writes that it will pass the case to a team member to evaluate, hands off with the summary and logs that the transfer happened because of a special-terms request. If the system lookup had failed instead, the response would be the same: say the value will be confirmed and bring in a human. At no point does the AI guess.

How to roll it out: map, limit, pilot, hand over

The method we use in customer service projects follows four moves, and they apply well to price questions.

• Map: document how prices are communicated today, who answers, where the number comes from, which questions come up most and which exceptions exist. This snapshot is the baseline for comparison later.
• Limit: define what the AI can query (read-only access to the official price list), what it can say and what it never does, such as negotiating, changing a value or promising a deadline. Also limit the data it collects. The article on guardrails for AI agents shows how to validate input, output and actions.
• Pilot: pick one real flow, such as the price of a specific service, and run it with logs of every answer, the source consulted and the handoffs. Agree on a success metric beforehand, for example how many price answers matched the price list and how many cases were handed off for the right reason.
• Hand over: at the end of the pilot, deliver the operation with the rules documented and decide from the logs whether to expand, adjust or stop.

If the company does not yet have WhatsApp connected to a system, the step before this is understanding how to integrate AI into your company's WhatsApp. For the anatomy of agents that call tools and know when to stop, Anthropic describes patterns in its guide Building effective agents.

One last precaution: review real answers regularly. Prices change, promotions end, and the official price list needs an owner who flags when something has changed. The AI is only as correct as the source it consults.

Frequently asked questions

Can the AI give an approximate price when it cannot find the exact value? Only if the company has officially published a range and the AI makes clear it is a range, with the conditions. Estimating on its own is the way to create expectations the team will have to undo. Without a source, the best move is to say the value will be confirmed and hand off.

Can I put the price list directly into the AI's instructions? It works in small tests, but it tends to go stale without warning, and the AI keeps answering with the old value. The safer approach is to query the system or a controlled price list with an owner and an effective date. That way, changing a price does not require rewriting the agent.

What if the customer only wants the price and does not answer the questions? When the price is fixed, the AI answers directly and asks for nothing. When it depends on variables, the AI explains why it is asking and can offer the official range, if one exists, or pass the customer to a person. If the customer goes quiet after the quote, the article on following up on unanswered quotes shows how to pick the conversation back up without nagging.

Can the AI grant a discount? We recommend it does not. A discount involves margin, the relationship and a commercial decision, which should stay with a person. If the company has objective, approved rules, such as a fixed discount for upfront payment, the AI can state that rule, but anything outside it goes to a handoff.

Sources

• OWASP: Top 10 for LLM Applications
• AWS: What is retrieval-augmented generation
• Anthropic: Building effective agents
• ANPD: Autoridade Nacional de Proteção de Dados