AI selling on WhatsApp on Black Friday with control

Key takeaways

• A closed offer is the document listing product, price, terms, stock, delivery time, discount authority and approved links; without it, the AI should not handle conversations.
• Under the CDC (Brazil's Consumer Protection Code, Law 8.078/1990, art. 30), a sufficiently precise offer binds the supplier: a price invented by the AI becomes the company's obligation.
• Start in copilot mode; move to autopilot when the human correction rate stays below 5% for 3 consecutive days.
• The AI pauses when a human takes over and only resumes by explicit decision; a conversation with two "salespeople" at the same time loses the sale.
• Measure the AI by confirmed sales and by time until a human takes over, not by number of messages.

What a closed offer is and why the AI needs one

A closed offer is the set of information the AI can state without checking with anyone: products on promotion, the price of each one, payment terms, available stock, delivery time, discount authority (how much it can grant on its own) and the approved payment links. Everything outside that set is "I don't know, I'll confirm with the team."

The reason is legal before it is technical. Under the Consumer Protection Code (Law 8.078/1990), art. 30, every sufficiently precise offer binds the supplier and becomes part of the contract; under art. 35, if the company refuses to honor it, the consumer can demand enforced performance. An AI that replies "yes, it's in stock for R$ 299" when the price is R$ 399 has created a valid offer. Art. 37 also prohibits misleading advertising, including by omission.

On Black Friday the offer changes several times a day. Define who updates the knowledge base and how fast. Rule of thumb: if the offer changed and the base was not updated within 15 minutes, put the AI in copilot mode until it is synced. How to build the offer itself is covered in how to build a clear offer for WhatsApp.

Copilot or autopilot: how to choose the mode

Copilot mode is where the AI drafts the reply and a human approves, edits or discards it before sending. Autopilot mode is where the AI sends on its own and the human only steps in when called. The choice depends on three variables: the risk of a wrong answer, volume and the maturity of the knowledge base.

Transition rules:

• If the human correction rate (replies edited or discarded ÷ replies suggested) stays below 5% for 3 days, release autopilot for that type of question.
• If the rate rises above 10% on any day, go back to copilot mode and review the base before switching it back on.
• If the customer asks for a discount above the authority, the AI states its limit and escalates; it does not negotiate.

Tools such as RASTRO's AI sales agent operate in both modes and pause when a human takes over the conversation; the decision about the mode still belongs to the media buyer and the client.

When the AI should pause and hand off to a human

The AI should stop replying in five situations, and each one needs a human owner with a resumption deadline.

1. A human took over the conversation. Any message sent by a person on the team pauses the AI in that conversation. Resumption is the agent's explicit decision, not automatic.
2. The customer asked for a human. "I want to talk to someone" ends the AI's involvement immediately; insisting is the fastest way to get blocked.
3. Question outside the closed offer. Price of an unlisted item, unconfirmed stock, delivery time for a city outside the table: the AI says it will confirm and opens a pending item for the team.
4. Discount request above the authority. The AI states what it can do and escalates.
5. Sign of a complaint or cancellation. WhatsApp purchases have a 7-day right of withdrawal under art. 49 of the CDC; the AI should not argue against that right.

A reasonable target for a human to take over a pending item opened by the AI: 15 minutes during business hours. The capacity calculation for that is in WhatsApp SLA and queue on Black Friday.

How to test the AI before releasing volume

Test with written scenarios, 20 to 30 simulated conversations per type, before switching on the media, with questions asked by someone who did not configure the base.

Mandatory scenarios:

• "I want to buy, send me the link": the AI must send the approved link for the right product, without restarting discovery.
• "Is there a discount?": a reply within the authority and nothing beyond it.
• "Is it in stock?": only confirms if the base says so; otherwise, it checks with the team.
• "Will it arrive before Christmas?": delivery time from the table or escalation.
• "I want to talk to a person": escalates immediately.
• Two undecided options: asks only for the information needed to choose (at most 2 questions before the price, as in lead qualification without stalling sales).

Approval criterion: zero replies with invented price, stock or delivery time across all scenarios. A single error of that kind fails the test, because on Black Friday it repeats at scale.

Hypothetical example: what the AI changes in team capacity

A hypothetical example, with made-up numbers to show the math. A store receives 400 conversations per day during Black Week, with 3 agents who each handle 50 conversations per day (capacity: 150).

1. Without AI: 400 − 150 = 250 conversations per day without adequate service (62% of the total).
2. With AI in autopilot: the AI gives the first response in 100% of cases and handles the closed-offer questions on its own. In testing, 30% of conversations ask for a human or go outside the offer: 400 × 0.30 = 120 conversations reach the team.
3. Remaining capacity: 150 − 120 = 30 conversations of slack; the human queue clears.
4. Correction rate (measured in copilot mode the week before): 12 replies edited out of 300 suggested = 4%. Below 5%, autopilot is cleared.
5. Sales: 280 conversations handled by the AI generate 22 confirmed sales (7.9%); the 120 handled by humans generate 18 (15%). Humans convert more, but serve a third as many. Total: 40 sales versus the roughly 22 that 150 human conversations would yield on their own.

The correct reading: the AI did not sell better than the team; it prevented 250 conversations from going unanswered. Comparing the AI's conversion rate with the humans' without considering volume leads to the wrong conclusion of "turn the AI off."

Which metrics prove the AI is under control

Measure the AI by the same metrics as a salesperson, plus three control metrics, reported together.

"Confirmed sale" means payment recorded in the CRM, not a link sent. Messages exchanged, conversation length and "engaged leads" do not go into the report as results.

Common mistakes when putting AI to sell on Black Friday

• Switching on autopilot on day 1 without copilot mode. The team discovers the errors through customers, at peak volume.
• Granting discount authority with no cap. "You can negotiate" becomes the lowest possible price in every conversation.
• Promising what the system has not confirmed. Credit approval, stock and delivery time only with data; a promise without data becomes a complaint. How to avoid it is in stock and delivery time without wrong promises.
• Sending proactive messages without opt-in. The WhatsApp Business messaging policy requires consent for proactive contact and, according to Meta's pricing documentation (2025), free-form messages are only valid within the 24-hour window opened by the customer; the AI only re-engages those who agreed to be contacted.

Frequently asked questions

Sources

• Consumer Protection Code (CDC), Law 8.078/1990 (arts. 30, 35, 37 and 49)
• Meta, WhatsApp Business Platform pricing documentation (2025)
• WhatsApp, WhatsApp Business messaging policy