Black Friday Customer Service With AI: A Practical Checklist

Black Friday Customer Service With AI: A Practical Checklist

To prepare Black Friday customer service with AI, decide before the peak what the agent can answer on its own (delivery times, shipping, order status, return rules), connect it to current stock and price sources, set sending limits, and define a clear path for a human to take over. AI absorbs repeated questions well, but only if the promotion rules are written down and there is a plan for when something goes off script.

Key takeaways

• The AI should answer only what is documented: promotion terms, returns, shipping and delivery times must be written before they go into the agent's knowledge base.
• Stock and prices change during the campaign, so the agent should query the live source and never rely on text copied the night before.
• Outbound messages have limits around consent, pacing and number quality. Sending without control is the shortest path to losing the channel.
• Every agent needs handoff rules that pass a short summary to a human, plus a switch to pause the automation.
• A rehearsal with real questions before the peak is worth more than any adjustment made during the campaign.

Why does Black Friday break customer service that works the rest of the year?

On a normal day, support varies little: questions repeat, rules rarely change and the team knows the exceptions. During a promotion, almost everything changes at once. New prices arrive, special conditions apply, products sell out within hours, delivery times shift and customers are more anxious than usual.

That creates three risks that AI can make worse if it is switched on without preparation. The first is answering with outdated information, for example confirming a product that has already sold out. The second is promising what the policy does not allow, such as a return outside the window or a discount that does not exist. The third is generating more outbound messages than the channel can handle, which hurts the reputation of your number.

The good news is that most of these problems are about process, not technology. Whoever writes the rules, sets the limits and tests before the peak greatly reduces the chance of surprises. For the broader question of what is worth automating and what should stay with people, read the guide on AI customer support: what to automate and what to keep human. If your business has predictable busy dates, how to prepare an AI agent for seasonal peak demand covers the same preparation in a more general way.

Which questions should you anticipate, and how do you build the answer base?

Start with your history. Gather support conversations from earlier promotions, emails, social media comments and the questions your sales team hears most. Group them by topic and pick the ones that repeat most. That list is your demand map, and it does not depend on any outside data.

In most stores, the expected questions fall into similar groups:

• Order status and delivery time, including what changes during the campaign.
• Product availability, size, color and voltage.
• Promotion rules: which items are included, whether there is a coupon, and whether it can be combined with other offers.
• Payment methods, installments and what to do when a payment is declined.
• Exchanges, returns, withdrawal rights and warranty.
• Address changes, cancellations and invoice copies.

For each group, write a short official answer with a source and an owner. If the answer depends on real-time data (status, stock), record that the agent must query the system rather than answer from memory. If the question requires judgment, such as a deadline exception, record that it goes to a person.

A common mistake is dumping one enormous document with every rule into the agent and expecting it to cope. Short, current documents organized by topic work better, and Anthropic discusses this principle of simplicity and well-defined tools in its material on building effective agents. For guidance on writing the instructions themselves, see how to write instructions for an AI support agent.

How do you define return, delivery and shipping rules before the peak?

AI does not create policy. It repeats the policy you wrote. That is why the most important work before Black Friday happens outside the agent: deciding and documenting.

For purchases made outside a physical store, Brazil's Consumer Defense Code provides a right of withdrawal. Because the interpretation of specific cases can vary, have the final text reviewed by whoever handles legal or compliance at your company. The agent should not give opinions on consumer law. It should state the approved policy and escalate doubtful cases.

A sequence for closing the rules:

• List each type of case: withdrawal, defect, wrong product, delay, lost parcel, size exchange.
• For each case, write the rule, the internal response deadline and who approves exceptions.
• Define what the agent can do alone (explain the procedure, open a ticket) and what only a person can do (approve a refund, grant an exception).
• Write the sentence the agent uses when it does not know, for example: I'll ask a team member to confirm this and get back to you here.
• Review the promotion policy item by item, because special offers often have their own rules.

If your company also answers complaints in public, align the official text with what will be said in those channels. The guide on handling consumer complaint sites with AI shows how to keep a record and human review in place, which prevents contradictory answers across channels.

How do you connect stock and prices to the AI without promising what does not exist?

Stock is where AI can help the most and also where it hurts the most when it is wrong. A customer who gets confirmation for a sold-out item is more irritated than one who was told from the start that the item is unavailable.

There are three levels of integration, from safest to riskiest:

• Inform only: the agent checks stock and price in the system and answers, without creating an order or reserving anything.
• Inform and guide: besides answering, it sends the correct product or cart link, where checkout validates stock at that moment.
• Act: it reserves products, changes orders or applies discounts. This should exist only with strict rules, human approval for non-standard cases and a record of every action.

For a first Black Friday with AI, the second level is usually enough. If your operation uses Bling, Omie or Tiny, the article on integrating AI with Brazilian ERPs via API details the options and the risks of each.

Practical rules for stock and price:

• The agent queries the source at the moment of the question. No tables copied into the prompt.
• If the lookup fails or is slow, the default answer is to say it will confirm, not to guess.
• Low-stock items should use cautious wording, such as subject to availability at checkout.
• Price and promotional terms come from the same source the website uses, to avoid mismatches between channels.

What sending and action limits should the agent have?

There are two kinds of limits. The first covers outbound messages, the ones your company starts. The second covers what the agent can do inside a conversation.

For outbound WhatsApp messages, the rule of thumb is to send only to people who gave consent, with relevant content and at a controlled pace. Waitlist notices, cart reminders and offer-start alerts can make sense, as long as the customer agreed to receive them. For details on approved templates, consent and best practices, see WhatsApp outbound messages: templates, opt-in and best practices. And since a badly planned sending spike can damage the quality of your number, also read how to avoid a WhatsApp number ban when automating with AI.

A list of limits to set before the campaign:

• How many outbound messages per contact during the period, with a clear ceiling.
• The allowed sending window by time of day.
• Authorized and prohibited segments, for example people who asked not to be contacted.
• Sending pace in batches, with an automatic pause if blocks or complaints increase.
• Actions the agent never performs alone: refunds, cancellations, changes to payment data, granting discounts.
• A step and time limit per conversation, so an agent cannot get stuck in a cycle. The article on AI agent infinite loops and step limits covers this, though you do not need to go to that level right now.

On personal data, apply the LGPD principle of collecting only what is necessary. The agent does not need a full ID document to give a delivery time. The ANPD, Brazil's data protection authority, publishes guidance that helps explain your obligations, and it is worth recording who at the company answers privacy questions during the campaign.

What contingency plan should you build for when something goes wrong?

Contingency is not pessimism, it is knowing your points of failure. In a campaign, the most likely scenarios are: the website or order system slows down, stock diverges from what the agent reports, the human queue grows beyond capacity, and the agent starts giving inappropriate answers.

Build the plan with four pieces:

• A switch: a simple way, known by more than one person, to pause the agent or reduce it to triage-only mode. Test this switch before the peak.
• A holding message: an approved text for when the system is unstable, saying the team has high volume and when the customer can expect a reply, without promising a deadline you cannot meet.
• A staffing rota: who takes the human queue, in which shifts, and who decides when to pause outbound messages.
• A log: every transferred conversation, every exception and every failure is noted for the next day's review.

To know when a conversation should leave the agent, use objective signals: a request to talk to a person, an angry customer, a payment or refund topic, a third attempt without resolution, or information the agent could not find. The article on sentiment analysis in support AI and when to hand off to a human goes deeper on this decision.

Observability helps here. If you log what the agent answered, from which source and when it transferred, you can spot a bad pattern on the first morning of the campaign. OWASP also maintains a list of risks for applications built on language models, including instruction manipulation and excessive agency, which works as a review checklist for anyone connecting an agent to real systems.

How do you apply map, limit, pilot and hand off before the campaign?

The method we use for AI projects follows four steps, and they fit a Black Friday preparation well.

• Map: collect the questions, systems and current rules, and note how support works today so you have a point of comparison.
• Limit: decide which data the agent sees, which actions it can take and which messages it can send.
• Pilot: test one real flow, preferably the most repetitive one, with conversation logs and an agreed metric, such as the rate of questions resolved without rework.
• Hand off: give the team the runbook, the switch and the list of who decides what, then decide whether to expand, adjust or stop.

An illustrative scenario: an online homeware store decides to use the agent for only three flows, order status, return rules and availability of the promotional showcase items. In the days before, the team builds a list of real questions from past promotions and tests the answers with colleagues who play difficult customers. They find that the agent confused the return rule for sale items with the one for regular items, and they fix the text. During the campaign, any question about refunds goes straight to a person with a summary of the conversation. This example is illustrative only and does not represent a client result.

To evaluate the agent before the peak, create a set of test questions with the expected answer and run it after every rule change. Anthropic and OpenAI publish material on evaluations (evals) that explains how to structure this kind of test. To measure the pilot itself, the article on how to measure an AI automation pilot offers metrics and decision gates.

What if there are only a few days left before the campaign?

When time is short, cut scope instead of skipping steps. A lean, safe version is worth more than an ambitious, untested one.

• Limit the agent to answering frequent questions and reporting order status.
• Leave anything that touches money or customer records to the team.
• Do not start new outbound sending. Use only lists that already have clear consent.
• Write the holding message and define the switch.
• Run a test with real questions and fix wrong answers before opening.
• Assign someone to read conversations in the first hours and adjust what is needed.

After the campaign, do a simple review: what the agent resolved, what it transferred, where it made mistakes and what the team had to redo. That becomes the base for the next stage.

Frequently asked questions

Can AI answer on its own throughout Black Friday? It can answer documented, low-risk questions on its own, such as delivery times, order status and published rules. Refunds, exceptions, serious complaints and sensitive data should go to a person. The best approach is to start with a few flows and expand as the logs show the agent getting them right.

Do I need to tell customers they are talking to AI? It is good transparency practice and helps set expectations, especially when there is an option to reach a person. Confirm with whoever handles legal and privacy at your company how to present it, and follow what the ANPD and the AI legislation under discussion in Brazil indicate.

Can I use AI to send promotional messages to my whole contact base? That is not recommended. Outbound messages should go to contacts who gave consent, with relevant content and at a controlled pace, using the templates the platform requires. Indiscriminate mass sending raises the risk of blocks and damage to your number's reputation.

How do I know if the agent is making mistakes during the campaign? Log the conversations and review a sample in the first hours, paying attention to transfers, complaints and answers about stock and price. Define in advance which signals call for a pause, such as contradictory answers or a rise in customers asking for a human, and keep the switch within reach of more than one person.

Sources

• Anthropic: Building effective agents
• OWASP Top 10 for LLM Applications
• ANPD: Autoridade Nacional de Proteção de Dados
• OpenAI: Evals guide