Customer service: what it is, types, the rules, and how to build yours

What is customer service? The one-sentence definition

Customer service is the set of interactions in which a business receives a person, works out what they need and takes the request through to a solution. If someone asks you to define customer service in one sentence, that is it: receive, understand, resolve. It applies before, during and after the sale, at the counter, on the phone, on WhatsApp, by email or with an AI agent.

A customer service department is the same activity seen as a company structure: the people, channels, deadlines and systems that make it happen. The terms vary by country. In the US, customer service and customer support are often used as synonyms. In Brazil, the acronym SAC has a legal meaning with its own rules, and Spain's law speaks of customer attention services. When a rule applies, check what the local text means by the term.

Good customer service comes down to three questions, in this order. Was the problem solved? How long did it take? How much work did it take the customer? Friendliness helps, but it does not make up for an unsolved case. A kind reply that solves nothing creates a second contact, and the company pays for that service twice.

Official data shows where this hurts. Brazil's government complaint portal, Consumidor.gov.br, closed 2,328,863 complaints in 2025, according to the annual bulletin of the National Consumer Secretariat. Billing and disputes made up 52.14%, contracts and offers 12.56%, and customer service itself 11.91%. The average resolution rate was 81.43%, with a response in six days on average. Only companies that signed up voluntarily appear, so the picture is partial, but it is official.

The effect also shows up in customer loyalty, the decision to buy again from the same company. The CX Trends 2026 study, by the Brazilian platform Octadesk with Opinion Box, surveyed 2,000 consumers: 85% say response speed strongly influences whether they buy from the same company again. The most cited criteria for good service were speed (51%), accurate answers (43%) and kindness (40%). It comes from a vendor in this market, so read it with that in mind.

Customer service, support, help desk, success and experience: what's the difference?

The five terms get used as synonyms, but they point to different things, and the confusion gets in the way when you need to decide who answers for what. Customer service is the umbrella: every conversation with someone who buys or wants to buy from you, including the question they ask before the sale.

Customer support is the part of customer service that fixes technical problems with use: the system that won't open, the integration that stopped, the device that won't turn on. It takes product knowledge and often a second technical tier. A help desk is the formal channel for information, questions, complaints, disputes, suspensions and cancellations. Brazil's 2026 decree makes the line explicit: service aimed at offering and selling products falls outside its help desk rules (art. 3, § 4).

Customer success is the work of following an account so it gets results from the product, without waiting for a problem to surface. It is proactive and more common in subscriptions and business-to-business sales. Customer experience, or CX, is the sum of everything a person goes through with a brand: ad, website, price, delivery, invoice and service. Service is one part of it, and it cannot fix on its own a delivery that is always late.

Here is one fictional customer of a subscription business. He asks in the chat which plan suits him, and that is pre-sales service. The app won't open, and that is support. The invoice arrives doubled and he disputes it, and that goes to the help desk. The company notices he stopped using the product and reaches out to help, and that is customer success. The impression he keeps of all this, along with the product, is the customer experience.

In a small company, the same person does almost everything, and that is fine. What needs to be written down is who decides in each case: who approves a refund, who answers a complaint, who reaches out to the customer who went quiet. When you measure, separate the reasons, because a queue mixing sales, technical support and complaints tells you nothing about any of them.

What are the types of customer service?

Guides tend to list types by mixing channel with method. It is clearer to split them along four axes, and every interaction has a position on each one. The first is timing: pre-sales, which answers questions and helps people choose; sales, which closes the order; and post-sales, which handles delivery, use, returns, billing and cancellation.

The second is who starts it. In reactive service, the customer comes to you. In proactive service, you tell them before they ask: a late order, a service outage, an invoice about to fall due. One good proactive notice prevents a queue of identical questions. You can start with a single notice, for whatever generates the most contacts today.

The third is rhythm. Synchronous means real time, such as phone, in person and chat with the person waiting. Asynchronous is the conversation each side answers when they can, such as email and, most of the time, WhatsApp. This changes your staffing math: on the phone, an agent talks to one person at a time; over messages, the agent alternates between several conversations, and the customer does something else while waiting.

The fourth is who handles it: a person, an automated system or both, with self-service and AI up front and people behind them. The combined model only works when the handoff is simple. Self-service includes help centers, FAQs and order lookup without talking to anyone, and it suits people who only want the information.

Customer service channels: which to open and how to choose

Choose the channel by the type of problem and by what your team can sustain. The phone suits urgent cases and people who prefer to talk, but it ties up one person per call. WhatsApp and other messaging apps suit everyday cases, because they take photos, voice notes and documents and keep the conversation. Where your customers already message you, this is often the first channel worth organizing.

Instagram, Messenger and Telegram reach people who arrived through social media or prefer that app, and the question tends to be short. Email suits long topics, attachments and anything that needs a paper trail. Website chat catches the question at the moment of purchase. In-person service is still needed when the case depends on seeing the product or on a signature.

SMS, RCS and push are notice channels more than conversation channels: access codes, confirmations, orders on the way, invoices. They support proactive service and bring the person back to the right channel when they are not in the app. A help center also belongs on the list, because it is the channel that creates no queue.

The practical rule is not to open a channel your team cannot staff. An Instagram profile that answers days later is worse than having no channel, because the customer wrote and was left waiting. Open one at a time, with an owner and a deadline, and say on your website which channel handles what and when. If your volume is on WhatsApp, the WhatsApp customer service guide covers the official number, triage and the 24-hour window.

With two or more channels, the real problem appears: the customer tells the story on Instagram, repeats it on WhatsApp, and nobody knows it is the same person. Multichannel means having several channels. Omnichannel customer service means having them in one queue and one record, with the history together. For many small teams, the gain comes less from opening channels and more from joining the ones they already have.

Human customer service in practice: bad message, good message

Human customer service means treating the person as someone with a specific problem, and not as a ticket number. In practice, that means using what they already told you, owning what is yours, saying what happens next and giving a deadline. Emoji and “valued customer” don't do that work. A short, exact reply signed by someone does.

A delay. Bad: “Dear customer, your request has been received and will be reviewed shortly.” Good: “Hi Emily. Order 48213 left our warehouse yesterday and the carrier expects delivery by Thursday. If it doesn't arrive, I'll open a claim with them and message you Friday morning. I'm Julia and I'll stay on your case.” The second one says where the order is, what happens if it goes wrong and who is answering.

A company mistake. Bad: “We apologize for any inconvenience caused.” Good: “The double charge was our mistake. I canceled the second one today and the receipt is attached. The amount comes back on your next invoice.” “Any inconvenience” avoids saying what the mistake was. Saying what went wrong takes fewer words and closes the matter. Only promise a deadline you can keep.

A refusal. Bad: “Unfortunately that isn't possible under our policy.” Good: “The coupon expired on October 10, so I can't apply it to this order. What I can do is hold the item until tomorrow at today's price, if you'd like to think it over.” Saying no is part of the job. The customer needs the concrete reason and, when there is one, a real alternative.

A transfer. Bad: “I'll transfer you to the responsible department. Please explain the issue again.” Good: “I'm passing you to Mark in billing, who can change the invoice. I've told him what you explained, so you don't need to repeat anything. He'll reply within 30 minutes.” Making customers repeat the story is the easiest failure to eliminate, and all it takes is a well-written internal note.

The first automated message. Bad: “To continue, enter your ID number and contract number.” Good: “I'm the store's virtual assistant, an AI. I can check orders, invoices and returns. If you'd rather talk to a person, type agent at any time.” The second says what it is, what it can do and how to get out. For people writing after hours, there are templates in the guide to greeting and away messages.

What the law requires of customer service, and from whom

The rule depends on where you sell and what you sell. In the US, the FTC's Mail, Internet, or Telephone Order Merchandise Rule covers orders placed remotely. You must ship within the time you state, or within 30 days if you state none. If you learn you can't, you must tell the buyer without being asked and offer a choice: accept the delay or cancel for a prompt refund.

The same FTC guide covers refunds. For cash, checks, money orders or payment through a third-party creditor, the refund is due within seven working days after cancellation, and store credit cannot replace it. Silence can count as consent only for a first delay of 30 days or less. Your team needs a delay template that says this plainly, because “we apologize for any inconvenience” does not offer the buyer that choice.

Subscriptions are where many guides are out of date. The FTC's 2024 click-to-cancel amendments are not in force: the Eighth Circuit vacated them in July 2025, and on February 12, 2026, the FTC restored the earlier rule text. In March 2026 it opened a new comment process. The Restore Online Shoppers' Confidence Act (15 U.S.C. 8403) still requires online negative option sellers to provide simple mechanisms to stop recurring charges.

In Europe, GDPR Article 22 gives people the right not to be subject to decisions based solely on automated processing that have legal or similarly significant effects. Article 50 of the AI Act, applicable since August 2, 2026, requires that people be told they are interacting with an AI system, unless that is obvious. Spain's Law 10/2025 goes further for basic services and large consumer companies, as the next section shows.

In Brazil, the Consumer Defense Code (Law 8,078/1990) applies to every business selling to consumers, and online stores must reply within five days under Decree 7,962/2013. Decree 13,135/2026, in force since November 2, applies to federally regulated suppliers that meet their regulator's criteria and to digital financial services supervised by the Central Bank. An ordinary retailer is not automatically covered. Check with counsel in each market where you operate.

What a strict customer service rule looks like: Brazil's 2026 decree as a yardstick

Even if you never sell in Brazil, Decree 13,135/2026 is worth reading because it turns vague ideas into numbers. Covered companies must offer a person from 8 a.m. to 8 p.m., seven days a week, on the phone, which remains mandatory, and on electronic channels, which expressly include chat and instant messaging. At least one electronic channel runs 24 hours a day. That adds up to 84 hours a week of human coverage, more than one shift can carry.

On electronic channels, the limit is five levels to finish a request, with no pushing the customer to another channel unless an in-person visit is essential. At any level, including the first, there must be an option to cancel right away and an option to reach a person, with the estimated wait stated. Starting service and reaching a person cannot depend on giving data first. AI may triage, route and resolve, as long as the human option exists.

On the phone, the wait is up to 20 seconds for at least 80% of calls answered and up to 300 seconds for the rest, unless a regulator sets otherwise. If a call drops, the company calls back within 300 seconds. For comparison, Spain's Law 10/2025 requires 95% of requests for personal attention to be answered, on average, in under three minutes, and bans answering machines and similar means as the only form of service (art. 8).

Deadlines and records. Brazil's decree sets a clear and conclusive response within seven calendar days, and resolution of an unrequested service or wrongful charge within 48 hours. Cancellation takes effect immediately, without requiring another menu, agent or channel. Every request gets a reference number, history goes out within five business days and stays available for six months, and messages are kept for 90 days. Ads are banned at any stage. Spain sets 15 working days, or five days for billing disputes (art. 17).

Comparing the decree with the 2022 text it replaced shows the direction of travel. It now names digital financial services. Human phone coverage rose from a floor of eight hours a day to 12 hours every day, and the human option reached electronic channels. Ads during the wait, once allowed with consent, are banned. Five levels, the 20- and 300-second waits and the 48-hour limit are new. Record keeping went from two years to six months, and sending history from five calendar days to five business days.

If no such rule applies to you, it still works as a yardstick. Four points suit any operation and cost little: a visible way to reach a person, no asking for data before hearing the problem, no making the customer repeat the story, and a reference number they can quote. The 84 hours are a business decision. If you are not required to offer them, shorter hours stated clearly beat a promise your team cannot keep.

Customer service metrics: CSAT, NPS, CES, wait time, handle time and FCR, with formulas

Six metrics cover almost everything. The numbers in the examples below are fictional and only show the math; they are neither targets nor market averages. CSAT is satisfaction with one interaction. CSAT = responses rated 4 or 5 ÷ total responses × 100. With 160 high ratings out of 200 responses, CSAT is 80%. The CSAT guide covers the question and the calculation in detail.

NPS measures willingness to recommend the company, not a specific interaction. The question, from 0 to 10, asks how likely the person is to recommend the company to a friend or colleague. According to Bain, 9 and 10 are promoters, 7 and 8 are passives, and 0 to 6 are detractors. NPS = % promoters − % detractors. With 50 promoters, 30 passives and 20 detractors out of 100 responses, NPS is 30.

CES measures how much effort the customer spent to get something done. The metric was introduced by CEB, now part of Gartner, in a 2010 Harvard Business Review article. There is no single scale: IBM gives examples from 1 to 5 and from 1 to 9. CES = sum of ratings ÷ number of responses. With 410 points across 100 responses on a 1 to 5 scale, CES is 4.1. Keep the same scale, or the comparison loses its meaning.

Average wait time is how long people wait until service starts. Wait time = sum of wait times ÷ number of interactions. With 1,200 minutes across 300 interactions, it is 4 minutes. Average handle time is how long service takes. Handle time = sum of interaction durations ÷ number of interactions. On the phone, the math is direct. Over messages, a conversation stays open for hours and the number misleads. Measure the agent's active time, or use first response time and resolution time.

FCR is first contact resolution. FCR = cases resolved with no further contact ÷ total cases × 100. With 240 cases out of 300, FCR is 80%. Write down what counts as a further contact, for example the same customer coming back for the same reason within seven days. Without that rule, each person on the team measures it differently, and closing a conversation early is enough to push the number up.

Read metrics in pairs, because each one alone can mislead. Low handle time with low FCR usually means people closing early. High CSAT with few responses may only sample those who were happy. Good wait time during business hours can hide the overnight queue. Start with three, first response time, FCR and CSAT, and add another when someone will act on it.

A customer satisfaction survey template you can copy and use

Ask right after the case is solved, in the same channel, with few questions. A customer satisfaction survey template that fits in a conversation: “Your conversation with Julia is closed. From 1 to 5, how satisfied were you? 1 is very dissatisfied and 5 is very satisfied.” After the rating: “Was it easy to get what you needed? Reply yes, somewhat or no.” Finally: “Want to tell us what we could have done better? This is optional.”

The three questions give you CSAT, a simple version of effort and the reason. Don't ask for a name, email or any data you already have, because the answer is already tied to the conversation. Don't add an offer: Brazil's decree bans advertising on help desk channels. Spain's 2025 customer service law goes further and bans satisfaction surveys before the case is resolved (art. 18). Even where no rule applies, that is a sound habit.

Watch for three traps. Don't let agents choose which conversations get the question, because the sample turns into praise. Always read the score next to the number of responses. And treat every 1 or 2 as an open case: someone on the team talks to that customer before looking at the average. That follow-up conversation usually teaches more than the dashboard.

How to set up customer service on a small team

First, one queue. All channels land in the same place, and each conversation has one of four states: new, in progress, waiting on customer and resolved. That way, the question “what still needs an answer?” has an answer on screen, and nobody opens four apps. The WhatsApp team inbox guide shows how to put several people on the same number without relying on one phone.

Second, one owner per conversation and one owner per shift. The conversation owner replies until the end or passes it on with a note. The shift owner watches what has no owner and what is close to breaching, and redistributes. On a team of two or three, this role can rotate daily. Also write down who covers lunch, days off and weekends, because those gaps are where deadlines slip.

Third, deadlines by priority, instead of a single deadline. A starting point to adjust to your business: urgent, such as a wrongful charge, fraud or a service outage, gets a first response in 15 minutes and a solution the same day; normal gets one hour and two business days; low gets the same day and three business days. This is a suggestion, not a market average. Where a legal deadline applies, it sits above these.

Fourth, a knowledge base. Read your last 100 conversations, note the reason for each one and write the answer to the 20 questions that come up most, with the right steps. Each answer has an owner and a review date. Use the same base for the team and for the AI, or the two versions will drift apart. A canned reply is a starting point: the agent adjusts the name, the order and the case before sending.

Fifth, the handoff to a person. Define the triggers: the customer asks, the base has no answer, two attempts failed, or the topic is billing, cancellation or a serious complaint. Every handoff carries a note in this format: “Customer wants: X. Already checked: Y. Missing: Z. Promised: what and by when.” Whoever receives it reads the note and continues without asking again. The rule also applies to the bot: without a note, the handoff isn't ready.

Sixth, a 30-minute weekly routine. Look at the five reasons that generated the most contacts, the five conversations with the worst scores and those that missed their deadline. Leave with one action: fix a text in the base, warn customers before they ask, or take a recurring problem to the area causing it. In Brazil's official bulletin, more than half of complaints are about billing. In your business, the top reason may also start outside customer service, and that is where it gets solved.

Common customer service mistakes

Measuring only speed. When the target is handle time, the team learns to close conversations. The case comes back, the company pays twice and the customer repeats the story. Measure resolution together with time, and make it clear that nobody is rewarded for quickly closing a case that returns.

Hiding the way to a person. A bot with no path to an agent blocks those who need help most. WhatsApp's messaging policy, updated on September 23, 2026, allows automation within the 24-hour window as long as there is a ready, clear and direct path to human support, such as a transfer in the conversation, a phone number or an email. Under Brazil's decree and Spain's law, the human option is an obligation for covered companies.

Asking for sensitive data in the conversation. The same policy bans asking for full payment card numbers, financial account numbers or identity document numbers. To find the customer, the order number or the phone number of the conversation itself is usually enough. Never ask for a password, code or card number, and say so on your website, because every service channel ends up imitated by scammers.

Sending the customer to another channel. “Call our 800 number” or “open a ticket on the website” hands the work back to someone who already came to you. On electronic channels covered by Brazil's decree, that is banned unless an in-person visit is essential. Everywhere else, it is a reliable way to lose patience and customers. Solve it where they are, or take the case yourself to where it needs to go and come back with the answer.

Promises with no owner and ads in the middle of a problem. “We'll look into it and get back to you,” with no name or deadline, leaves the customer not knowing what to expect. Offering a promotion to someone complaining irritates them, and it is banned on help desk channels under Brazil's decree. And opening a new channel while the last one's queue still has no owner only moves the wait somewhere else.

How to choose a customer service system: criteria, and per-seat vs per-message pricing

Before any demo, write down three numbers: conversations per month, channels used and people answering. Then assess seven criteria. Real channels, with the official API for WhatsApp, because the WhatsApp Help Center treats unofficial apps as a breach of its terms, and a WhatsApp support system built on one puts your number at risk of a ban. A queue with automatic routing and transfers with notes. Visible deadlines, with a warning before a breach.

The other four. A customer record with the history of all channels and easy export; under Brazil's decree, covered companies keep messages for 90 days and history for six months. Reports with the metrics above, by reason and by person. AI that answers from your knowledge base, says it is AI and hands off to a person with context. And role-based permissions, security and data location, which matter for GDPR and for Brazil's LGPD.

The pricing model changes the bill more than the price list suggests, and there are at least three models. Per seat, you pay a license per person, as at Zendesk, according to its pricing page read on November 2, 2026, which charges AI per resolution. It is predictable with a stable team and grows with each hire. Per active contact, as at Octadesk, the same day's reading shows unlimited users and plans by DAU, a contact with a conversation in a 24-hour window, with templates charged separately.

Per message, you pay for what is sent, and team size stays out of the bill. That is the CCX Messages model: a monthly fee by volume, messages and AI interactions, with published prices and no per-seat fee. The bill rises with volume, so long conversations and campaigns weigh more. No model suits everyone. A large team with low volume feels per-seat licenses more, and a lean team with high volume feels usage-based pricing more.

Do the math in three scenarios, with your own numbers: a normal month, a peak month and a doubled team. Ask the vendor what counts as a conversation, what going over the allowance costs, whether templates and AI are extra, whether consulting is mandatory, whether there is a lock-in, and how you take your data when you leave. Per-seat licenses have a side effect: they can push teams to share logins, and the record of who answered what disappears. There is a comparison with Octadesk, with sources and dates.

Test with real conversations before signing. Put two people on the system for a week, force a transfer, let a conversation get close to its deadline and ask for a report for the period. Check whether the history opens beside the conversation and how many clicks it takes to pass a case on. The CCX Messages shared inbox page shows how the queue works there, and the page for support teams shows the rest of the operation.

Where AI helps in customer service, and where it should not decide

AI helps with what is repetitive and has a written answer: order status, invoice copies, opening hours, return policy, product questions. It also helps agents by summarizing a long conversation, suggesting a reply from the knowledge base and tagging the contact reason, which feeds the weekly routine without a spreadsheet. And it covers the overnight hours, as long as it makes clear when a person will reply.

There are ground rules. Every agent introduces itself as AI at first contact, answers from the company's knowledge base and hands off to a person when unsure, when the customer asks or when the topic is sensitive. Article 50 of the EU AI Act makes that disclosure a legal duty in Europe, and Brazil's decree allows AI for triage, routing and resolution as long as the human option exists. The guide to the AI agent for WhatsApp shows how to set one up.

AI should not decide what touches money or rights: denying a refund, refusing a cancellation, ending a contract, declining credit, judging a serious complaint or a health case. In those cases, it gathers the facts and hands over to a person. Where GDPR Article 22 allows such automated decisions, it requires at least the right to obtain human intervention, and Article 20 of Brazil's LGPD gives people the right to request review of decisions based solely on automated processing.

The company answers for what the bot says. In 2024, in Moffatt v. Air Canada, a British Columbia tribunal held the airline liable because its chatbot said a bereavement fare could be claimed retroactively, which another page on its website contradicted. The airline argued the bot was effectively a separate entity, and the tribunal rejected that. It is a small-claims decision from Canada, but it is a fair warning: test your AI with hard questions before switching it on.

Measure AI the way you measure the team. Resolution rate without a person, read together with reopens and with the score from those it served. High resolution with high reopens is a bot closing conversations. At CCX Messages, the AI answers from the knowledge base and transfers to a person with the context of the conversation. Whatever the system, start with few topics and widen only when the numbers confirm it.