Tone of voice by situation: what to write when a customer is angry

Four adjectives describing your brand are the most useless thing in the world at 9pm, when a customer has just written "where is my order, you said 3 days and it has been 6" and the agent on chat duty has about 30 seconds to type something. "Friendly, professional, warm, trustworthy" does not tell that person what to write. What does help is a situation matrix with actual wording in it.

Why do brand voice documents so often go unused?

Most of the brand voice material we are handed at the start of a project stops at the descriptive layer: 4 adjectives, a paragraph about what the company believes in, sometimes a pair of opposites like "warm but never over familiar". That layer is fine for choosing a direction, but it is not a working tool. Nobody on chat duty can translate "professional" into a concrete sentence in 30 seconds while a conversation is heating up. Everyone translates it their own way, and the brand ends up sounding different depending on who is on shift.

The gap sits between the adjective and the sentence. Closing it does not need more theory, only a table of real situations with real sample lines. That is where this differs from building a brand voice for AI agents at the strategic level: that piece answers what the voice is and how to get it into the machine, while this one answers a narrower and more urgent question, which is what to type into the chat box right now.

What goes into one row of the matrix?

Each situation should be described in exactly five cells, no more. Cell one is the situation, written in the customer's language rather than internal jargon: "customer wants to buy something we are out of" instead of "OOS handling". Cell two is the emotion the customer is carrying, because the same event with a different emotion needs a different reply: a disappointed customer needs reassurance, an annoyed one needs action. Cell three is the goal of the reply, and you pick only one: save the sale, cool things down, or protect the policy. Cells four and five are the line to send and the line to avoid, written out word for word, with one short line of reasoning.

That reasoning line is the part teams tend to drop, and it is the part that decides whether the matrix survives. Someone who understands the reason will improvise correctly when a situation appears that is not in the table. Someone who only memorised the sample line will paste it into the wrong context.

A condensed view of the first 5 rows. Your internal version should spell the sample lines out in full instead of summarising them like this.

Five hard situations and the exact wording

Out of stock. The customer is deflated because they had already decided to buy. The goal is to keep that intention alive, not to explain why the warehouse is empty. Say this: "That style is sold out in your size right now, and the next batch lands on the 22nd. I can put one aside for you, or the beige version is still in every size, same fabric and same price." Avoid: "It sells out so fast, sorry!" The reason is that the avoid line turns your own supply problem into a humblebrag and leaves the customer with no next step at all.

Price pushback. The customer is questioning the value, not necessarily their own budget, so stay away from anything that touches affordability. The goal is to make clear what they are paying for. Say this: "I understand, it does sit above the usual range. That price covers installation at your place and a 24 month replacement warranty. If you do not need the installation part, there is a lighter package. Want me to send a quick comparison?" Avoid: "You get what you pay for" and "our products are for customers who care about quality". The reason is that both lines imply the customer is either wrong or not the right kind of buyer, which is the fastest way to lose the sale.

Late delivery. The customer is annoyed and has stopped trusting the last promise they were given. The goal is to restore the feeling that someone is accountable. Say this: "Your order is 3 days past the date we promised, and that is on us. I have just checked it, the parcel is at our east depot and goes out for delivery today. If it is not with you by 5pm, reply here and I will sort it out straight away and refund the shipping." Avoid: "The courier is overloaded this season." The reason is that people buy from the brand, not from the courier, so a line of blame solves nothing and quietly tells the customer this is not our problem.

Refusing a request. The customer is hopeful and will be let down by a no. The goal is to keep the relationship while keeping the policy. Say this: "I am not able to discount that any further, unfortunately. What I can do is include the 6 month servicing package and get you a priority installation slot this week." Avoid: "That is our policy." The reason is that quoting policy closes the conversation, whereas the "cannot do A, can do B" pattern leaves the customer somewhere to go. If there genuinely is no B, say so plainly and briefly rather than talking around it.

A public complaint. There are two audiences here: the person complaining and the few hundred people scrolling past. The goal is to show the second group how you handle things, not to win the argument with the first. Say this, publicly and short: "You are right, we sent the wrong colour and that is our mistake. I have messaged you directly to arrange the exchange today and we are covering the return." Avoid: "Please DM us" on its own, or a long paragraph explaining your internal process. The reason is that pushing straight to DMs looks like hiding, while the long process explanation turns the post into a debate stage.

How do you shift register per channel without losing the personality?

The matrix tells you what to say. The channel decides how long and how formal it gets. The rule is that the content and the attitude stay identical, and only the length, the rhythm and the level of formality move. Chat and messaging replies belong in 2 to 3 sentences with the action in front of the explanation, because the customer is reading on a small screen and waiting. Email allows the full version: one line taking responsibility, a short paragraph on the cause at a level the customer cares about, a paragraph on the fix and the timing, and a sign off from a real person with a name.

A live stream or a video reply is considerably more casual, but one thing must not move: how directly you admit fault and how much you commit to. A simple check is to read all three versions of the same situation out loud. If they sound like three different people, you have gone too far. If they sound like one person speaking in three different rooms, you have it right.

How do you turn this matrix into instructions for an AI agent?

A situation matrix is one of the rare documents that humans and machines can share almost unchanged. For an AI agent, each row becomes a block of instruction: the situation description so the agent can recognise it, the say-this line as a worked example, the avoid-this line as a hard prohibition, and the reasoning line so the agent can handle variants that were never listed. The only real addition is that a machine has to be told explicitly when to stop and hand over to a person: when compensation is demanded, when there is a legal angle, or when the customer raises the same complaint a second time.

A good instruction block looks like this: "When a customer reports a late order, confirm how many days late it is from the order record, take responsibility in one short sentence, state the current status and one specific time, and ask whether they want the shipping refunded. If they ask for compensation beyond the shipping fee, hand over to a human." The common failure looks like this: "Be empathetic with the customer and handle it professionally." The second version contains nothing the machine can execute, and what comes back is the polite empty paragraph that every reader immediately recognises as machine written. This is the part we build with clients when we deploy AI agents for business, and on larger engagements it sits inside our brand and customer experience services.

You do not need 20 situations to start. Take the 5 hard cases in this article, open the last 30 days of chat history, copy out the replies that were actually sent, and sort them yourself into say-this and avoid-this. A first matrix written in a single afternoon is enough to be useful, and the rest accumulates over time. If you want to go deeper for your own industry, see the profile of Trinh, Brand Voice Designer.

FAQ - Frequently asked questions