Chatbot branding: tone of communication and visual style
Bot archetype: not based on a designer's taste, but on the type of risk in the industry
Choosing a character for a bot is not a question of "what we like," but a calculation of what emotion the client feels at the moment of contact. For fintech, it is anxiety about money, for e-commerce—impatience and desire for quick results, for B2B—the need for a competent interlocutor who does not waste time.
| Industry | Suitable archetype | Character traits | What to avoid |
|---|---|---|---|
| Fintech | Guardian / Sage | Calmness, precision of formulations, minimum humor in financial matters | Familiarity, emojis in messages about money, jokes about risks |
| E-commerce | Friend / Cool Guy | Vivacity, speed of response, light irony outside of controversial situations | Excessive formality, bureaucratic language, long greetings |
| B2B | Expert / Ruler | Structure, references to data, respectful distance | Flirtation, emotional exclamations, colloquial slang |
An important caveat: an archetype sets the basic tone, not the only acceptable one. A bot of a fintech company can be warm during onboarding and extremely dry when discussing overdue payments—and this is correct if the transition between modes is pre-defined.
Static tone versus dynamic adaptation: why the first approach is outdated
Most tone of voice guides describe one set of rules for all situations. This works for simple FAQ bots, but fails in real conversations, where the client's emotion changes throughout the dialogue. A mismatch between style and topic—for example, a friendly joke in response to a complaint—reduces the bot's perceived competence. The number is large, and it explains why, after one failed response, a client loses trust in the entire bot, even if the other answers were accurate.
Working version — a module for sentiment analysis of messages (Sentiment Analysis), which switches the bot between pre-prepared 'personalities' within a single dialogue:
- anger or strong irritation in the message — the 'Empathetic Manager' branch is activated: shorter phrases, acknowledgment of the problem, without attempts to sell anything additionally;
- questions about tariffs, specifications, deadlines — the 'Clear Salesperson' branch: numbers, specifics, minimum rhetoric;
- neutral small talk — the basic brand tone without changes.
A similar principle of long-term memory is implemented in Yandex's AI characters mode — the bot is capable of holding long dialogues and remembering previous topics, which makes tone switching more natural: the system understands that the client has already written irritably five minutes ago, and does not reset the context to neutral.
Abstract brief versus systematic prompt: where the tone of voice 'floats'
A classic mistake — describing the bot's character with adjectives («friendly, but professional, with a touch of humor») and conveying this to developers without further formalization. The neural network will maintain the set tone for the first ten dialogues, but on the eleventh complex request it will start 'composing' — because abstract traits do not give it firm boundaries.
Formalization is based on five elements, and the omission of even one element guarantees a style drift:
- Role and context — who is speaking and in what situation (not 'bank bot', but 'support staff who sees the client's payment history').
- Tone criteria — specific parameters: phrase length, emoji allowance, formality level on the scale.
- Hard restrictions — what you can never say: promises, legal formulations, diagnoses.
- Formatting rules — lists, paragraphs, greetings, signature.
- Examples matrix — pairs of 'bad answer / good answer' on real, not fabricated cases.
Without the examples matrix, the neural network relies on general perceptions of 'friendliness' from the training data, not on the brand's specifics — and ultimately all bots on the same technology sound alike.
Mascot as a trust amplifier: when anthropomorphism works, and when it hinders
A high level of anthropomorphism in an avatar increases trust in the bot — this is a confirmed effect. But it has a reverse side: excessive 'humanity' of the character when discussing sensitive topics (health, debts, personal data) increases the client's anxiety about social evaluation and reduces their willingness to self-disclose. People are more willing to admit to a problem to a faceless interface than to a character with a face and name, if the topic is shameful.
A telling example of a mascot successfully integrated into multi-channel communication is SberKot: over 8 years, the character has gathered 13 million active subscribers in the chatbot and 4 million in a separate community, stickers with him have been sent over 3,5 billion times, and about 80 million Russians know him. In the year of launching the card with the cat's image, every second bank customer decorated their card — the mascot works not only in dialogue but also as a standalone marketing asset on souvenirs, merchandise, and cards.
A separate issue is the gender and voice of the assistant. Female voices are often chosen by default in voice assistants (Amazon Alexa, Google Assistant, Apple Siri), and studies show that they are perceived as more responsive and empathetic, although this may be linked to gender stereotypes rather than universal preference. This is not a universal rule for any brand, but for support and care scenarios, a female voice or feminine traits of the character statistically perform better than masculine ones.
Transition to a live operator: where the style is most likely to break
The moment of transferring the dialogue from the bot to the human is the weakest point in most implementations. The bot was polite and concise, while the operator writes long paragraphs with bureaucratic language — the client physically feels like they have entered another service. The tonal gap at this point devalues all the previous work on the persona.
What is needed for a seamless transition:
- an explicit phrase about changing the interlocutor — 'I am a real specialist' or a similar phrase, without trying to pretend to be a continuation of the bot;
- retaining the same level of formality and phrase length that the bot used;
- passing the dialogue context to the operator so that he does not re-ask what the client has already explained to the bot — this is felt as a gap not so much stylistic as logical;
- single visual layer: if a bot had an avatar and brand colors in the chat interface, the operator should not be presented with a 'state' window without brand identifiers.
Typical mistakes in tone of voice for neural networks
The error is not in the absence of a brief — almost all companies have one. The error is that the brief is written for people, not for the model.
- Adjectives instead of rules. “Warm and human” tells nothing to a neural network about phrase length and acceptable topics for humor — measurable criteria are needed.
- One tone for all scenarios. Without a switching module (the so-called Router based on Sentiment Analysis), the bot responds equally enthusiastically to both a question about a promotion and a complaint — losing competence by 25–30%, as mentioned earlier.
- Lack of strict prohibitions. Without a list of what cannot be said, the model begins to 'fill in' the answer in the brand's spirit, but not based on facts, on a complex query.
- Ignoring the handover point to the operator. The brief often ends where the human begins, creating a gap described above.
- Testing only on 'convenient' dialogues. The persona is usually tested on standard scenarios, but not on conflictual or borderline topics — precisely there it often spreads out.
Calibrating the persona on real audience — not a one-time exercise before launch, but an ongoing process: collected logs of dialogues with low ratings must be regularly run through a matrix of examples and updated, otherwise the bot will gradually start sounding like it did a year ago, while the brand will not.
Bot archetype: not based on a designer's taste, but on the type of risk in the industry
Choosing a character for a bot is not a question of "what we like," but a calculation of what emotion the client feels at the moment of contact. For fintech, it is anxiety about money, for e-commerce—impatience and desire for quick results, for B2B—the need for a competent interlocutor who does not waste time.
| Industry | Suitable archetype | Character traits | What to avoid |
|---|---|---|---|
| Fintech | Guardian / Sage | Calmness, precision of formulations, minimum humor in financial matters | Familiarity, emojis in messages about money, jokes about risks |
| E-commerce | Friend / Cool Guy | Vivacity, speed of response, light irony outside of controversial situations | Excessive formality, bureaucratic language, long greetings |
| B2B | Expert / Ruler | Structure, references to data, respectful distance | Flirtation, emotional exclamations, colloquial slang |
An important caveat: an archetype sets the basic tone, not the only acceptable one. A bot of a fintech company can be warm during onboarding and extremely dry when discussing overdue payments—and this is correct if the transition between modes is pre-defined.
Static tone versus dynamic adaptation: why the first approach is outdated
Most tone of voice guides describe one set of rules for all situations. This works for simple FAQ bots, but fails in real conversations, where the client's emotion changes throughout the dialogue. A mismatch between style and topic—for example, a friendly joke in response to a complaint—reduces the bot's perceived competence. The number is large, and it explains why, after one failed response, a client loses trust in the entire bot, even if the other answers were accurate.
Working version — a module for sentiment analysis of messages (Sentiment Analysis), which switches the bot between pre-prepared 'personalities' within a single dialogue:
- anger or strong irritation in the message — the 'Empathetic Manager' branch is activated: shorter phrases, acknowledgment of the problem, without attempts to sell anything additionally;
- questions about tariffs, specifications, deadlines — the 'Clear Salesperson' branch: numbers, specifics, minimum rhetoric;
- neutral small talk — the basic brand tone without changes.
A similar principle of long-term memory is implemented in Yandex's AI characters mode — the bot is capable of holding long dialogues and remembering previous topics, which makes tone switching more natural: the system understands that the client has already written irritably five minutes ago, and does not reset the context to neutral.
Abstract brief versus systematic prompt: where the tone of voice 'floats'
A classic mistake — describing the bot's character with adjectives («friendly, but professional, with a touch of humor») and conveying this to developers without further formalization. The neural network will maintain the set tone for the first ten dialogues, but on the eleventh complex request it will start 'composing' — because abstract traits do not give it firm boundaries.
Formalization is based on five elements, and the omission of even one element guarantees a style drift:
- Role and context — who is speaking and in what situation (not 'bank bot', but 'support staff who sees the client's payment history').
- Tone criteria — specific parameters: phrase length, emoji allowance, formality level on the scale.
- Hard restrictions — what you can never say: promises, legal formulations, diagnoses.
- Formatting rules — lists, paragraphs, greetings, signature.
- Examples matrix — pairs of 'bad answer / good answer' on real, not fabricated cases.
Without the examples matrix, the neural network relies on general perceptions of 'friendliness' from the training data, not on the brand's specifics — and ultimately all bots on the same technology sound alike.
Mascot as a trust amplifier: when anthropomorphism works, and when it hinders
A high level of anthropomorphism in an avatar increases trust in the bot — this is a confirmed effect. But it has a reverse side: excessive 'humanity' of the character when discussing sensitive topics (health, debts, personal data) increases the client's anxiety about social evaluation and reduces their willingness to self-disclose. People are more willing to admit to a problem to a faceless interface than to a character with a face and name, if the topic is shameful.
A telling example of a mascot successfully integrated into multi-channel communication is SberKot: over 8 years, the character has gathered 13 million active subscribers in the chatbot and 4 million in a separate community, stickers with him have been sent over 3,5 billion times, and about 80 million Russians know him. In the year of launching the card with the cat's image, every second bank customer decorated their card — the mascot works not only in dialogue but also as a standalone marketing asset on souvenirs, merchandise, and cards.
A separate issue is the gender and voice of the assistant. Female voices are often chosen by default in voice assistants (Amazon Alexa, Google Assistant, Apple Siri), and studies show that they are perceived as more responsive and empathetic, although this may be linked to gender stereotypes rather than universal preference. This is not a universal rule for any brand, but for support and care scenarios, a female voice or feminine traits of the character statistically perform better than masculine ones.
Transition to a live operator: where the style is most likely to break
The moment of transferring the dialogue from the bot to the human is the weakest point in most implementations. The bot was polite and concise, while the operator writes long paragraphs with bureaucratic language — the client physically feels like they have entered another service. The tonal gap at this point devalues all the previous work on the persona.
What is needed for a seamless transition:
- an explicit phrase about changing the interlocutor — 'I am a real specialist' or a similar phrase, without trying to pretend to be a continuation of the bot;
- retaining the same level of formality and phrase length that the bot used;
- passing the dialogue context to the operator so that he does not re-ask what the client has already explained to the bot — this is felt as a gap not so much stylistic as logical;
- single visual layer: if a bot had an avatar and brand colors in the chat interface, the operator should not be presented with a 'state' window without brand identifiers.
Typical mistakes in tone of voice for neural networks
The error is not in the absence of a brief — almost all companies have one. The error is that the brief is written for people, not for the model.
- Adjectives instead of rules. “Warm and human” tells nothing to a neural network about phrase length and acceptable topics for humor — measurable criteria are needed.
- One tone for all scenarios. Without a switching module (the so-called Router based on Sentiment Analysis), the bot responds equally enthusiastically to both a question about a promotion and a complaint — losing competence by 25–30%, as mentioned earlier.
- Lack of strict prohibitions. Without a list of what cannot be said, the model begins to 'fill in' the answer in the brand's spirit, but not based on facts, on a complex query.
- Ignoring the handover point to the operator. The brief often ends where the human begins, creating a gap described above.
- Testing only on 'convenient' dialogues. The persona is usually tested on standard scenarios, but not on conflictual or borderline topics — precisely there it often spreads out.
Calibrating the persona on real audience — not a one-time exercise before launch, but an ongoing process: collected logs of dialogues with low ratings must be regularly run through a matrix of examples and updated, otherwise the bot will gradually start sounding like it did a year ago, while the brand will not.
Recent articles
MegaFon, BMW, VTB and more 240+ companies. See how it looks in the showcase reel:
Our branding agency is approached to order:
Branding agency services
