Who's talking when your agent talks?

Every assistant has a personality, even if nobody designed it. If it interrupts, it seems rude. If it doesn’t remember what you just said, it seems scatterbrained. If it answers everything with the same enthusiasm, it seems fake. We attribute human traits and qualities to anything that talks to us, and we do it in seconds.

Clifford Nass and Scott Brave studied this for years at Stanford: we react to machine voices the way we react to people, with the same social expectations (Nass and Brave, 2005). Your agent is going to have a personality either way. What you can decide is which one.

1. Start with the relationship

Before the adjectives, decide what role the agent plays toward the person talking to it. An expert giving advice, an assistant carrying out tasks and a companion in a leisure activity are not the same thing. The role shapes almost everything else: how formal it is, how much it takes the initiative, how much it allows itself to joke.

When we created Victoria, the synthetic voice for sports news, with PRISA, we started with a workshop with the AS and Carrusel Deportivo teams to define who Victoria was before thinking about how she sounded.

2. Pick a few traits, and their limits

Three or four traits are enough. The useful part is pairing each one with what it must not become: friendly, but not overfamiliar; with a sense of humor, but no jokes when someone has a problem; confident, but able to admit it doesn’t know something. That second half is what keeps it from becoming a caricature.

3. Decide how it speaks

Personality shows in concrete decisions:

Formal or informal address (in Spanish, tú or usted).
Short sentences or long explanations.
Technical or everyday vocabulary.
Whether it uses colloquial or regional expressions.
Whether it makes small talk or gets straight to the point.

Write them down with examples. “Friendly” means different things to each person on the team; a sample answer doesn’t.

4. Define the limits

A personality is also defined by what it doesn’t do. How it responds when it doesn’t know something. How it says no to a request it can’t handle. How it reacts to an insult. Which topics it avoids. And how it makes clear it’s an AI, something the EU AI Act has required since August 2026 of systems that converse with people, unless it’s obvious (Regulation (EU) 2024/1689, Article 50).

These hard moments say more about a personality than a hundred friendly answers.

5. Make the voice tell the same story

If the agent speaks, the voice is the part of its personality people perceive first. Timbre, speed, intonation and pauses convey character before a single word is understood. Phonetician Oliver Niebuhr, at the University of Southern Denmark, has spent years measuring which features of intonation make a voice sound charismatic, and has applied them to synthetic voices to see how they change listeners’ reactions (SDU).

What matters is consistency. A friendly, casual text read in a solemn voice sounds odd, and vice versa. Nass and Brave found that people penalize these contradictions between what’s said and how it’s said. Microsoft, in its guide to creating custom synthetic voices, recommends exactly that: first define the voice’s persona, its character and emotional tone, and keep it consistent across all recordings (Microsoft Learn).

6. Choose gender and accent on purpose

The default voice isn’t neutral. For years, almost every assistant had a female voice by default, and in 2019 UNESCO warned that this reinforced stereotypes, as we explain in an assistant repeats the prejudices in its data. There are alternatives: Q, a voice designed to be perceived as neither male nor female, was tested with more than 4,600 people in Europe (Dezeen, 2019). Accent communicates too: a voice with an Andalusian, Canarian or Mexican accent says something different from one with a standard accent. At Monoceros we’ve worked on both fronts. With LLYC we created Free the Voices, a bank of synthetic voices that don’t conform to the gender binary, made from the voices of LGBTQ+ people. And in Fonos there are voices with accents other than the standard one, such as Izar, with an Andalusian accent, or Libertas, one of the Free the Voices voices, with a Mexican accent. There’s no right answer, but it should be a decision, not a box nobody touched.

7. Adapt the tone without losing the identity

The same personality can speak differently depending on who it’s talking to and the situation. In our evaluation of a real estate portal’s two help chatbots, we defined a common base for both, but with different degrees of warmth: the one serving professionals who come back often could be more familiar than the one handling one-off questions (real estate portal case study). The same goes within a single conversation: the tone isn’t the same when sharing good news as when handling a complaint.

8. Write it down and test it

All of the above fits in a short document: the role, the traits and their limits, how it speaks, what it doesn’t do, how it sounds, and sample answers for the most common situations and the hardest ones. With a language model, that document becomes a large part of its instructions.

And you have to test it. Language models tend to drift back to their default tone in long conversations or in situations the instructions don’t cover. Review real conversations looking for those shifts, and ask the people who use the agent how they’d describe it. If their words resemble the ones in your document, the personality works.

Designing the personality of assistants, agents and voices is a central part of our work in conversation and agent design and custom synthetic voices.