AI-ul vocal pentru întreprinderi nu are nevoie de mai multă personalitate. Are nevoie de empatie prin design

Lucrez în domeniul vocal din 2017, suficient de mult timp pentru a fi trecut printr-un ciclu de hype și pentru a vedea a doua undă câștigând o reală impuls. Deși acest al doilea val este tehnic mai impresionant, discuția în jurul lui s-a redus la cifre. Foarte puțin din aceasta se referă la modul în care oamenii vorbesc cu adevărat și aș spune că am pierdut conexiunea pasionată cu lingvistica și psihologia care caracteriza multe dintre fazele timpurii ale vocii.

Răsfoiți anunțurile și veți vedea valori ale latenței, rate de rezolvare și afirmații ambițioase despre ce pot face agenții. Toate acestea sunt importante, dar un agent cu voce frumos articulată este inutil dacă nu poate finaliza sarcina. În mod notabil, produsele vocale pentru întreprinderi se bazează din ce în ce mai mult pe același pool de modele de bază, ceea ce înseamnă că viteza este, de fapt, un element de bază.

Diferența se află într-un alt loc. Ce știe agentul? Ce poate face? Cum se simte interacțiunea în timp ce acționează? Cel mai important, clientul trebuie să depună efort suplimentar pentru că agentul este o mașină?

Aici începe designul vocal condus de empatie. Nu mă refer la adăugarea unei voci calde sau a unei fraze simpatice. Mă refer la proiectarea în jurul intenției apelantului, a încărcăturii cognitive și a stării emoționale probabile. Agentul ar trebui să știe ce știe deja afacerea, să explice ce face, să se recupereze curat când greșește și să înceteze să vorbească când sarcina este finalizată.

A warm synthetic voice that asks for the same account number three times has not made the call any easier and the warmth is not doing any work.

Empatia nu este o setare a vocii

A voice can sound thoughtful without the system behind it understanding the problem or being able to do anything useful. A customer may be calling about an outage, a disputed invoice, or an implementation blocker that cuts across several teams. If the agent only knows the FAQ, a warm or polished tone will not make the experience empathetic. The agent needs to know who is calling, what they have bought, what the last interaction on their account was, and what it is allowed to do to help the customer.

An enterprise voice agent needs the right customer and product context, with the right permissions, so it can answer and act while the person is still on the line. The caller does not need search results read aloud. They may need the refund issued, the record corrected, or the right specialist brought in.

I see empathy as both a design discipline and a systems requirement. Conversation design determines how the agent behaves. The underlying architecture determines whether it knows enough to be useful. I think of the combined result as the product texture of voice: what the system knows, how it uses that knowledge, how it handles uncertainty, and how the person feels when the conversation ends.

Vocea creează un contract social

Voice is not a neutral interface. I was on a call with a client recently and the demo agent had a female voice. Within a minute, people in the room were calling it “she.” Nobody announced that decision, it was just a natural extension of the conversation.

This effect predates generative AI. In 1994 research on computers as social actors, Clifford Nass, Jonathan Steuer and Ellen Tauber showed that people apply social responses to computers even when they know they are interacting with a machine. Voice intensifies that response through timing, confidence and tone.

That changes how failure feels. For example, a text box that returns the wrong link is irritating, and a voice that talks over you, misunderstands the same point twice, and insists it has solved the issue feels rude. The technology has not developed feelings. The caller has.

Trust is already thin. A 2025 global study led by the University of Melbourne with KPMG surveyed more than 48,000 people across 47 countries. It found that 66 percent use AI regularly, but only 46 percent are willing to trust it. Every voice interaction either earns a little trust or spends it.

There is also an ethical tension. The industry often treats the most human-sounding agent as the best one. But people do not want to be tricked. When I design an agent, I want it to identify itself, explain what it can access, and state what it is going to do before it does something consequential. Transparency must be part of the conversation.

Proiectează pentru ureche, nu pentru ochi

A screen is forgiving. People can scan, reread, and scroll back. Speech disappears as it is heard, which makes cognitive load a practical design constraint rather than an academic footnote.

Nelson Cowan’s review of working memory describes a central store limited to roughly three to five meaningful items in young adults. Yet voice systems still read out long menus and stack several questions into one turn. Most people cannot hold all of that in place, especially when they are distracted, driving, or already annoyed.

A text chatbot cannot simply be given text-to-speech and called a voice product. Native voice design uses shorter turns, puts important information first, and gives the caller room to interrupt. It also recognizes that not every gap in a conversation needs filling, so while speed matters, native voice design can “read” when a conversation is finished.

My background includes sociocultural linguistics, which looks at how language works between people and within social settings. That is useful in voice design because it gives us precise language for things like pacing, prosody, turn-taking and repair. “Make it sound natural” is not a design instruction. Being able to describe how a conversation should work gives teams something they can actually build against.

Ce arată designul condus de empatie

Empathy-led design is not about making an agent sound warmer or more human. It means shaping the conversation around the person’s situation, reducing the effort required to carry out a task, and handling failure without adding to their frustration. In practice, that starts with a few basic design choices.

Începe cu motivul pentru care a sunat persoana

People do not call to navigate a company’s taxonomy. They call because something is broken, late, or confusing. A caller who says, “I need to stop this delivery,” should not be pushed through a generic order-status script. The agent should establish which delivery and whether it can still intervene.

Folosește contextul pentru a elimina repetiția

Empathy is often most visible in what the system does not ask. If the business already knows who the customer is, which product they use, and what happened previously, the agent should use that approved context rather than making the caller reconstruct it.

The same applies inside the business. A salesperson between meetings should be able to ask what changed in an account, dictate the follow-up and have the update written back to the right system. The value is staying in motion without losing context or creating more administration.

Proiectează repararea înainte de calea fericită

Every voice system will mishear someone. Accents, background noise, unfamiliar names, and interruptions make that unavoidable. The product is defined by what happens next. The worst recovery pattern is to repeat the same question in the same words. A better response explains what the agent understood and narrows the uncertainty. If the issue requires judgment or access the agent does not have, it should involve a person before the caller has to demand one.

Respectă predarea și ieșirea

Handing a caller to a person is not a failure. Handing them over without the context of the conversation is. If the agent transfers the call without passing on what the customer said, what it identified, and what it has already tried, the customer has to start again. That is exactly the work AI should remove.

The same applies when the issue is resolved. Sometimes the best voice experience is one question, one accurate answer, and a click. Callers will not necessarily thank the agent or confirm that their query has been resolved before hanging up. If they have what they need, the interaction has worked.

Măsoară efortul, nu teatrul de performanță

A perfectly resolved one-shot call can look like a bounce because the customer hangs up without confirming satisfaction. If success means completing a long script, teams will optimize for longer conversations and call it engagement.

I would rather measure repeat calls, corrections, avoidable transfers, and whether the requested action was completed. Broader satisfaction scores like NPS still matter, but the conversation itself reveals where the customer had to work harder than they should have.

In februarie 2026, Gartner reported that 91 percent of customer service and support leaders were under executive pressure to implement AI. That pressure makes it even more important to measure whether automation is actually removing work. A voice agent that creates repeat calls, unnecessary transfers or longer interactions may look efficient on paper while simply moving cost elsewhere. The economics improve when the system resolves the right work, avoids unnecessary turns and keeps customers from returning through a more expensive channel.

Capacitatea și empatia merg mână în mână

The enterprise voice market is moving beyond natural speech. What matters now is whether an agent can understand the caller’s needs, access the right business context, act precisely and safely, and respond appropriately when the conversation goes off script.

Empathy-led design should not make an agent more talkative. Done properly, it removes repetition, shortens explanations and prevents the caller from having to manage the machine. Sometimes delight is simply the feeling that something difficult became easy.

The most important question I ask when building any AI product is brutally simple: why would anybody bother? Voice earns its place when it makes something noticeably easier, rather than giving people more work to finish.

Before launching a voice agent, listen to failed conversations, not just polished demos. Did it understand why the person called? Did it know enough to solve the problem? When it failed, did it make the next step easier or harder?

If the answer is unclear, more personality, lower latency, and a more human voice will not rescue it.