By Jim Kent

Would you use your AI assistant more if it had an attractive face, that you controlled and remembered everything that you told it? For most organisations, artificial intelligence has so far been a largely text-based experience. Large language models sit behind chat windows, customer service bots, internal copilots or enterprise search tools, answering questions and generating content at scale. Increasingly, however, a new layer is being added to this interaction, a face.

One of the fastest-growing areas in AI is the ability to attach photo-realistic human faces to chatbots and digital assistants. In most cases, the underlying intelligence remains unchanged. These systems are still powered by the same large language models, often accessed via API or hosted internally, provided governance and data-handling rules allow it. What changes is not the model’s reasoning capability, but how users perceive and engage with it.

Humans are instinctively responsive to faces, and this is not a new insight. Some of the most advanced experimentation with realistic digital humans has taken place outside the corporate world. For years, the adult entertainment and companion-AI sectors have acted as early laboratories for embodied AI. While controversial, they have demonstrated how visual presence, memory and emotional pacing dramatically increase engagement, even when the intelligence itself remains unchanged.

So how do these systems actually work?

At the core sits an existing large language model, accessed via API or self-hosted. On top of this is a carefully engineered system prompt defining personality, tone, boundaries and behavioural rules. Much of what users perceive as “character” is the result of prompt engineering rather than intelligence.

Memory is where the experience starts to feel convincing. Without it, the illusion quickly collapses. Most platforms therefore add a dedicated memory layer, typically using vector databases. This combines short-term conversational context with longer-term memory of preferences, names and past events, allowing the system to reference earlier interactions and appear consistent over time.

Emotional behaviour is orchestrated, not understood. State machines, heuristic rules and lightweight agent logic are used to pace responses, mirror emotional tone and decide when to express empathy or restraint. This creates the appearance of emotional awareness without any genuine understanding.

Around all of this sits a layer of safeguards. Content filters and behavioural constraints are applied to prevent illegal activity, policy breaches or harmful interactions, driven as much by legal exposure as by ethics.

Finally, a visual layer is added on top, a face avatar that speaks, listens and reacts.

This is where regulation becomes critical. Under the EU AI Act, systems that interact directly with humans are subject to transparency obligations, particularly where there is a risk of deception or emotional manipulation. An AI system presenting itself with a human-like face or voice must clearly disclose that it is artificial. In sensitive contexts such as customer service, healthcare, education or finance, additional requirements around risk management, human oversight and accountability may apply.

Crucially, the Act does not regulate intelligence alone, but impact. A face does not make an AI system more capable, but it can make it more persuasive. For organisations, this shifts responsibility from model selection to design choices. How human-like should an AI be? Where is the line between assistance and influence?

The next generation of face avatars is already emerging. London-based start-up Synthesia recently raised $200 million at a valuation of $4 billion, reflecting growing confidence that visual interfaces will become a standard way of interacting with AI. Users can already generate realistic avatars from a single photograph, including versions of themselves.

Putting a face on an LLM does not make it smarter. It makes it more human-seeming. For organisations, the challenge is to use that power responsibly, improving access and engagement without blurring the line between simulation and understanding, and without underestimating the regulatory implications that come with it.