Photo source: Interactive Powers

Video, the Last Frontier of Conversation

If AI can already read and speak, why aren’t artificial agents routinely serving customers by video? It isn’t a camera problem. Nor a bandwidth problem.

AI can already see. The hard part is holding a visual conversation in real time, understanding the context and responding the way a human interaction demands.

There are three reasons.

1. Seeing is not the same as understanding

An LLM processes text extraordinarily well. New voice models can already listen and respond with latencies close to those of a natural conversation.

But in video there is much more than words.

A face, a gesture, a pair of hands, a document, a product. Everything changes while we talk. A scene like that moves around 40,000 times more data than text.¹

Human conversation doesn’t wait for each message to finish. We observe, listen, react and correct on the fly — with turns of around 200 ms between speakers.²

It isn’t enough to recognise what appears on screen. You have to understand what is happening while it is happening.

2. It isn’t just a question of power

We can add more computing capacity. But that doesn’t guarantee an AI will understand a conversation better.

The challenge is integrating image, voice, context and response without breaking the rhythm of the interaction.

It already happens with voice. To answer in real time, native voice models think less: in a mathematical reasoning test, OpenAI’s text model solves 75% of the problems; its real-time voice version, 6%.³

Multimodal research is moving fast. Even so, interpreting an image or answering questions about a video is not the same as sustaining a visual conversation with the reliability a real service requires.

The frontier isn’t that AI can see. It’s that it can converse while seeing.

3. Conversation is learned by interacting

A child doesn’t learn to communicate by watching sped-up videos. They learn by looking, acting and getting responses from other people.

Context matters. Experience matters. And so does training data: the largest dataset of face-to-face conversation gathers around 4,000 hours.⁴ A four-year-old has lived around 16,000 hours awake.⁵

In customer service, images can also contain personal data that must be handled with safeguards.

It isn’t enough to pile up videos. AI has to learn from diverse, contextualised and relevant interactions.

And in the meantime?

Here’s what matters for any company: we don’t need to wait for AI to do everything in order to benefit from video.

• ✅ AI can transcribe, summarise, retrieve information, extract data and assist the agent.
• ✅ The person can see, understand the context, bring judgement and resolve what requires human intervention.

It isn’t AI or Human Intelligence (HI). It’s AI + HI, connected in the same conversation.

The goal isn’t to put an AI in front of a camera. It’s to make sure every interaction is resolved with the right technology — and that a person can step in when it truly adds value.

Because the future of customer service isn’t about automating every conversation. It’s about never missing a chance to understand and resolve.

🤔 Is your customer service ready to see what the customer needs to show — or does it still depend on what they manage to explain over the phone?

¹ Typical WebRTC rates: text ~40 bit/s, Opus voice ~32 kbit/s, 720p video ~1.5 Mbit/s. ² Stivers et al., PNAS 2009. ³ Voice Evaluation of Reasoning Ability (VERA), ICLR 2026: GPT-5 text 74.8% vs GPT-Realtime voice 6.1%, on competition math problems. ⁴ Meta FAIR 2025, Seamless Interaction: 4,000+ hours of face-to-face interaction. ⁵ Yann LeCun, Lex Fridman Podcast #416.

Interactive Powers - Streamline your business communications