Why the chat bubble grew a face

The chat bubble in the corner of your website has been there long enough to feel like furniture. In that time it learned to answer instantly, in any language, at three in the morning. It never learned to look anyone in the eye. This is the story of why we built Bubblio around a face instead of a text box, and why the hard part was not the video.

The bubble stayed the same while everything behind it changed

When live chat first spread, the bubble usually connected you to a person. Then came canned replies, then decision trees, then large language models. Through all of it the interface never moved: the same circle in the corner, the same greeting, the same typing indicator. The costume stayed human while the substance became software.

That gap matters more than it looks. A typing indicator is a promise that someone is composing a thought for you. When the thing behind it is a deflection engine, customers feel the mismatch before they can name it. And they adapted.

Customers learned to type "human" first

Text bots were mostly built to deflect, because deflection was the metric their buyers paid for. The public numbers show the strain: vendors claim resolution rates around 76%, while real-world figures for text bots are reported to cluster between 42% and 53%. Customers never read those studies. They simply learned that the fastest path through a bot is to ask for a person in the first message.

The clearest tell comes from inside the category. Fin, the AI support agent that was Intercom until a May 2026 rename, is a serious product with real strengths: distribution, breadth of integrations, a brand support teams already know. Yet users report that Fin's own customer support is rated 3.2 out of 5 on Trustpilot, and that the failure cited most often is not being able to reach a human. Those strengths make the complaint more telling, not less: when reports say a product this established struggles to connect customers to people, the model itself is due for a rethink.

Years of deflection taught customers that the bubble is an obstacle. We wanted to build one that is the destination.

What changes when support has a face

Bubblio puts a realtime AI video person in the bubble. A visitor clicks and talks to it face-to-face, the way they would on a support call, except it answers in seconds and never puts them on hold. Three things change immediately.

• Presence. A face signals that someone is here now. Not a ticket number, not a queue position, but a conversation that has already started.
• Accountability. The agent answers from knowledge crawled off your own site, cites its sources, and refuses to guess when it does not know. A face makes hedging visible, so we made honesty the default behavior.
• Attention. Talking is faster than typing. A two-minute conversation replaces a form, a confirmation email, and a day of waiting.

The refusals are not dead ends. Every question the agent declines to answer lands in a "Customers are asking" feed, so you see exactly what your site fails to explain. Deflection hides that signal. A face that admits ignorance hands it to you.

Video only works if it never fails weirdly

Here is the honest objection to video support: video calls fail in more ways than text. Avatar providers have outages. Corporate networks block WebRTC. Visitors deny the microphone prompt. If any of those produce a dead bubble, the face costs you trust instead of earning it.

So most of our engineering went into failure, not features. Bubblio runs three avatar providers under the hood with automatic failover at connect time: if the first provider is down, the next takes over. If video cannot connect at all, because the mic was denied or the network will not carry it, the bubble degrades to plain text chat. The visitor always gets an answer; the only variable is the medium. The same rule applies to billing: when a prepaid minute pack runs out, the bubble degrades to text rather than going dark.

The human never left the loop

We did not build this to remove humans from support. We built it to stop making customers queue for them. Every Bubblio conversation has a path to a person: an async inbox with email follow-up when nobody is around, and live takeover when someone is. You can pause the AI mid-call and type as yourself. You can also puppeteer: your typed words are spoken by the avatar's voice, mid-video-call, so the customer keeps one continuous conversation while you steer it.

That inverts the industry's arrangement. The bot is not a wall in front of your team. It is your team's front door, and you can step through it the moment a conversation deserves you.

Receipts are the same philosophy, applied to the bill

Honesty about failure has a billing corollary. Fin's published price is $0.99 per resolution, and a conversation counts as resolved if the customer goes silent for 24 hours. Estimates place 25–45% of billed resolutions in that assumed, no-confirmation bucket, and users report bills jumping from $4k to $9k a month. We wrote up the mechanics in the assumed-resolutions post and the full cost teardown.

Our version is resolution receipts. Outcomes are measured from transcripts and shown per conversation; anything we cannot measure is counted as unmeasured, never guessed. The dashboard reads "X% resolved · measured on N", and N is right there. Pricing follows the same shape: prepaid minute packs at $49 for 100 minutes, $199 for 500, $599 for 2,000. No per-resolution fee, no overage (spend cannot exceed what you bought), no contract.

The wedge is open

Nobody in mainstream customer-support software offers realtime video support agents today. Fin's realtime modality is voice only. The enterprise agents, Sierra at a reported ~$15.8B valuation and Decagon at a reported ~$4.5B with contracts reported around $400k a year, are priced for enterprises, not for a founder pasting a snippet. Voice platforms like Retell (around $0.10 a minute all-in) and Parloa stay voice-only. Beyond Presence offers video-avatar chat, but without tool-calling, failover, or an analytics platform behind it. Meanwhile Salesforce agreed in June 2026 to acquire Fin for roughly $3.6B. The category is consolidating around text and voice while the video seat sits empty.

To be fair about who should stay put: if your support depends on a deep stack of helpdesk integrations, or your volume justifies an enterprise contract with a dedicated team behind it, Fin and the enterprise vendors are rational choices. The wedge we are taking is everyone else: teams who want support that feels like a person and is priced like a utility. Install is about five lines of code.

Bubblio is in private beta. No resolution tax. No overage. No contract. Live in 5 minutes. If you are curious what resolution-priced billing looks like at your volume, run the calculator; if you want a face in your bubble, request access.