The phone tree is getting a face, and a memory

The phone tree only accepts the answers it expects

Traditional IVR systems route callers through scripted decision trees that only accept expected answers, and the failure shows up in the numbers. Tavus reports that 45 percent of customers in a 2024 Gartner survey said self-service systems failed to understand their intent. In a 2026 survey, 87 percent of customers said companies using generative AI for service must provide access to a human agent. m:…1439

Tavus's answer is the Personified Application Layer, or PAL, an agent that sees, hears, remembers and responds in real-time conversations. The PALs use a four-model loop of timing, perception, response and expression. Sparrow-2 is the timing model that waits through hesitant pauses, Raven-1 is the perception model that identifies mismatches between words and tone, and Phoenix-4.5 is the expression model that renders live facial expressions. Tavus reports that Sparrow-2 achieves 92.4 percent end-of-turn recall and 97.4 percent interruption recall in preliminary tests, figures that come from the company itself. Raven-1 audio perception operates under 100 milliseconds, and Phoenix-4.5 renders PAL faces with audio-to-video latency under 130 milliseconds. m:…1439 m:…1441

Why feeling seen matters

The case for video rests on what voice-only systems drop. Automated support systems often rely on transcript-only input, which ignores tone and expression, and fixed timing rules in voice agents often cut off customers mid-thought. People register timing shifts of 100 milliseconds or less, so a pause slightly too long, or an interruption slightly too early, is felt before it is understood. AI customer experience systems often fail to detect emotional cues like frustration or hesitation, and only 27 percent of customers would try a chatbot again after a negative experience. m:…1441

The early evidence for video is a case study, not a broad trial, and it should be read that way. Careflick's companion PALs for seniors found that 97 percent of users preferred video over voice-only interactions, and Careflick users had a 54 percent activity rate at day 30, with average daily engagement of 24 minutes. Salesforce developed Piper, an AI sales development representative, using Tavus's technology. Tavus offers a voice-only mode on its CVI stack for teams to validate AI performance before adding video, and HIPAA compliance is available for Tavus Enterprise plans. m:…1439 m:…1441

Most churning customers hear nothing first

Lorikeet's research puts a price on the silence that video agents are meant to fix. It reports that 87 percent of customers heard nothing from their provider before they decided to leave, and that the average customer who leaves a fintech, insurance or health provider takes about 826 dollars in revenue with them. Most customers do not leave in a moment: 72 percent take a week or longer to leave after first thinking about it, and 79 percent did something their provider could see before they left, such as reduced usage or fee inquiries. m:…0276

What works, according to the same research, is narrow. Fixing a problem was the only kind of contact with a higher stay rate than no contact, while sales pitches were linked to the lowest. Yet 56 percent of customers who contacted their provider about a problem had to do so more than once, and 30 percent had to explain their situation again each time they contacted their provider. Four in five customers, 84 percent, say companies only pay attention to them when they try to leave. This suggests the failure is not the channel but the memory: a customer who must repeat the problem to every new person has little reason to stay. m:…0276

The same agent, inside the building

The same ideas are showing up away from the contact center. Convai has launched an on-premise AI agent platform for training and on-the-job assistance, which can be deployed on-premise, in a private cloud, or in an air-gapped environment. It supports a one core agent architecture, so a single agent configuration can serve web apps, 3D simulations and physical assistance tools. Assembly Studio generates step-by-step training procedures from CAD models, and Vision Assistant provides real-time guidance for physical work using manuals and documents. It integrates with existing enterprise language models via OpenAI-compatible endpoints, uses MCP to connect agents to internal tools and data, and includes Mindview, which traces the full prompt behind any agent response. m:…6304

Convai is SOC 2 compliant, available through AWS Marketplace and Azure Marketplace, and its NeuroSync provides real-time facial animation and lip sync for avatars. The deployment choice is the enterprise answer to the same trust question the service agents raise: the agent can see, remember and speak, but the data stays on the company's own servers or in its own cloud. Tavus's video agents in service and Convai's on-premise agents in training point in the same direction, toward agents that notice timing and tone, remember what happened, and keep a human within reach. m:…6304

What to watch

The open questions are practical. The Tavus accounts flag the long-term scalability of real-time video AI agents in high-volume support environments as unresolved, and the cost comparison between video agents and traditional IVR has not been provided. The Careflick preference figures come from a single case study, and the Sparrow-2 recall figures are preliminary tests reported by Tavus. AI voice agents improve operations by routing calls based on intent, reducing steps to resolution and providing context to human agents, which suggests the near-term win may be the handoff rather than the face: an agent that listens properly, remembers the problem, and gives the human a complete brief. The phone tree is being replaced, but the test that matters is whether the customer still has to repeat the problem. m:…1439 m:…1441