Software that listens, speaks, sees and understands.
Voice AI is the technology that lets software hold a spoken conversation: understanding speech, reasoning about it, replying in a natural voice and, increasingly, appearing on video. It now runs inside phone calls, meetings and contact centres, alongside the people it supports.
Not long ago this meant speech recognition and a scripted bot. Large language models changed what is possible: agents that understand what a caller means, keep context, take actions in other systems and answer in a voice that sounds human, with delays short enough to hold a real conversation.
The same technology now works on the human side of the conversation too: coaching an agent while they talk, translating a call as it happens, and turning every conversation into searchable intelligence. That makes it a discipline in its own right, with its own engineering, governance and risks.
Then and now
If your picture of Voice AI is a few years old, this is what moved.
| It used to mean | It now means | |
|---|---|---|
| Self-service | Touch-tone menus and keyword bots | Customer voice and video agents that hold a natural conversation |
| Speech | Transcription after the call | Real-time understanding, speech-to-speech and live translation |
| Insight | A few calls a month, listened to by a person | Conversation intelligence across every interaction |
| Coaching | A review meeting a week later | Guidance during the conversation, and AI role play before it |
| Risk | A misheard menu option | Hallucination, impersonation, consent and regulation |
The landscape
AI agents that answer and make calls, understand intent, take actions in back-office systems and hand over to a person with the context intact.
AI agents with a face: real-time, lip-synced video for service, onboarding, training and kiosks, where seeing someone changes how the conversation goes.
Understanding every conversation at scale: topics, sentiment, intent, outcomes and risk, across calls, meetings and messages.
Guidance for the person in the conversation while it is happening: prompts, answers, tone and compliance nudges, with AI role play for practice beforehand.
Speech translated as people talk, in calls, meetings and service conversations, increasingly in the speaker's own voice.
The engines underneath: speech recognition, speech synthesis, speech-to-speech models, and the trade-offs between accuracy, latency and cost.
Keeping AI agents accurate, accountable and lawful: testing, observability, consent and disclosure, and defending against synthetic voices used for fraud.
Connecting AI to the telephony, contact centre and business systems it needs, and letting several agents work together on one task.
What is emerging
The developments the Voice AI programme is built around, and the ones every exhibitor will be asked about.
Models that listen and reply in audio directly, cutting the delay that made earlier voice bots feel robotic.
Voice agents gaining real-time video and the ability to see what a customer shows them on camera or screen.
Live translation moving from a meeting feature you switch on to a layer inside calls and service conversations.
Standards such as agent-to-agent protocols and MCP letting AI agents from different vendors pass work between them.
Cloned voices used against contact centres and finance teams, with detection and liveness checks developing in response.
Disclosure, consent and accountability rules for AI that talks to the public, and the audit trails to show what an agent actually said.
Voice AI sessions
Every PACKETLOSS session is Level 200 to 400. These show the depth, not the final programme, which members help shape.
Who should come
For the people designing, building, governing and buying AI that talks.
For exhibitors
Visitors to the Voice AI zone want to talk to your agent, not watch a video of it. Expect technical questions about latency, accuracy, handover, data handling, and how the agent behaves when a customer goes off script.
What works: live demonstrations, technical people on the stand, and a straight answer about which platforms you work with. Nothing at PACKETLOSS is Level 100, and that includes the stands.
The other disciplines
The phone system moved to the cloud. The job got bigger.
Explore →From answering contacts to resolving customers.
Explore →Where work happens, and increasingly where AI does too.
Explore →Communications as building blocks.
Explore →PACKETLOSS 2028 · Voice AI
Two days. Every platform. No beginner content.
Register your interest, put a session forward for the Voice AI programme, or ask about exhibiting.
The next PACKETLOSS is in 2028. The date, venue and exhibitor packages are announced once confirmed.