Synthesia Spent a Decade Recording Avatars. Now It Wants Them to Talk Back.

Synthesia built its business on a simple loop. You write a script, pick an avatar, and get a presenter video back. The company says that loop has run through the largest enterprise video platform in the world for the better part of a decade, and it got Synthesia to a reported $4 billion valuation with more than $100 million in annual recurring revenue by 2025.

On October 1, 2026, the loop changed direction. Synthesia launched Sessions, a platform for two-way conversations with a lifelike avatar. Two session types went live the same day, with a third announced as coming soon.

Roleplay Sessions puts the user in a conversation that is supposed to be difficult. The avatar plays a customer, a candidate, or a report, and it pushes back. Sales teams practice discovery and objection handling. Managers rehearse the conversation they have been avoiding with an employee. Support teams simulate an escalation. After each attempt, the system coaches and scores the participant. New capabilities let people adjust a scenario to their exact situation before practicing, and let companies build custom roleplays for their own recurring conversations.

Survey Sessions turns a form into an interview. The avatar asks the company's questions one to one, follows up on what it hears, and hands back structured insight. Synthesia lists employee feedback, customer insight, exit and churn interviews, NPS follow-up, and product research as the intended uses. Responses come back as themes and sentiment across respondents, with any individual answer openable as a transcript, an audio file, or a CSV.

Expert Sessions is the one still on the roadmap: an avatar trained on a company's product documentation and internal processes that walks someone through a task step by step and answers questions as they work.

The argument for this shape of product

Synthesia's CTO Peter Hill wrote the launch post. His framing is a staffing bottleneck rather than a technology story. There is never enough of the right person to go around, whether that person is a manager coaching every rep before a hard call, a researcher interviewing every customer, or a product expert guiding every user who gets stuck. The usual substitutes are recordings instead of rehearsals, forms instead of interviews, and help documentation instead of hands-on guidance.

Hill's biography makes the argument easier to take seriously than the average launch post. He spent nearly 25 years at Amazon and AWS, leading teams behind Kindle, Fire, Alexa, and services including Amazon Connect and WorkSpaces, before a stint as CEO and CPO at Wildlife Studios. He has shipped conversational systems at scale before.

The early data Synthesia is willing to publish is thin but pointed. Across early deployments, 78 percent of learners attempted a roleplay more than once, and of those who came back, 83 percent improved their score on a later attempt. Read that carefully. The number that matters is not the improvement rate, which is measured on the subset that chose to return. It is the 78 percent return rate. A training tool that people voluntarily repeat is a different product from one they complete because someone assigned it.

Synthesia also says thematic insights are meant to show which skills drive performance rather than only who scored highest, and that each employee gets suggested roleplays to close their own gaps. That is a shift from reporting to diagnosis, and it is the part of the pitch that learning and development teams will actually evaluate.

Two routes, and why the split matters

Synthesia is selling the same capability two ways. The Interactive Avatar API is for companies that want to build their own avatar product and control the stack. Sessions is the ready-made platform for the conversations that make up ordinary work. The company describes the API as a hosted pipeline, meaning a team can bring its own language model, speech-to-text, and text-to-speech rather than running the rendering stack itself.

That split is more interesting than it looks. Businesses that already have a preferred model provider do not want to be told which one to use, and businesses that have neither the appetite nor the headcount for integration work do not want to build anything. Offering both keeps Synthesia from having to pick a single customer profile. Session-level integration opens the full-stack API on the same day, which suggests the company sees the API as the quieter, larger business over time.

Compliance details are worth noting because this product touches employee conversations and customer interviews, which means recordings, transcripts, and personal data. Synthesia lists SOC 2 Type II, ISO 27001, 27701, and 42001, plus GDPR compliance.

What the launch does not settle

Three questions sit under the surface of this release.

The first is whether interaction quality holds up outside a demo. A roleplay only produces skill transfer if the avatar's pushback is realistic and varied. An avatar that agrees too easily teaches nothing, and one that argues on rails gets recognized within two attempts. Synthesia has not published independent evaluation of how its roleplay counterparts behave across scenarios, and its deployment figures describe repeat use rather than learning outcomes.

The second is the record-keeping question. A score attached to an employee's practice attempt is a data point about a person, created by an AI, stored in an enterprise system, and potentially visible to a manager. 78 percent of learners coming back looks healthy in a press release. It also means most participants produced a scored record of a conversation they handled badly. Where that record lives, who can see it, and how long it is retained are policy decisions every buyer will have to make, and the answers differ sharply between, say, a sales enablement team and an HR department.

The third is what happens to the human specialists whose scarcity is the entire premise. If Synthesia is right that a trained avatar can absorb the repetitive conversations at a fraction of the cost of a human call, then the coaching moment that needs a real person becomes the expensive one. That is a reasonable trade for the company's customers. It is also, unavoidably, a story about which parts of a job get automated first.

Where this leaves enterprise AI

Sessions is a useful marker because it moves an avatar product from content generation into workflow. A one-way avatar video competes with a camera, a studio, and a deadline. A two-way avatar competes with a manager's calendar, and the comparison is no longer about production cost. It is about whether an organization can run a thousand practice conversations a day and keep the results.

Synthesia's bet is that most organizations cannot, and that the avatar that talks back is a cheaper answer than hiring. Whether the level of interaction quality justifies the substitution is what the next year of deployments will show. The repeat-use number is a promising opening signal. It is not yet evidence that the practice made anyone better at the job.