Synthesia’s AI training platform is moving beyond videos into live coaching

Synthesia built its name on a seductive promise: corporate training videos in minutes, not months, for pennies on the dollar. That pitch carried the company to a $1 billion valuation and a client list spanning half the Fortune 500. But the British startup's newest product admits something its earlier marketing never would — videos don't teach. They inform. They demonstrate. They sit passively on an LMS while employees click through at 2x speed. The real learning happens when someone actually does the thing, fails, gets feedback, and tries again.

Roleplay Sessions, launched Wednesday, is that admission made tangible. An AI avatar runs a sales rep through a pricing objection. It plays a disgruntled customer threatening to churn. It simulates a direct report pushing back on a performance review. Then it scores the employee against a rubric the company designed. The avatar talks back. It pushes back. It remembers what was said three turns ago and uses it. This is not a video. It is a flight simulator for conversations that cost companies millions when they go wrong.

The strategic shift is sharper than the product announcement suggests. Synthesia's avatar and voice technology remain proprietary — that was always the moat. But the reasoning layer powering the conversation is OpenAI's. By wrapping that commodity intelligence in rubrics, performance data, and managerial analytics, Synthesia is repositioning itself from an AI video company into a performance-management platform with a video front end. The distinction matters. Video generation is a feature racing toward zero marginal cost. Proving that a sales force improved its objection handling by 23 percent last quarter is a budget line item nobody cuts.

Victor Riparbelli, Synthesia's CEO and co-founder, puts it plainly: "Video performs way better than text or sending out a document. But for most things, we learn the best by actually practicing something rather than just reading it." He cites a meta-analysis of learning research. The industry has known this for decades. Corporate training has ignored it for just as long because practice at scale used to require human role-players — expensive, inconsistent, impossible to schedule across time zones. AI avatars solve the logistics. The rubric layer solves the accountability.

Early traction suggests the market agrees. One of Europe's three most valuable companies. A top-five Fortune 100 constituent. One of the world's largest recruitment firms. These are not design partners getting free access for logo rights. They are scaled commercial rollouts. The two dominant use cases — sales training and leadership development — both sit in soft skills territory where video has always failed hardest. You cannot lecture empathy. You cannot slide-deck negotiation. You have to do the awkward thing, watch it go sideways, and adjust.

Synthesia's consultant team will still build bespoke programs from a client's existing training documents. That service revenue smooths the adoption curve. But the platform's long-term value accrues in the data exhaust: granular performance maps of entire teams, benchmarked against each other and against the rubric. "From a managerial executive layer, we're seeing huge interest in mapping out the talent in their company in a way that's been really difficult to do before," Riparbelli said. If your entire sales force practices the same scenario, you suddenly know who closes, who crumbles, and who needs coaching before the quarter ends.

Skepticism has purchase points. The scoring rubric is only as good as the human who wrote it. AI evaluators hallucinate. They reward sycophancy. They miss nuance that a seasoned manager would catch. Synthesia says it mitigates this through prompt engineering and human-in-the-loop calibration, but the company has not published validation studies comparing AI scores to expert human raters. Until it does, the analytics layer carries an asterisk. Enterprises buying Roleplay should demand that evidence before tying compensation or promotion decisions to the output.

Competition is gathering. Microsoft's Copilot simulations, Salesforce's Einstein role-plays, a dozen well-funded startups building conversational AI for specific verticals. Most lack Synthesia's avatar fidelity — the lip sync, the micro-expressions, the voice cloning that makes the practice feel visceral rather than robotic. That sensory realism is not cosmetic. It sustains the suspension of disbelief long enough for the employee to actually try. But avatar quality is a depreciating advantage. The reasoning layer will commoditize faster. The rubric-and-analytics layer is where the switching costs live.

Synthesia knows this. The "Sessions" branding signals a platform play beyond Roleplay. Job interviews and candidate screening are next, per TechCrunch's exclusive. Screening is a ruthlessly practical wedge: high volume, clear rubrics, immediate ROI if it reduces bad hires. It also generates training data for the scoring engine — every screened candidate becomes a labeled example of "good" or "bad" responses. That flywheel could compound.

The broader signal: enterprise AI budgets are entering the prove-it phase. CFOs are done funding pilots that produce pretty demos and no measurable behavior change. They want instruments that show the needle moving. Synthesia's pivot from content generation to performance verification mirrors what every AI application layer will eventually face. The winners won't be the best model wrappers. They'll be the ones who make the output auditable, the improvement visible, and the budget defensible.

Roleplay Sessions is Synthesia's down payment on that thesis. Whether it pays off depends on whether the rubrics hold, the scoring withstands audit, and the managers actually use the data. The technology is ready. The organizational discipline is the variable. That has always been the variable. Synthesia just built a product that finally makes it visible.