In June 2023 I wrote a post explaining why we built Video Canvas the way we did. Personalised video platforms at the time rendered in the cloud, so a viewer's data had to leave the firm before the video could be made. We took a different route: render every personalised element in the browser, offer the whole thing self-hosted, and keep customer data inside the customer's own infrastructure.
We called it data sovereignty: the firm's customer data never leaves the firm's control, not for processing, not for rendering, not for anything. At the time that read as a security preference. Three years later it turns out to have been the decision that matters most.
Nothing about that architecture has needed to change. The player is still ours, built in-house, responsive across every screen. The recap exercise still turns watching into understanding and still gives firms evidence that it happened. Personalised elements still render in the browser, from names and figures to charts, tables and dashboards. Self-hosting is still a set of files a firm drops into its own infrastructure, with nothing calling home. Data sovereignty is still the default, not an option.
What has changed is the pressure on those choices. In 2023 data sovereignty protected against breaches. In 2026 it protects against something bigger: the pull of generative AI, and the fact that no financial services firm can safely follow its customers there without handing their data to a third party.
People have learnt to take their financial questions to AI. They go because they want two things they are not getting from the mostly static experiences financial services provides: to understand their own position, and to have it explained in terms of their own circumstances. They put in their details and get back a conversation, in plain language, about themselves. Our own consumer tracking shows how deep this now runs. Over five waves and more than a thousand responses since 2024, 56% of consumers now use AI for their finances, and 89% expect their providers to be using it too. Only 45% say they actually want that. Expectation is running ahead of comfort.
The same research shows what people want instead. Consumers want personalised, interactive explanations of their own financial position, and their expectation that providers will deliver them has risen from 38% to 52% in two years. Almost a third say they want both: the ease of AI and the reassurance of interaction.
Firms know all of this. And they have largely done nothing, because putting customer data anywhere near a generative model brings a hallucination risk, a data leakage risk and a regulatory conversation nobody wants. The FCA's Mills Review sharpened the point this summer: be more personal, and be able to evidence that every customer understood every word.
What sits behind Video Canvas is not a language model, and it does not use one. It is a body of a firm's approved financial communications, compressed into a single asset, with the player drawing from it at runtime to respond to the individual in front of it.
The comparison to a small language model is tempting, and it is wrong in the one place that matters. A language model generates. It predicts the next word, and in a regulated industry that is the risk: a sentence nobody has read, said to a customer about their money. Our asset cannot generate anything. Every possible response is authored, reviewed and approved before the asset exists. Nothing is predicted, nothing is invented, and nothing can be said that has not already been read and signed off by a human.
What we share with the idea of a model is only this: a large body of knowledge, compressed into one asset, able to respond to any individual's situation the moment they arrive. That part is useful. We kept it, and we left the generation out.
A Customer Language Module is a single, self-contained asset holding every approved thing a firm can say about a topic, in that firm's own voice. It is calibrated to how they describe a pension, how they explain risk, how they talk to their customers, and it is signed off once by their compliance team. It is deployed inside the firm's own infrastructure, exactly as Video Canvas has always been, so the sovereignty story holds end to end: the module lives in the firm's estate, the viewer's data is read in the browser, and nothing about either is sent anywhere. It can only speak pre-approved messages.
The Video Canvas player takes an individual viewer's data and assembles, live in the browser, a personalised interactive video from that module. Chapters are selected, figures rendered, journeys shaped to the person watching. There is no inference, no model weights, no possibility of a sentence that nobody has seen before. Every word was approved before it shipped.
To the viewer it feels like a conversation with someone who knows their circumstances. To the risk team it is a finite catalogue they approved once and can evidence forever. Both are true at the same time, and I do not know of anything else that makes them true at the same time.
If there is no AI in the finished module, where is the AI? In how we make it.
Think of a modern aircraft wing or a golf club head. Neither uses AI in flight or on the tee. Both are the product of thousands of hours of simulation no human with a drawing board could match. The remarkable part is not the carbon; it is the construction.
Prism, our agentic publishing engine, plays that role for us. Researching, generating and validating an exhaustive, compliant corpus in a firm's own voice used to take a studio months. Prism lets us do it at a scale and fidelity that hand-built production cannot reach. AI is everywhere at build time. It is nowhere at delivery, and customer data never touches it. That is the whole design, and it is what makes the data sovereignty claim more than a hosting arrangement. There is no model at runtime for the data to leak into.
We did not set out in 2023 to answer the AI question. We set out to keep customer data where it belonged. But the two turn out to be the same problem. A firm that keeps its communications and its customer data inside its own estate, and that can only ever speak words it has approved, is a firm that can give its customers the personal, conversational experience they have learnt to expect, without taking the one risk it cannot afford.
That is the safe version of the thing everyone in financial services is afraid of. Three years on, it is what we made.
"Great video content is so engaging for members, especially when it is deeply personalised to them. We chose to partner with Video Canvas, not just for great tech, but for their creativity and their strong focus on security."
Chris Connelly
Chief Strategy Officer
