London startup Kult has brought AI into the game itself: it understands what players are doing and how far they have progressed, then answers questions through voice and a second screen. We think this is more worth watching than another chat window, because AI is shifting from “answering questions” to “understanding context and providing continuous assistance.”
What this is
Kult is not building a general-purpose chatbot, but a companion that provides information around the game in progress. Under the hood, it uses Gemini Flash to handle each turn of conversation and a more advanced model to perform one deep-digest pass per game. Its game knowledge comes from an external repository, which the model retrieves and cites. RAG—retrieving information first and then having the model answer based on it—is not a showcase feature here; it is the foundation for reducing outdated information and unsupported “confident” claims.
The design also exposes the practical barriers facing consumer AI products: responses must be fast enough, answers must match the specific game, version, and player progress, and the cost of each call must remain low enough to sustain long-term operation. Kult argues that supporting hundreds of games can still be commercially viable because adding a new game can be largely automated by models rather than maintained by hand, title by title.
Industry view
Our view is that Kult demonstrates a viable path for vertical AI products: general-purpose models handle understanding and interaction, industry data delivers accuracy, and product design brings answers into the interface users are already operating. Google Cloud provides model choice, knowledge-base capabilities, and deployment infrastructure, allowing a small team to assemble a service that fits its use case without training its own foundation model.
But this path is not secure. Kult's article was published by Google Cloud and carries an obvious platform-vendor perspective. It does not disclose user numbers, retention, conversion to paid plans, or unit economics, nor does it show that support for “hundreds of games” has been validated by real players. Rapid game updates, licensing rights to knowledge, voice-recognition errors, and cloud inference costs could all compress margins. General-purpose models may not remain cheap over time either: the greater the usage, the harder it becomes to control inference costs and latency.
Impact on regular people
For enterprise IT: This offers a reference point for enterprise software: an AI assistant does not need to live in a separate chat window; it is easier to create value by entering the workflow employees are already using. The prerequisite, however, is reliable business data and continuous control over inference costs.
For individual work: More software will gain the contextual ability to “know what you are doing,” automatically suggesting next steps across writing, spreadsheets, and design tools. The key shift will move from “giving better answers” to “switching windows less and repeating background less.”
For the consumer market: Gaming may become one of the first categories in which users develop a habit of paying for AI companionship. But players are ultimately buying immediate, reliable help, not the model itself. If latency or the error rate rises even slightly, the companion quickly becomes an interruption.