Reply, an artificial intelligence firm, has introduced its Model Factory concept in London, offering businesses the capability to develop specialised AI models tailored to their proprietary data and expertise. This initiative aims to provide companies with greater control over AI systems by allowing them to customise, govern, and continuously improve models that align closely with their operational needs.
The Model Factory operates through dedicated physical locations known as Houses of Models, with the London site joining an existing facility in Turin, Italy. These centres serve as collaborative spaces where Reply engineers work alongside client experts to convert organisational knowledge into domain-specific generative AI models.
Daniele Vitali, a partner at Reply who oversees the project, highlighted the benefits of this approach. He explained that clients bring valuable information often dispersed across various systems or retained informally within staff expertise, which Reply engineers then help transform into trainable AI models. This process includes identifying data gaps and supplementing them with synthetic datasets or verified external information when necessary.
One example of application is in healthcare, where a client’s two decades of specialised cancer treatment knowledge is being encapsulated into a model designed to assist other clinics. Such models can create new revenue opportunities by scaling expertise beyond its original setting. Additionally, Reply targets scenarios requiring automation at large scale, where general-purpose AI models may become prohibitively costly or ineffective due to their breadth. Specialised models offer comparable performance tailored for specific tasks while reducing running costs.
Clients also benefit from enhanced sovereignty over their AI assets. Vitali emphasized that control extends to all aspects of the model lifecycle: from data collection and training procedures to access rights, deployment environments, and ongoing updates. Ownership remains with the customer, who retains full rights over the datasets and the AI model itself, including any synthetic data derived from customer information.
The development process involves starting with existing open-weight models, which are then fine-tuned using a mix of techniques such as supervised learning, reinforcement learning, and optimisation to meet precise quality, speed, and cost targets agreed upon with clients. Rigorous evaluation phases ensure that models are only deployed once these benchmarks are satisfied.
Post-deployment, Reply assists clients in monitoring user interactions to identify opportunities for further refinement. The structured approach allows versions to be rolled back or adjusted to maintain or improve performance, providing an audit trail of model iterations and their respective data inputs. This systematic tracking also supports compliance with regulatory frameworks like the AI Act.
Reply’s Model Factory aims to convert the investment in AI into durable organisational assets—customised models and reusable datasets—that can evolve alongside business needs and technological advances. This process not only captures and preserves specialised knowledge but also enhances the efficiency and economics of AI integration within enterprises.