Summary
- EQT Ventures and FirstMark co-led Remanence's $30m seed round.
- The company's platform converts AI agent activity into training environments for specialised open-weight models.
- Early customer work spans legal services, customer support and enterprise software in Europe and the US.
Paris-based artificial intelligence startup Remanence has raised $30 million to develop technology that converts the operational records generated by enterprise AI agents into specialised models that businesses can own and improve.
The seed round was co-led by EQT Ventures and FirstMark, with participation from Zero Prime, Emblem, Drysdale, Global Founders Capital and Yellow.
Remanence is developing software that examines how AI agents complete business tasks and uses that information to train models adapted to an organisation’s particular processes. The company is working with customers in legal services, customer support and enterprise software across Europe and the United States.
The investment will support further research, engineering recruitment and computing infrastructure as Remanence develops an alternative to relying exclusively on general-purpose models supplied through external AI services.
Agent activity becomes training material
Enterprise AI agents can interact with applications, retrieve information and carry out sequences of actions intended to complete a task.
These activities produce execution records containing instructions, tool calls, observed results, errors and subsequent attempts to correct them.
Such information is commonly used to monitor performance or investigate failures. Remanence wants to make it part of the training process itself, allowing models to learn from the work that enterprise agents are already undertaking.
The company’s technology analyses execution traces and uses them to construct simulated environments based on operational tasks.
Models can then practise those activities through reinforcement learning, receiving feedback according to whether their actions produce acceptable results.
For example, a model carrying out an administrative process might need to retrieve the correct records, follow a particular sequence of operations and produce an output consistent with an organisation’s rules.
The training process can evaluate different attempts and adjust the model’s behaviour in response to the results.
Remanence’s proposed system involves identifying relevant tasks, building environments in which those tasks can be practised and establishing methods for assessing successful completion.
That final element is particularly important. An automated system cannot reliably improve merely by collecting large volumes of activity if it cannot distinguish appropriate outcomes from errors.
Training data may also contain incomplete records, unusual decisions or actions that should not be reproduced. The system therefore needs controls over which examples contribute to training and how successful behaviour is defined.
Specialised models offer another approach
Many organisations currently build AI applications around general-purpose models, adding instructions, retrieval systems and external software tools to make those models useful for specific business activities.
This approach allows companies to use increasingly capable technology without developing their own models. However, repeated workflows may continue to depend on an external service whose pricing, availability and underlying behaviour are controlled by another supplier.
Remanence is using open-weight models that can be adapted to specialised tasks through further training.
Customers retain ownership of the resulting model weights, providing greater flexibility over where the technology is deployed and how it is subsequently developed.
That ownership does not eliminate all external dependencies. The underlying models remain subject to their applicable licences, while training and deployment still require computing resources, specialist software and appropriate technical support.
Nevertheless, businesses may gain more control over model development than they would through a service that permits only limited customisation.
The commercial attraction depends partly on whether specialised models can complete repeated tasks more efficiently than larger general-purpose systems.
An application handling a substantial volume of similar transactions may benefit from a smaller model trained to perform those operations reliably.
More varied work may still require the broader capabilities of a frontier model, particularly where tasks involve unfamiliar situations or extensive reasoning.
Enterprises may therefore use specialised and general-purpose systems alongside one another rather than replacing one category entirely.
Performance depends on reliable evaluation
Reinforcement learning introduces costs and operational requirements beyond those associated with ordinary AI inference.
Training environments must represent the activities a model is expected to perform, while evaluation systems need to identify successful outcomes accurately.
If the simulated environment omits important restrictions or differs substantially from actual business systems, a model may perform well during training without producing reliable results in production.
The problem becomes more complicated where success involves professional judgement or depends on information that is not fully captured in software records.
Legal workflows, for example, may include repetitive document preparation alongside decisions requiring contextual interpretation and specialist review.
Customer service systems must handle common requests efficiently while recognising when circumstances require escalation to an employee.
In both cases, organisations need to establish which activities can be assessed automatically and which require additional supervision.
Models trained continuously on new activity also require monitoring for unintended changes in behaviour.
Improvements on familiar tasks may be accompanied by poorer performance on unusual cases, while changes in the underlying business applications can make previously successful procedures unreliable.
Performance evaluation must therefore consider accuracy, reliability and the cost of maintaining the system, rather than counting completed training cycles.
Former H Company researchers lead development
Remanence was founded by Pierre-Louis Cedoz, Hamza Benchekroun and Axel Moyal, who previously worked together at French AI developer H Company.
The founders held senior responsibilities in post-training work associated with H Company’s open-weight Holo models.
Their backgrounds include experience across AI research organisations and technology businesses, providing expertise in adapting models after their initial training.
Remanence intends to bring those techniques into enterprise environments, where customers may not have the resources to develop and maintain specialised model training systems internally.
The company is building its team principally in Paris, with additional activity in London and San Francisco.
Its investors include executives and founders associated with established enterprise software companies, although their participation does not independently establish the performance or commercial adoption of the technology.
The company has disclosed customer activity across several sectors but has not published comprehensive revenue figures or independently verified comparisons showing the performance of its specialised models against alternatives.
Its new financing will support the infrastructure needed to generate those results, including development of simulation environments, reinforcement learning systems and evaluation methods.
For enterprise customers, the economic calculation will involve the costs of training and operating specialised models against the continuing expenditure associated with general-purpose AI services.
Information governance will also affect deployment decisions. Agent records can contain confidential business information, making access controls, retention arrangements and the handling of training data central to the proposed system.
Remanence is seeking to turn operational experience into a reusable technical asset rather than leaving it solely within application logs and external service interactions.
The commercial evidence will emerge as customers test whether that process produces measurable improvements in task completion, reliability and overall operating costs across sustained deployments.












