The battle for legal AI will not be decided by the best model, but by control of the workflow

Introduction

In the space of a few years, the French legal ecosystem has seen a remarkable surge of activity around generative artificial intelligence. Incumbent legal publishers, established legaltechs, specialised start-ups and international players are already offering solutions aimed at lawyers, in-house counsel, notaries, judicial officers (commissaires de justice), in-house legal departments and public administrations. This momentum is healthy. It shows that the French market is not condemned to be a passive recipient of innovations from the United States or the United Kingdom.

But it sometimes masks the real strategic shift under way. The first question raised by legal AI was: which model gives the best answer to a legal question? The next one will be different: which player controls the working environment in which the question is formulated, the documents are added, the answer is produced, the validation is carried out, the client is informed and the data is reused?

In other words, the battle will not be fought only at the level of the model. It will be fought at the level of the workflow. And that distinction is decisive for understanding the risks and opportunities facing French legal players.

1. The first wave: AI as a legal production tool

The first wave of AI adoption in law focused mainly on generation and research. Professionals tested the models to draft emails, produce summaries, reword clauses, prepare outlines for legal opinions, summarise court decisions or explore a document database. This phase was logical: it allowed the time savings and the limits of the tool to be measured very quickly.

It also allowed several French players to position themselves. Legal publishers began integrating generative features into their platforms. Legaltechs accelerated the development of profession-specific assistants. New start-ups offered drafting, analysis or automation tools. The State itself experimented with uses of generative AI to improve its relationship with users of public services and to assist its civil servants.

This phase nevertheless has a limit: it remains centred on producing an isolated output. Yet legal work cannot be reduced to an answer. It is made up of sequences, checks, interactions, versions, validations, exchanges with clients, constraints of professional ethics and evidentiary requirements. It is this chain that the most ambitious players are now seeking to capture.

2. The second wave: AI as a workflow layer

The arrival of agentic solutions marks a turning point. An agent is not just a chatbot you put a question to. It is designed to carry out a task over several steps: open a case file, analyse documents, compare versions, generate a table, identify anomalies, prepare a draft response, organise the exhibits and sometimes interact with other applications.

For the legal sector, this development is decisive. The most time-consuming tasks are rarely simple question-and-answer exchanges. They are workflows: contract review, due diligence, litigation management, handling internal requests, compliance, producing reports, filing documents, preparing for meetings, tracking obligations, responding to tenders, knowledge management. An isolated model can help. An integrated workflow can transform the organisation.

This is why the announcements around Claude Cowork, platforms such as Harvey, solutions integrated into Microsoft’s suites or products developed by the major legal publishers must be read as a strategic signal. Providers no longer want merely to supply the intelligence. They want to supply the working interface.

3. Why control of the workflow is worth more than raw performance

In a professional market, model performance is necessary, but rarely sufficient. A firm does not buy an AI because it scores well on a general-purpose benchmark. It adopts it if it fits into its practices, respects its constraints, shortens its turnaround times, improves its quality control and does not create disproportionate new risks.

Value therefore lies in the combination of several elements: access to the right sources, structuring of internal data, integration with existing tools, compliance, security, traceability, usability, deployment capability, team training, support and measurement of return on investment. A very high-performing but poorly integrated model will remain a gimmick. A slightly less spectacular model, but one deeply integrated into the life of the firm, can become indispensable.

We therefore need to move away from an overly technological view of the market. The legal market will not be won by the best model in the abstract, but by the best chain of trust. That chain begins with the data, continues with the tool, passes through human validation and ends with delivery to the client. Whoever controls that chain controls part of the value created.

4. The risk of platformisation of the legal market

Platformisation refers to the moment when a player no longer merely sells a tool, but becomes the environment in which others work. In law, that risk is now real. If a general-purpose provider manages to integrate the documents, the templates, the data, the applications, the client exchanges and the working habits, it can capture the daily relationship between the professional and their subject matter.

This risk is not theoretical. The major AI providers have considerable financial resources, a capacity for very rapid innovation, direct access to users and global distribution power. They can enter a vertical market without having to patiently build all the assets of the specialised players. When a general-purpose model adds a legal plugin, it does not immediately replace a legaltech. But it changes market perceptions and the value attributed to existing players.

The February 2026 episode around Anthropic’s announcement on legal workflows showed this. The markets reacted not because Claude had instantly replaced the legal publishers, but because the announcement signalled that frontier-model players were ready to move up the value chain. They are no longer content to be the invisible infrastructure. They can become direct competitors at the application level.

For French players, the stakes are clear: it is no longer enough to have a good legal AI engine. They need to own access to the workflow, to domain data, to the relationship of trust and to distribution channels.

5. The strengths of French and European players

French players are not starting unarmed. They have an intimate knowledge of the professions, the rules of professional ethics, French sources, the practices of the courts, confidentiality constraints and the expectations of firms. This proximity is a decisive advantage, provided it is turned into a product and commercial advantage.

The French legal market is also highly fragmented. The needs of an international business law firm, a mid-sized law firm, a sole practitioner, an in-house legal department, a notary or a public administration are not identical. This diversity makes total domination by a single tool difficult. It opens up space for vertical, specialised, well-integrated and well-supported solutions.

European players can also draw on a narrative of sovereignty, data protection and compliance. But that narrative will not be enough if it turns into a mere marketing argument. Sovereignty is only worth something if it comes with product quality, genuine support, integration capability, proven security and a user experience comparable to that of the best international players.

6. The three battles to be won

To secure a lasting position, French legal AI players will have to win three battles.

The first is the battle for the use case. Solutions that are too general will be squeezed between consumer models and office suites. The strongest players will be those that address specific business problems: clause review, generation of specialised contracts, litigation management, sector-specific compliance, case preparation, case-law analysis, support for the in-house legal department, knowledge management.

The second is the battle for data. General-purpose models know a great deal, but they do not necessarily know the internal documents, a firm’s own doctrine, a client’s usual positions, the validated templates, the negotiation histories or the constraints specific to a profession. Organising, cleaning, protecting and exploiting this domain data will become a major source of differentiation.

The third is the battle for distribution. In law, adoption cannot be decreed. It comes through referrers, professional bodies, in-house legal departments, professional networks, integrators, training, references and partnerships. The best tool can fail if it does not have access to the right decision-makers or is not backed by a credible growth strategy.

Conclusion

The legal AI market will not only be a competition between models. It will be a competition for access to real work. Legal professionals do not need a spectacular assistant that performs well in a demonstration. They need a tool that is reliable, integrated, secure, understandable and suited to the constraints of their profession.

This reality creates an opportunity for French players. They can build solutions that are closer to actual practice, more respectful of professional obligations and better suited to local uses. But they will have to move fast. For the general-purpose platforms have understood that law is a market with real purchasing power, document-intensive and ripe for automation.

The strategic question is therefore not: who will have the best model? It is: who will control the legal workflow of tomorrow?