The digital transformation of the most regulated sectors was built in successive layers, often at the cost of technical and organizational compromises that favored stability over agility. Finance, insurance and telecoms have thus accumulated complex systems, which are difficult to interoperate, where each evolution involves heavy trade-offs.
What appeared to be a barrier today becomes a point of support. For two years, these same environments have attracted a new generation of actors who no longer seek to replace existing systems, but to operate within their complexity.
Organizations structured by complexity
In these industries, a seemingly simple process, such as opening an account, processing a claim, modifying a contract, in reality requires a chain of tools, validations and business rules. Information circulates between several systems, often inherited, and rarely synchronized.
This fragmentation has the direct consequence of maintaining a strong dependence on human operations. Teams play an implicit orchestration role, compensating for tool limitations through coordination, interpretation and verification.
Attempts at transformation have long been based on the logic of recasting. Replace a core system, migrate to the cloud, unify databases. These projects still exist, but their horizon is long, their cost high and their execution uncertain. Innovation, in this framework, progresses in increments. This is where AI agents come into play.
A technology designed to operate in the existing environment
Unlike previous waves, AI is not introduced here as a frontal break. It is positioned as a layer capable of navigating between heterogeneous systems, interpreting unstructured data and triggering actions in different environments.
What changes is the ability to chain operations. Where a chatbot could answer a question, an agent can now follow a file, collect information, query internal databases, apply business rules and guide further processing. AI is no longer limited to the interface and is starting to intervene in operational logic.
Regulation as a structuring constraint
In finance or insurance, no automation can be considered without meeting traceability and explainability requirements. Each decision must be able to be justified, reconstructed, audited.
This constraint limits superficial approaches. It imposes a technical discipline that transforms the way we design systems. Agents cannot simply generate plausible responses; they must rely on identified sources, produce structured outputs and maintain a usable memory of their actions.
This framework slows down certain experiments, but it encourages the emergence of more robust solutions. Where other sectors can tolerate a degree of approximation, these require operational precision. This requirement, far from being an absolute obstacle, acts as a filter which selects more solid architectures.
The customer experience as an entry point
The customer experience constitutes a privileged field of experimentation. It concentrates both a large volume of interactions, strong pressure on costs and a growing expectation in terms of quality of service.
By automating part of the exchanges (processing of requests, collection of documents, contextualized responses), companies obtain immediate gains. But this first layer quickly reveals a limitation: to respond effectively, you must access internal systems, understand business rules and coordinate several steps.
The automation of the front then leads to a transformation of the back. What starts as an optimization of customer service gradually becomes a reorganization of internal processes.
A change in the nature of AI in business
AI is expected on its ability to execute, to be part of operational chains, to manage real cases with their exceptions and constraints. This move to execution introduces new requirements. It is no longer just a question of algorithmic performance, but of reliability, error management, consistency with existing systems. In critical environments, these dimensions become central.
This is what makes these industries both difficult to enter and particularly attractive. Complexity slows entry, but protects acquired positions.
A gradual but cumulative transformation
The introduction of AI agents does not result in an immediate disruption. It follows an accumulation logic, first on limited tasks, then on longer sequences, until covering entire workflows.
As these systems expand, they change the very structure of organizations. They reduce some friction and make dependencies visible. This process is slow, but difficult to reverse.
The paradox of rigid industries
What is emerging is a paradox. The most constrained sectors, those where innovation seemed the slowest, become ideal areas for in-depth transformation. Not because they evolve faster than others, but because the value of an improvement is higher there.
In these environments, every productivity gain, every delay reduction, every improvement in customer experience has a measurable impact. AI, by tackling the coordination of systems rather than their replacement, finds a lever adapted to this reality.
The transformation remains gradual, supervised, sometimes cautious. But it is part of a dynamic where complexity, long endured, gradually becomes a space for intervention.
Several players are already illustrating this evolution, with approaches that converge towards the same orchestration logic. In customer service, platforms like Intercom or Zendesk now integrate agents capable of processing complete requests by relying on internal systems, beyond simple conversational responses. In process automation, UiPath extends its historical RPA scope by adding AI capabilities to manage more complex and less structured flows. ServiceNow, for its part, is pushing deeper integration between operations management and intelligent agents, directly targeting the critical workflows of large organizations.
As these platforms evolve, a new category begins to emerge, even more transversal. Tools like Granola illustrate this shift towards a so-called “context capture” layer, where the challenge is no longer just to execute tasks, but to continuously structure the information resulting from interactions. By capturing conversations (meetings, internal exchanges, customer calls) and transforming them into usable data, these solutions provide agents with operational context. This material, long informal and dispersed, becomes a structured asset, which can be mobilized in processes. This approach appeals to investors: Granola has just raised $125 million (around €106 million) in a round led by Index Ventures, bringing its valuation to $1.5 billion. The boundary between note-taking, corporate memory and execution system then gradually blurs, paving the way for architectures where AI no longer just acts, but relies on a dynamic understanding of the organization itself.
Finally, more recent players like Notch are immediately positioning themselves on this intermediate layer, by developing platforms capable of coordinating customer interactions, document processing and business execution in highly regulated environments. These solutions differ in their entry point, but reflect a trend: AI is no longer inserted as an isolated overlay, it gradually becomes a component of the operational system itself.
Founded in 2021 by Rafael Broshi, Elool Jacoby and Yuval Peled, the startup recently raised $30 million in Series A, bringing the total funding to $45 million. The round is led by Headline, with participation from Lightspeed Venture Partners, Illuminate Financial, Jibe Ventures and Phoenix.
Other startups are following this trajectory, with complementary approaches. Some, like Sierra or Decagon, focus on automating customer support by integrating agents capable of resolving complex requests in direct interaction with internal systems. Others, like Cresta or Kore.ai, are working on contact center augmentation, combining real-time assistance and partial process execution. In a more automation-oriented vein, platforms like Adept or Cognition are exploring the ability of agents to manipulate existing software interfaces to perform complete tasks without heavy integration.
This very fragmented landscape converges towards the same logic: going beyond the simple generation of responses to take charge of operational sequences. As these solutions gain depth, they redefine the boundary between tool and system, no longer positioning themselves as a support layer, but as an execution mechanism integrated into operations.
