After two years of rapid expansion of artificial intelligence agents in companies, what was still perceived as a field of experimentation is gradually becoming a field of exploitation in its own right. The agents were first deployed in more or less convincing POCs. Able to hold a fluid conversation, handle simple requests, automate part of customer support or internal assistance, they have found their place in relatively controlled environments. If this phase allowed organizations to measure the potential, it also made it possible to identify the limits.
Because an agent that operates within a restricted framework does not guarantee its robustness once integrated into the heart of operations. When he intervenes on a customer journey, he no longer just responds, but qualifies, guides, proposes, and sometimes makes consequential decisions. Even partially autonomous, it becomes a sensitive component of the information system and, by extension, of the relationship between the company and its customers, which means that at this level, errors are not allowed.
From a promising tool to an operational asset
In many large groups, discussions around AI agents have therefore gradually left the field of innovation to join that of governance.
During the initial phase, the focus was on the ability of agents to understand an intention, to produce a fluid response, to reduce the volume of tickets. While these indicators are of course relevant, they become insufficient when the agent is called upon to operate on a large scale.
Thus, CX, IT, legal, security and risk departments are now wondering how to supervise a system that continuously learns and interacts with customers, sometimes in critical areas.
Attention is focused on three essential points: first of all reproducibility, that is to say the ability to maintain consistent behavior in varied contexts; then observability, which allows us to understand what the agent does, on what basis and with what effects; finally corrigibility, essential for detecting a deviation, adjusting a rule, testing a modification, then redeploying it.
The agent thus goes from a simple tool to an operational asset, and is subject to the same requirements as the other bricks of the information system.
CX as a laboratory for industrialization
If this maturation is today more visible within customer services, it is no coincidence. The high volume of interactions and the repetition of scenarios make CX a privileged field of observation. It also highlights the less visible costs of agentic AI. Automation does not eliminate complexity and requires continuous human supervision, regular training cycles, exception support, integration with existing tools, as well as constant alignment with commercial policies and regulatory constraints.
A competition that moves to the management layer
A problem which has not escaped the eye of investors, who are more interested in startups offering a management layer capable of transforming the agent into an manageable system than in those which promise the smoothest conversations.
Parloa illustrates this point perfectly. The German company does not only highlight the performance of its agents, but above all the capacity of its platform to design, simulate, supervise and develop them within a controlled framework. An approach in line with the expectations of large organizations faced with reliability and responsibility issues.
2026, a year of normalization rather than rupture
Talking about the end of “experimental” agents in 2026 does not mean the end of testing or innovation, but rather a change in framework. The experiment will be part of more formal systems, with validation criteria and supervision tools.
Companies will favor platforms capable of orchestrating several specific agents, with the challenge of making them evolve in a manner consistent with their constraints and objectives. From this perspective, the management layer becomes significant.
Specialized players, emerging platforms
The competitive landscape for enterprise AI agents is rapidly shaping, but has not yet stabilized. In Europe, the market remains fragmented between players specialized by vertical or use. In Germany, Parloa and Cognigy (acquired at the end of the year by the American Nice) occupy a position in complex contact center environments, while in the United Kingdom PolyAI has established itself in voice and large-scale conversational agents. In France, the ecosystem remains more dispersed, with players positioned on hybrid approaches, specific technological bricks or sectoral use cases, without a platform standard having yet established itself. In the United States, competition is playing out on another level, with startups like Sierra, which are alongside the initiatives of large software and cloud publishers, starting with Salesforce, Microsoft and Google, which are gradually integrating agentic capabilities into their existing platforms.
A Series D round that changes the dimension of PARLOA
Founded in 2018 in Berlin by Malte Kosub and Stefan Ostwald, PARLOA develops an AI agent management platform for large companies. The company employs around 380 people and has offices in Berlin, Munich and New York. It says it works with several major international accounts, including ALLIANZ, BOOKING.COM, SAP and SWISS LIFE.
Parloa announces a fundraising of 310 million euros in Series D, led by GENERAL CATALYST, with the participation of EQT VENTURES, ALTIMETER CAPITAL, DURABLE CAPITAL PARTNERS and MOSAIC VENTURES. This operation values the company at around $3 billion and brings the total capital raised to more than 480 million euros in less than four years.
