Patrick Joubert proposes systematic oversight of AI agents

For Patrick Joubert, entrepreneur with his startup Rippletide, the main obstacle to the adoption of AI agents is not performance but the lack of an operational safeguard. He emphasizes an independent control layer, based on a systematic evaluation of actions before execution, as a lever for trust and production deployment.

Independent control makes agency more deterministic

The advocated approach aims to shift agents from a probabilistic and risky logic to a deterministic, controlled, explainable operation that aligns with the specifics of the company. It is presented as having already proven its effectiveness and as the key to unlocking use cases that can produce quick benefits. According to this vision, these qualities foster renewed trust, considered the main gain from operational control. However, the opportunities opened by agents remain conditioned on two combined requirements: trusting them and controlling them.

Pilots deemed solid but too unreliable for production

Many organizations consider their AI agent pilots promising but insufficiently reliable to make the leap to production. Doubts, reinforced by increasing uncertainties about costs and the human aspect, weaken the momentum for adoption and widen the gap between advanced and hesitant companies. The primary cause is linked to the probabilistic nature of language models: even optimized, an agent can make significant errors. The issue is thus characterized as less technical than managerial and operational, and does not justify a blank check or an exclusive race for performance.

A "professional conscience" filters the actions of agents

The stated priority is to avoid any harmful acts for the company by systematically evaluating proposed actions and blocking those that are inappropriate or contrary to regulations, culture, ethics, or long-term strategy. In addition to the connectors and actuators that give agents access to the world, it is proposed to equip them with a "professional conscience": an independent control infrastructure from the model, which synthesizes, in the form of ontology, the written and unwritten rules of the organization. Taking context into account, this layer evaluates the legitimacy and relevance of each action and issues a green light for automatic execution, an orange light to require prior human control, or a red light to prohibit. Such pre-execution evaluation is presented as necessary to prevent plausible mistakes, such as a transfer exceeding a delegation threshold or anti-money laundering checks, or the recommendation of a medication for unauthorized use.

With trust, use cases expand beyond "super-RPA"

In a climate of distrust, professions restrict AI agents to use cases where an error would have minimal consequences, favoring the automation of simple tasks that are easy to control and manually correct, and primarily seeking productivity gains to justify an investment in "super-RPA." Conversely, when trust is established and agents remain aligned with the company's goals, creativity and boldness can flourish. An approach fueled by trust would allow for the full exploitation of their analytical, reasoning, and action capabilities, described as capable of overcoming barriers of time, cost, labor, and knowledge, and reinventing methods, customer relationships, products, and business models.

The message and its author

This perspective and the proposals it articulates are championed by Patrick Joubert, engaged with AI agents through his startup Rippletide. He points out barriers to adoption and argues that these tools can multiply creativity and productivity, provided they are managed through an evaluation of their actions before execution.