AI Agents: Seven IT Leaders Call for Regulation of Their Actions

Seven technology leaders interviewed about the priorities for 2027 recommend limiting the actions of AI agents and renewing tests when their model changes. Their recommendations also cover measuring outcomes, preparing data, and choosing between external and self-hosted models.
Defining Responsibilities and Rights of Agents
At Sedgwick, a specialist in claims management, Sean Safieh leads the information systems for global platforms and digital solutions. He describes agents that no longer just summarize or recommend: they guide work and trigger transactions. A sales team, for example, might allow an agent to independently validate a transaction. For this leader, these capabilities present a different operational risk, particularly in complex organizations.
Sean Safieh contrasts the generally identifiable responsibility of a person with that of an agent, where it must be determined who authorized access and what actions are permitted. He advocates for explicit oversight, auditable and explainable actions, as well as reliable traceability. At Sedgwick, he specifies that AI governance falls under the business, not just IT control: it touches on risks, compliance, and the trust placed in tools. Before increasing an agent's autonomy, he recommends treating it as a corporate identity and defining responsibilities with business and risk management.
For Boris Kolev, Global Technology Officer at JA Worldwide, an NGO dedicated to youth education, the change also lies in the actions permitted, not just the reliability of responses. According to him, modifying data, engaging in a transaction, communicating externally, or activating another system transforms an error into operational risk. He calls for identity management, permissions, and audit trails tailored to agents. Spending caps, escalation procedures to a supervisor, and a possibility for human intervention should, in his view, precede their deep integration into critical processes.
Retesting Systems When Their Model Changes
Frank Meltke, CEO of the digital transformation consulting firm Contraco, warns that an already approved system can change its model without new authorization. Providers continuously replace and update their models, he claims. An assistant validated in the spring could therefore behave differently in the fall while retaining its name, rights, and role in processes. He believes very few companies trigger a new review in such cases.
Le brief IA que les pros lisent chaque soir
Les 7 actus IA du jour, décryptées en 5 min. Gratuit.
Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.
Choisissez votre rythme
Gratuit · Pas de spam · Désabonnement en 1 clic
His recommendation is to record the model and its version for each system in production, then ask providers to announce changes in advance. Initial acceptance tests should be replayed with every change, he insists, especially when a tool can approve an operation, make a purchase, or modify records without human intervention.
Daniel Herrington, Director of AI at insurance comparison site The Zebra, calls for the development of synthetic assessments to verify how agents act on data. He considers companies that have not invested in this area to be already behind. He deems it naive to rely on traditional protections such as two-factor authentication or CAPTCHAs and predicts that security platforms will ultimately lose this confrontation.
Measuring Outcomes and Preparing Data
AI is among the top concerns for information systems leaders for 2027. For Michael Corrigan, who holds this position at World Insurance Associates, an investment must address a specific business or IT problem. Before deploying the tool, the company should define what it will measure and the criteria that will be used to judge its effectiveness. The same indicators should then allow for a comparison of the previous situation to the new results, he recommends.
Mojgan Lefebvre, Chief Technology and Operations Officer at insurer Travelers, emphasizes the importance of data preparation. She reports that an experienced specialist on her team considers their proper organization a prerequisite for using AI. Quality, classification, and understanding must be worked on, she explains. She also highlights the need to provide the business context of the data, as its interpretation by AI differs from that of humans.
Distributing Requests Among Models
Andy Boettcher, Director of Innovation at DoubleTrack, an IT services and consulting company, anticipates a combination of large cutting-edge models and open models hosted by companies themselves by 2027. In this scenario, routers would direct each request according to classification rules. He presents the management of this ensemble, rather than a single model, as a challenge to prepare for.
The leader states that he uses three open models alongside a cutting-edge model on his machine and finds their performance comparable to the latter. He recommends defining now which data and questions can be entrusted to each destination. The goal he sets is to keep the most sensitive work within the company.
Brief IA — L'actualité IA en français
L'essentiel de l'actualité de l'intelligence artificielle, décrypté et expliqué chaque jour.