The AI agent is no longer just the discreet assistant who summarizes a meeting or prepares a memo. It is now becoming integrated into the company’s actual workflows: customer service, transactional banking, financial planning, cybersecurity, sales, collections, and recruitment. This shift changes responsibility more than it does the technology itself. Because as soon as a system can take action, decisions must be made about who authorizes it, who monitors it, and who is held accountable for its errors.
From Assistant to Employee Under Supervision
The simplest definition of an agent can be summed up in just a few words: software capable of utilizing a language model, reasoning within a given framework, and, above all, taking action. That’s the key difference. A co-pilot makes suggestions. An agent carries them out.
IBM reports that it is carrying out more than 500 projects related to these solutions, representing approximately $10 billion in revenue. In cybersecurity, the company states that it has 28 agents capable of conducting threat investigations, handling 70,000 investigations per month for its clients. This is therefore no longer an experimental field; it is now entering production lines.
But the software only creates value when the company is willing to take an unflinching look at its own processes. IBM says it broke down its organization into 490 workflows and then overhauled 70 of them, representing $25 billion in expenses. Three years later, those expenses had reportedly been reduced to about $20 billion—a savings of $4.5 billion. The lesson is stark: buying licenses doesn’t change anything. Redesigning work, however, does.
The promise is not autonomy, but architecture
In regulated sectors, the AI agent cannot be a sleek “black box.” At BNP Paribas, the question is not what general level of autonomy to grant the machine, but rather what level of autonomy to grant to which process, and within what controlled framework.
The example of the self-driving car is illuminating. On an open road, the technology is impressive. In the city, things get complicated: pedestrians, ambiguous signals, and unpredictable behavior. A business is more like a city than a test track. Incomplete data, access rights, exceptions, vulnerable customers, compliance, costs, audits—this is where an agent must prove their maturity.
The model itself is no longer always the core of the problem. The real issue is now the framework: agent identity, permissions, action logs, observability, revocation rights, rollback, and the separation between slow governance and fast execution. An agent is neither an employee nor a traditional application. It is becoming a new player in the information system.
Details
A bank agent should only be able to act at the intersection of three permissions: what it claims to be able to do, what internal policy allows it to do, and what the entity behind the request—whether a human, an application, or another agent—is itself authorized to request. Responding to this intersection in real time is one of the real technical challenges of agent-based AI in production.
Customer Service: A Visible Test of Acceptability
For the general public, the first interaction with these agents will rarely take place in an executive committee meeting. It will happen over the phone, in a chat, when dealing with a bill, a complaint, a loan application, or a connection issue.
The question “Do customers want to talk to an AI?” is almost the wrong one to ask. No one really wants to contact customer service. Users want their problem solved—quickly, clearly, and without being passed around. If the agent can do that, resistance decreases. Wonderful AI has observed that resolution rates increase over time for the same agents, a sign that usage is becoming more normal as trust is built.
Even more troubling: some sales agents are already outperforming humans when the timing is just right. In the credit sector, an agent who detects that a customer is approaching their credit card limit and offers a loan at just the right moment can close more deals than a human advisor. Not because of greater empathy, but because of precise timing. Business intelligence is thus shifting from rhetoric to timing.
The New Grammar of Management
Within the company, Pigment describes an initial wave of clear benefits: faster modeling, accelerated analyses, and financial scenarios generated in parallel. In some cases, modeling time can be reduced by 80%. But the next phase is more strategic: improving the quality of decision-making, not just the speed of execution.
This requires a shared infrastructure. An agent cannot be reliable in a fragmented organization, where each team experiments with its own tool, using its own data and following its own rules. It needs a single source of truth, access rights, traceability, and audit capabilities. In short, a Data House comes before an Agent House.
Managers, too, are taking on new roles. They no longer simply assign tasks. They set goals, monitor progress, approve decisions, and strike a balance between human and digital work. Companies no longer just recruit talent; they integrate agents into their operational systems.
What Must Remain Human
The speakers agree on one point: any task that is repeatable, structured, and controllable can be delegated to an agent. These include meeting minutes, analyses, literature reviews, first-level support, scenario preparation, recruitment assistance, and security investigations.
But the final decision, the intuitive judgment, the relationship of trust, and certain evaluation steps remain human. Recruitment is one example. AI can speed up certain steps, as in the case of an American hospital that was able to hire nurses 12 days faster thanks to the partial automation of the process. But choosing someone to join a culture, picking up on subtle nuances, or sensing hesitation remains a leadership decision.
In the healthcare or insurance sectors, the line is even clearer: an agent can provide guidance, help find a healthcare provider, and handle administrative requests. The agent must not make a diagnosis or substitute his or her judgment for that of a healthcare professional.
The next transformation of work will therefore not resemble a complete disappearance of the human element. It will more closely resemble a redistribution of tasks: speed, memory, and coordination will be left to machines; discernment, intention, and responsibility will be left to humans. It remains to be seen whether companies will be able to establish this framework before rushing to delegate tasks they do not yet fully understand.
Cette publication est également disponible en :
