For years, corporate AI was presented as an assistant: it answered a question, summarized a document or prepared a draft. The AI agent changes that relationship. It can break down a goal, select tools, query archives, update systems and deliver a result. The decisive shift is not from person to machine, but from suggestion to action.
The partnership announced by IBM and OpenAI on August 13 brings frontier models and products into the platform IBM Consulting uses to develop enterprise workflows. The declared objective is to apply AI to core operations, software modernization and security. The value promised is no longer limited to generating content. It lies in allowing AI to operate inside processes with economic, organizational and sometimes legal consequences.
An agent can open a case, review a contract, modify a database, arrange a purchase or coordinate other specialized agents. Each action may appear small; in sequence, they form a decision. A conversational interface mainly shows the final answer. An operational system must make it possible to reconstruct which sources were consulted, which tools were used and which step turned a probability into an action.
This is where the question of signature begins. When an employee prepares a document, we know who drafted, reviewed and approved it. When an agent moves through ten applications and composes the outcome autonomously, authorship is distributed among the people who set the objective, configured the system, own the data and authorized execution. Saying “AI did it” erases the chain of responsibility the company must preserve.
Human oversight cannot be reduced to one final approval button. Control must be designed in advance through spending limits, differentiated permissions, prohibited data, escalation thresholds and moments when the agent must stop and request confirmation. Useful autonomy is not the absence of rules; it is the capacity to act within a legible perimeter.
Errors also change form. A chatbot may produce false information that is read and perhaps corrected. An agent can use that information to send a message, change a price or block an account. The error leaves language and enters the process. Audit trails, reversible actions and escalation procedures therefore become part of the product rather than administrative additions.
WHO SIGNS? is not a question against automation. It is the condition for using it seriously. When AI enters a company’s central processes, every agent needs a mandate, a perimeter, a record of its actions and a person or function accountable for the outcome. The future of work will depend not only on what agents can do, but on how clearly their decisions can be attributed.