AI is transforming procurement, but without governance and reliable data, it increases risks.
The real risk to procurement isn’t AI, it’s AI without governance
What supervised agentic autonomy really looks like
AI becomes problematic when it acts without a clear framework. The consequences are then well known: erroneous assumptions, errors of judgment, excessive dissemination of information, and decisions that are visibly arbitrary or impossible to trace.
The risk is all the greater as many companies are now increasing the number of specialized AI agents, each dedicated to a specific task. Taken separately, on limited areas or specific tasks, these tools can be effective. But on a large scale, they often create complexity that is difficult to control: who controls what? What data is circulating? Who is responsible for the decisions made by AI?
The challenge is therefore not to deploy more and more agents, but to build a truly supervised AI; a single AI for all purchasing, and not a swarm of specialized agents, configured independently of each other, and assembled together without a common governance policy. A single, informed and governed agent, inheriting the permissions of the person they work for. Purchasing expertise is formalized in the form of instructions that the AI follows. Every action is recorded. The instructions define the situations where the AI can act independently, and those where it interrupts to rely on human judgment. The AI executes. Humans guide, arbitrate, supervise and enrich its capabilities.
Your AI is only as reliable as your data
Despite the desire to leverage AI more meaningfully, 74% of procurement leaders say their data is not ready for AI. Inaccurate, inconsistent, insufficient or obsolete data provides fertile ground for risk, especially as
80% of CPOs plan to deploy generative AI over the next three years.
AI agents work with the data provided to them. When supplier data, tender history, contracts, or purchasing transactions are scattered across different software or databases, it becomes almost impossible to define clear decision logic based on precise information. In this case, AI can produce results that appear reliable, but are not in reality because they are based on a partial or incorrect view of the situation.
Compartmented data poses an operational problem and thus breaks the principle of “supervised autonomy”. Purchasing usually follows a relatively linear process: sourcing, contracting, procurement and payment. However, data is often scattered and inconsistently structured. As long as this information remains fragmented, AI cannot have a global and reliable vision. Clearly, the quality of decisions made by AI directly depends on how the data is organized.
Build a reliable supplier file to allow AI to act safely
In many other professions, companies already have a centralized view of the customer. On the other hand, in purchasing, this unified approach on the supplier side has long been neglected. It is, however, the necessary starting point to establish a coherent structured data repository. AI is now highlighting this weakness. Much of the information critical to purchasing passes through suppliers, but many systems still view them as peripheral elements of the process. For AI to be reliable, it is essential to have a single supplier repository that brings together in one space history, obligations, performance, financial exposure, transactions and risks over time. This supplier core must be considered as a fundamental element of the infrastructure. A unified, deduplicated and hierarchical view, where each interaction enriches the same repository instead of creating a parallel one. When designed well, trends emerge naturally. The risks appear earlier. The relationships become readable. Teams stop managing data and finally use it.
Teaching your agents what’s true
The hybrid human-agent model only works if both rely on the same reliable data. Most purchasing teams already have the raw materials, but agents cannot work with raw and inconsistent data.
Agentic AI requires a semantic layer that defines: the meaning of each field, what is included or excluded, how the data is accessed and the conditions under which it can be considered reliable.
A parallel layer of human governance defines what the AI is allowed to do and who is responsible for each action, depending on the situation. The agent inherits access controls from the user it is working with and every action of the agent is recorded, attributable and auditable. Finally, human validation mechanisms define the situations where the AI can act alone and those where it must wait for human intervention.
With these two layers in place, AI operates within the same constraints imposed on your employees.
The dividing line
AI deployed without human supervision ends up becoming unreliable automation. A competitive gap is widening between organizations that operate AI that humans can monitor, audit and improve, and those that deploy AI that operates outside of this quadrant of control. These choices will determine whether agentic AI becomes a trusted decision-maker capable of evolving within a framework of supervised autonomy, or a risk for which no one assumes responsibility.
