AI Agents in Procurement: Who Is Accountable When Machines Make Sourcing Decisions?

The decision no one remembers making

Imagine that an AI agent monitoring a company’s supply base detects a sudden deterioration in one supplier’s financial position. It reviews open orders, checks alternative capacity, calculates the switching cost and recommends moving 30 per cent of future volume to a competitor. The recommendation reaches a category manager, who approves it between meetings. Three weeks later, the incumbent supplier challenges the decision. Its data was incomplete, the alternative supplier cannot meet a critical specification and the relationship damage is already real. Who owns the mistake—the category manager, the system, the data provider, the software vendor or the executive who authorised the technology? That question is about to become less hypothetical. Procurement has spent years adopting tools that analyse spend, draft requests for proposal and flag supplier risk. The next generation of systems is designed not merely to advise, but to pursue goals, coordinate tasks and act within defined boundaries. The attraction is obvious. Procurement teams are overloaded, data is fragmented and business stakeholders expect faster answers. An agent that can assemble a sourcing event in hours rather than weeks sounds less like science fiction and more like overdue help.

From assistance to delegated authority

There is an important difference between generative AI and agentic AI. A conventional assistant may summarise a contract when asked. An agent can monitor events continuously, decide that a response is required and initiate the next step. In procurement, that might mean inviting suppliers, requesting missing documentation, recommending an allocation, escalating a risk or negotiating within pre-approved limits. The World Economic Forum has described agentic supply chains as networks in which autonomous systems perceive, reason, negotiate and execute decisions across organisational boundaries. Once software begins interacting with supplier systems, logistics providers and financial platforms, it is no longer simply producing content. It is influencing the movement of goods, money and commercial commitments. Deloitte’s 2025 Global CPO Survey found procurement leaders increasing their focus on generative and agentic AI as they face growing complexity with constrained resources. Yet efficiency is only half the story. Every task delegated to an agent also transfers a degree of operational authority.

Not every procurement decision is a transaction

Procurement decisions often contain context that does not sit neatly in a dataset. A supplier may be temporarily more expensive because it invested in capacity for the buyer. A technically weaker bid may carry lower implementation risk. A small local supplier may matter to a national-content commitment. A contract clause that appears favourable may be unenforceable in practice. This is why the debate cannot be reduced to whether AI is accurate. Even a statistically sound recommendation can be commercially unwise. Procurement involves trade-offs between price, continuity, innovation, reputation, relationships and risk. Those trade-offs require judgement—and judgement requires someone to be answerable for the outcome.

Build boundaries before building agents

The sensible response is not to keep AI away from procurement. It is to define authority before deployment. Low-risk, reversible actions can be automated more confidently: collecting documents, checking fields, monitoring expiry dates or drafting routine communications. Decisions that commit spend, exclude a supplier, change an allocation or accept contractual risk require tighter thresholds and explicit human approval. Leaders should record what the agent saw, which sources it used, what rule or objective it followed and who approved the result. Permissions should be narrow rather than convenient. An agent that can read contracts does not automatically need the ability to issue a purchase order. The World Economic Forum’s 2026 cybersecurity outlook warns that AI agents can accumulate excessive privileges or propagate errors at scale when governance is weak. Procurement also needs a route for suppliers to challenge consequential automated decisions. Transparency will become part of good supplier governance, particularly when AI is used to screen, rank or disqualify vendors.

The CPO remains accountable

There will be pressure to describe failures as technology problems. That would be a mistake. A company cannot delegate responsibility simply because it has delegated activity. The CPO remains responsible for the operating model: what the agent may do, where a person must intervene, how exceptions are handled and whether the system’s incentives reflect the organisation’s values. The best procurement functions will not be those that automate the most. They will be those that know precisely what should be automated, what should be augmented and what should remain a human decision. The real leadership question is not whether machines can make sourcing decisions. It is whether the organisation has decided which decisions it is prepared to let them make.