Agentic AI in procurement: what changes for the CPO
A copilot drafts, suggests, and waits for a person to click; an agent takes a goal, works across systems, and completes the task without step-by-step prompting. For a CPO, that distinction changes two things beyond the tooling itself: governance has to account for decisions a system made on its own, and the measure of success shifts from adoption to exceptions resolved end to end.
An agent acts on a goal, a copilot waits to be asked
Most procurement teams already have some AI: a copilot that summarizes contracts, scripts that move fields between systems. All of it needs a person to initiate every action and check every output. An agent is different in one specific way. Give it a goal – clear this queue of blocked POs, resolve these invoice holds – and it gathers context from the ERP, the supplier portal, and email, decides on the correct resolution, and executes it, escalating only the cases it cannot settle. This capability is arriving fast: Gartner predicts 40% of enterprise applications will be integrated with task-specific AI agents by the end of 2026, up from less than 5% today. The practical differences from older automation are covered in how agentic AI differs from RPA.
Governance has to cover decisions a system made on its own
When a copilot suggests and a person approves, existing controls still work: the person is accountable and the audit trail points at them. When an agent releases an invoice hold or amends a PO on its own, auditors get new questions. What data did it read, what policy did it apply, why this resolution, and who set the thresholds for when it acts alone versus escalates? A CPO adopting agents needs those answers written down before the first audit: decision logs at the level of individual actions, explicit autonomy boundaries by exception type and dollar value, and a named owner for the agent's policy, the same way every approval queue has a named owner today.
The metric that matters is exceptions resolved end to end
Copilot programs get judged on adoption: logins, weekly active users. Those numbers say nothing about whether work got done. The metric that survives contact with an agent is throughput: how many exceptions were resolved end to end without a person re-doing the work, at what accuracy, at what escalation rate. An escalation rate above target usually means thresholds were set conservatively, which is fixable. A rework rate above target means the team is spot-checking everything and the agent has added a review step instead of removing one. Pricing follows the same logic: outcome-based terms fit a system judged on resolution, while per-seat licensing assumes adoption was the goal. For a grounded view of where this already performs, see what's actually working in AI for procurement today.
Most agentic AI failures are scoping failures
The same firm forecasting rapid agent adoption also predicts that over 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value, and inadequate risk controls. None of those reasons concern model capability. Projects fail when they start with "deploy agents" instead of a specific queue with a measurable backlog, or when the governance questions above surface mid-deployment and stall everything. The CPOs likely to land in the successful majority pick one exception-heavy workflow with a clear owner, define resolution and escalation rules up front, and expand only after the audit trail has survived a real quarter.
If you want to see what a scoped deployment looks like in practice, Fragment publishes its source-to-pay workflow catalog – from req-to-PO conversion and change orders through invoice exceptions, GL coding, and supplier master data – and you can walk through a live exception with the team. Fragment's agents work across your existing ERP, portals, and email with nothing ripped out or replaced, human-in-the-loop by default, and typically reach production in weeks.