Agentic AI Can Transform Procurement—But the Operating Model Comes First

HBR is right: procurement is unusually well suited to agentic AI. My push is that technology is not the first move. Leaders should redesign the team, challenge the commercial model, and write down decision rights before they scale an agent.

HBR is right. I would go further.

When Harvard Business Review published “Why Agentic AI Could Transform Procurement” on August 13, I agreed with the premise. The authors, including BCG leaders and an academic, argue that procurement is unusually agent-ready because the work is structured, financially measurable, and full of judgment-intensive tasks.

The evidence supports part of that case. Gartner names contract management, project scoping, supplier recommendations, and RFx drafting among emerging workflow uses. Peer-reviewed case research places spend classification among the advanced analyses supporting strategic purchasing decisions. If agents can absorb more of that preparation, buyers get more room for judgment. But that is only the technology case.

1. Change the org chart before you deploy the agent

BCG's own 2026 study surveyed more than 200 CIOs, procurement leaders, and technology-procurement buyers across North America, Europe, and Asia-Pacific. Its authors say organizations combining technology deployment with capability building, process redesign, and governance reported stronger outcomes than those focused on technology alone. That is a self-reported association, not proof. The direction is still worth taking seriously.

Drop agents into the same category silos, the same approval ladders, and the same deck-building handoffs, and you get a faster version of the old system. Useful? Maybe. Transformational? No.

My view is that a compact central team should own the agent platform, controls, standards, and reusable workflows. It should not own every category outcome. Category and cross-functional owners keep strategy, specifications, supplier selection, and awards. Agents carry more of the preparation, analysis, and monitoring. People concentrate on exceptions, commercial judgment, negotiation, and relationships.

2. Challenge pricing built around the old workflow

If a platform contract is priced by seats or modules, leaders should ask whether that unit still matches value when agents execute more of the workflow. A seat may remain simple to sell and budget. It is not automatically the natural unit of value. The better question is what work was completed, with what evidence, and who owns the result.

Gartner's warning matters here: many organizations are experiencing uneven ROI or falling short of expectations. It also says fragmented and low-quality data across procurement systems can hinder accurate outputs, while integrating stand-alone GenAI solutions with existing platforms is often complex. Buying an agent add-on for an unchanged stack does not remove those constraints. It can simply make the old architecture more expensive.

• What are we paying for: access, an executed workflow, or a verified outcome?
• Which baseline, volumes, and exclusions define that outcome?
• Who pays for bad inputs, rework, or exceptions outside the vendor's control?
• Who owns the enriched supplier, contract, and spend data—and can we take it with us?
• What audit rights, service levels, and termination support protect the buyer?

3. Write decision rights before you ask for trust

The durable model is hybrid, not hands-off. The spend-classification research proposes combining AI processing with human validation and adjustment. For systems classified as high-risk, the EU AI Act requires that they can be effectively overseen by natural persons while in use, including the authority not to use the system in a particular situation. I am not claiming ordinary procurement tools fall into that legal class. The useful operating principle is simpler: a person must be able to see, challenge, and stop the action.

Make that boundary operational. Let agents draft, normalize, monitor, and recommend inside declared policy. Keep awards, negotiation strategy, risk acceptance, financial commitments, and supplier relationships with named people. Our guide to AI in procurement maps that execution-versus-decision boundary across the full cycle.

• Execute: name the permitted data, systems, policy, thresholds, required evidence, and fallback before the agent acts.
• Recommend: name the accountable person, the review they perform, and the record retained when they accept or override the output.
• Approve or stop: reserve money, risk, strategy, supplier-facing commitments, and system shutdown for the authorized owner; preserve segregation of duties and escalation.

Trust follows control. Not the other way around.

A Kinaxis-sponsored IDC study, reported by CSCMP's Supply Chain Quarterly, makes the gap visible. Just 6% of respondents described their supply chains as autonomous at scale, while 41% expected that to become their core operating model within one to two years. In the same survey, 52% named trust in AI-driven decisions as a top barrier. The study covered more than 2,000 supply-chain leaders across nine markets, a broader population than procurement alone.

I would not put those expectations into a budget. Expectations are not deployments. But I would take the gap seriously. Trust does not arrive because the next model is better. It grows when everyone knows what the agent may do, what evidence it must leave behind, and who owns the decision.

So what actually changes?

What I would do on Monday

1. Choose one low-risk workflow. Start with historical replay or an approved shadow run alongside a live event. Complete the required security, privacy, legal, records, and category-owner reviews before using sensitive data.
2. Map every handoff. Mark where information is copied, summarized, normalized, checked, communicated, and approved.
3. Write the operating boundary. For each step, record whether the agent may execute, recommend, or must wait; name the owner, inputs, threshold, evidence, override, and fallback.
4. Reopen the commercial terms. Ask what you pay for when seats and manual touches fall, and how your enriched data leaves with you.
5. Take the baseline before the pilot. Record cycle time, rework, exception rate, and the share of output accepted without substantive correction. No baseline, no ROI story.

Do not buy an agent to preserve the workflow you already have. Scope the workflow, baseline, decision-rights matrix, and shadow plan in days. Move to production only after the controls and approvals match the data, supplier interaction, financial authority, and risk involved. Fast is good. Undocumented authority is not.