Why Procurement Is Lagging on Agentic AI—and How Leaders Can Catch Up

Procurement’s 9% Agentic-AI Problem

Agentic AI is spreading unevenly across enterprise functions. A survey of 385 organizations cited in a new Harvard Business Review analysis puts adoption at 35% in software development, 31% in IT operations and 26% in marketing—but only 9% in procurement.

That gap is notable because procurement has the clearest structural fit for autonomous agents: high-volume workflows, measurable financial outputs such as supplier terms and spend levels, and persistent manual friction in document validation, risk assessment and category classification.

The analysis draws on more than a decade of procurement technology change and interviews with procurement leaders and technology providers. It describes production deployments, including a large European energy buyer that uses AI agents to qualify 30,000–40,000 supplier applications a year, and an automotive OEM that now handles roughly 70% of procurement work through a chat-based agent orchestration interface.

The central finding is that the constraint is organizational, not technological. Most companies are trying to layer agentic AI onto existing operating models rather than redesigning those models around it—and that is why pilots often stall.

Why Procurement’s Strongest AI Case Hasn’t Scaled

The Energy Player’s Proof Point

The article’s most detailed case is a major European energy company with a 25-person team managing 18,000 active suppliers and 30,000–40,000 new applications a year. Each application can include up to 25 documents and more than 20 qualification steps, creating a structural backlog and reputational risk from missed compliance signals.

After deploying a set of AI agents in 2025, the company cut its backlog by 20% within six months. First-pass accuracy rose from 60% to between 70% and 80%, with a trajectory toward 95% at scale. One agent catches wrong category selections—an error in about 30% of applications—while others prioritize the queue by business need and scan structured and unstructured sources for tax, labor, fraud and executive-risk signals.

This is an interpretation, not a guarantee: the payoff came from embedding agents into a specific high-volume process and accepting human oversight rather than replacing the team.

The Automotive OEM’s Commercial Redesign

A major automotive original-equipment manufacturer took a different route. It began with text-based generative AI tender assistants in 2023, found the time savings could not be monetized inside legacy processes, and in 2024 rebuilt its procurement lifecycle around commercial outcomes—negotiation uplift, captured savings and EBIT impact.

Roughly 70% of procurement work now flows through one chat-based interface, with an orchestrator coordinating knowledge retrieval, tool execution and system integration. A controlled three-month A/B test produced speed gains, measurable negotiation uplift and a material EBIT contribution, measured with the same metrics as ordinary procurement performance.

Three Organizational Blockers

The analysis identifies recurring barriers. First, accountability and control: procurement decisions carry legal and financial consequences, so companies restrict agent autonomy to the point of neutralizing the benefit. Second, fragmented ownership: IT or central digital teams design the tools, while procurement remains a beneficiary rather than an owner, producing pilots that scale poorly. Third, data as an afterthought: inconsistent supplier master data, incomplete spend categorization and fragmented ERP records undermine agents before they can perform.

A related failure is the “belief stage” problem: local productivity gains from tender assistants are celebrated, but surrounding processes are not redesigned to capture the freed-up time, so accountability for value stays unclear.

The Common Thread

The companies pulling ahead make procurement the owner, target the highest-pain and highest-value processes rather than safe pilots, treat data as infrastructure, design graduated autonomy, build reusable “agentic factories,” engineer governance from the start, and measure commercial outcomes rather than adoption. The article’s core claim is that these choices compound—and that late entrants will find it increasingly hard to catch up.

Seven Procurement Moves from the Fast Adopters

• Make procurement the owner, not the beneficiary. In deployments that scale, procurement defines the use case, owns success metrics and holds accountability; IT enables but does not dictate. The trade-offs should be made by the people who live with the consequences.
• Start where the economic pain is greatest. Target supplier onboarding, RFP generation and compliance monitoring—not low-risk pilots—and focus on direct materials and complex services, where value concentration is higher than in tech-procurement tail spend.
• Treat supplier master data and spend taxonomy as infrastructure. Fix data quality before launching agents, accepting that this will slow the first visible deployment; it is the cost of compounding advantage rather than a retrofit after failure.
• Design graduated autonomy, not a binary switch. Move agents from advisory roles to human-in-the-loop decisions to supervised execution, keeping high-stakes activities such as negotiation in a coaching role where judgment remains human.
• Build capability, not isolated use cases. Create repeatable “agentic factories” with reusable components and shared governance so the team that builds the first deployment carries it into the next portfolio of agents.
• Build governance in from the start. Specify approval rights, audit trails, escalation paths and override mechanisms before go-live, and require every agent to have a clear investment case mapped to a procurement phase and a defined value lever.
• Measure commercial outcomes, not adoption. Track captured savings, negotiation uplift, cost avoidance, EBIT contribution and decision speed; treat cycle-time reduction, chatbot usage and pilot adoption as secondary.