AI in procurement: why 95% of deployments fail to deliver ROI

Procurement has the highest weekly generative AI adoption rate of any business function: 94% of procurement executives now use GenAI at least once a week, according to the AI at Wharton 2024 survey. Yet MIT Project NANDA, published July 2025, found that 95% of organizations deploying generative AI saw zero measurable return.

Not low return. Zero. The gap between adoption and value is not a technology problem. It is a deployment problem, and it follows a pattern that repeats across industries, toolsets, and organization sizes. Understanding that pattern is the first step to breaking it.

---

The adoption bubble and the return vacuum

Procurement AI adoption surged from 50% weekly usage in 2023 to 94% in 2024 — the largest increase of any business function surveyed. By early 2026, 73% of procurement organizations were either piloting or actively scaling AI, up from 28% in 2023. Deloitte's 2024 CPO survey found 92% of respondents planning or assessing GenAI capabilities. Gartner reported 73% of procurement leaders expected to adopt the technology by end of 2024.

Yet the same period produced sobering results. RAND Corporation's analysis of 2,400+ enterprise AI initiatives found that 80% of AI projects fail to deliver their intended business value — twice the failure rate of regular IT projects. Gartner predicts 60% of AI projects unsupported by AI-ready data will be abandoned through 2026, a rate that is already at 42% of U.S. companies. A 2025 PwC survey found only 12.5% of CEOs reported AI delivered both cost savings and revenue growth.

---

The six root causes that repeat

Across every study surveyed — MIT NANDA, RAND, Gartner, McKinsey, BCG — the same six root causes appear. None of them involve model capability.

• Undefined success criteria. 73% of failed AI projects had no agreed definition of success before the project started. Worse, 61% of enterprise AI projects were approved on projected ROI that was never measured after launch.
• Fragmented data, no quality pipeline. Only 12% of organizations report data of sufficient quality and accessibility for AI applications. 73% of data leaders identify data quality and completeness as the primary barrier — ranking it above model accuracy and computing costs.
• Integration failure. 88% of procurement leaders cite integration issues as a barrier to AI confidence. The model works in isolation. It fails when it must connect to the systems procurement actually uses.
• Siloed talent. Data science and procurement teams work in parallel, not together. The procurement SME who knows what data matters is not in the room when the model is designed.
• Technical debt from prototype code. Pilots that "work" in a Jupyter notebook are pushed to production without observability, monitoring, or guardrails.
• No post-deployment evaluation. Once deployed, models drift. Without continuous evaluation, the first sign of failure is a procurement stakeholder who stopped trusting the output and never told anyone.

---

What separates the 5% that deliver

The organizations that report measurable return share three structural characteristics. They do not deploy AI differently. They prepare for it differently.

Deloitte's 2025 Global CPO Survey identifies "Digital Masters" — the top quartile of procurement organizations by tech maturity — who achieve 3.2x ROI on generative AI investments, compared with 1.5x for followers. These organizations commit 20% or more of their digital budgets to AI and invest 70% of AI resources in people and processes, not just the technology layer.

McKinsey's 2025 global survey found that early AI adopters report $3.70 in value for every dollar invested, with top performers achieving $10.30 returns per dollar. The common factor: they started with the data, not the model.

---

The data readiness gap that kills most deployments

Informatica's 2025 CDO Insights survey found that data quality and readiness is the number-one obstacle at 43%. Only 12% of organizations report data of sufficient quality for AI applications. 92.7% of executives identify data as the most significant barrier to AI implementation, according to NewVantage 2024 data.

This is not a procurement-specific problem, but procurement feels it acutely because procurement data is notoriously fragmented: across ERP systems, P2P platforms, supplier portals, contract management tools, and spreadsheets. A Capital One/Forrester survey of 500 enterprise data leaders found that 73% identified "data quality and completeness" as the primary barrier to AI success, ranking it above model accuracy, computing costs, and talent shortages.

The organizations that will report AI-driven P&L gains in 2026, according to multiple studies, are the ones that stopped running new pilots and started fixing their data foundation first.

---

What this means in practice

For procurement leaders evaluating AI investments, the data supports a specific sequence of actions, not a list of vendors to evaluate.

1. Audit data readiness before any pilot. Run a structured assessment of data quality, completeness, and accessibility. If fewer than 30% of your spend categories have clean, structured data, do not deploy AI into them. Fix the data first. Expected outcome: a data roadmap that replaces the AI roadmap. Timeframe: 8-12 weeks.
2. Define a single measurable outcome per deployment. Not "improve efficiency" but "reduce PO processing time from 4 days to 1 day" or "increase contract compliance to 85%." Without a specific number, you cannot measure success. Expected outcome: a threshold for go/no-go decisions. Timeframe: before any code is written.
3. Allocate 70% of budget to process and people. The technology is the smallest cost. The work is in workflow redesign, change management, training, and governance. Budget accordingly. Expected outcome: deployment that survives the first 90 days. Timeframe: ongoing.
4. Plan for post-deployment evaluation before deployment. Define how you will measure drift, accuracy, and stakeholder trust. Assign an owner who reports quarterly on AI performance. Expected outcome: a model that degrades visibly instead of silently. Timeframe: before production go-live.
5. Run one deployment to production before starting the next. The organizations with the best results do fewer things better. One category, one workflow, one measurable outcome. Prove value, then expand. Expected outcome: a template for the next deployment. Timeframe: 60-90 days per cycle.

---

Frequently asked questions

---

Sources

1. Why 95% of AI Projects Fail and How Data Fixes It — SR Analytics, February 2026
2. Why 90% of Enterprise AI Implementations Fail — Talyx, 2026
3. AI Project Failure Rate in 2026 — Folio3 AI
4. State of AI in Procurement in 2026 — Art of Procurement
5. Generative AI in Procurement — Deloitte CPO Survey
6. The 2026 State of AI in Procurement — Global Survey Report
7. 200+ AI Statistics & Trends for 2025
8. 90% of Procurement Leaders to Adopt AI Agents — Icertis/ProcureCon