In short: Adoption of generative AI in procurement is close to universal in intent: Deloitte's survey of 250+ CPOs found 92% assessing or planning GenAI capabilities. Production deployment is a different story: MIT's NANDA research found that 95% of the enterprise GenAI projects it analysed showed no measurable P&L impact, while IDC data points to strong returns among organisations that get it right. The gap between those two numbers is the real story. This article looks at where GenAI genuinely changes procurement work, why most pilots stall before they reach the income statement, and what separates the minority that convert experiments into operating capability.
Generative AI in procurement has passed the point where the interesting question is whether to adopt it. Deloitte's Global Chief Procurement Officer Survey, drawing on more than 250 CPOs across some 40 countries, found 92% of procurement leaders assessing or planning generative AI capabilities, with around 37% already piloting or deploying.
The harder question is what happens after the pilot. MIT's NANDA initiative examined 300 enterprise AI deployments alongside 150 leadership interviews and found that roughly 95% delivered no measurable impact on profit and loss, with only about 5% achieving meaningful value. It's worth reading that figure precisely: it measures attributable financial impact among the projects analysed, not whether a pilot worked technically or whether users found it useful. But the direction is unambiguous. Nearly everyone is experimenting; comparatively few are shipping.
Meanwhile the upside for those who do get there is real. IDC research sponsored by Microsoft reported an average return of $3.70 for every dollar invested in generative AI, rising to around $10 among leading adopters. That study is vendor-sponsored and surveys adopters, so treat it as the optimistic end of the range. Even so, the spread between it and the MIT finding isn't a contradiction. It reflects how concentrated the returns are among organisations that redesign a process rather than bolt a tool onto an existing one.
GenAI offers real advantages across procurement and supply chain functions, with one caveat: each depends on data quality and process integration rather than on the model itself.
Enhanced decision-making. AI-powered analytics provide deeper insight into supplier performance, analysing thousands of data points to identify patterns in pricing, delivery times, and quality metrics.
Automation of routine tasks. From purchase order processing to invoice reconciliation, GenAI reduces manual processing time substantially, freeing procurement teams for strategic work.
Improved supplier relationship management. Through natural language processing, GenAI analyses years of supplier communications and contract data to surface relationship strengths, potential risks, and opportunities for deeper collaboration.
Predictive capabilities. By processing historical and real-time data, GenAI forecasts supply chain disruptions in advance and generates mitigation options based on past interventions.
Sustainability optimisation. AI models factor complex sustainability metrics into procurement decisions, aligning sourcing with ESG obligations that increasingly extend several supplier levels deep.
This is where the MIT findings become useful rather than discouraging. The research points to failure causes that sit upstream of the technology: data foundations that were never production-ready, unclear ownership of AI initiatives, and organisational resistance to changing how decisions actually get made. Model quality is rarely the binding constraint.
Three patterns recur among procurement teams that stall.
The data problem is deferred, not solved. A pilot runs on a curated dataset someone cleaned by hand for the demo. Production runs on the actual supplier master, with inconsistent naming across departments, contracts stored as email attachments, spend data split between ERP and spreadsheets. The pilot works; the rollout doesn't.
No baseline was captured. A significant share of "no measurable impact" findings reflects the absence of a measurement, not the absence of a result. If nobody recorded cycle times, error rates, or maverick spend before the pilot, there is no way to demonstrate the improvement afterwards, which makes funding the next phase considerably harder.
Individual use outpaces process change. Team members adopt AI tools personally to draft emails or summarise documents, which raises individual productivity without altering the workflow. That produces genuine time savings and no line on the P&L, because the underlying process still runs the way it always did.
As companies integrate GenAI into procurement, supplier relationships shift in three ways:
These depend on suppliers participating in the same data environment. Where supplier information arrives as PDFs and email threads, the analysis has nothing consistent to work with. That is why supplier onboarding and data standardisation tend to precede any useful AI capability rather than follow it.
The MIT research and the Deloitte data point in the same direction on priorities. Four areas matter more than tool selection.
Fix the data foundation before funding new use cases. Audit data quality, integration, and governance, and resolve the critical gaps first. This is unglamorous and it is the single most consistent differentiator in the research.
Assign clear ownership. Define who monitors performance, who acts on the output, and who escalates when something breaks. Initiatives without an owner drift regardless of how good the underlying technology is.
Measure before you build. Establish the baseline (cycle times, error rates, spend under management) so improvement is demonstrable rather than asserted.
Build capability alongside tooling. Upskilling procurement teams and sustaining a culture of continuous learning determines whether a deployed tool actually changes how work gets done.
Procurement leaders who get these right can move their function from cost centre to value driver. The question is no longer whether to adopt GenAI, but whether an organisation can move from experimentation to production before the advantage is gone.
For procurement leaders weighing where to begin, Fluenta One offers a practical starting point that delivers measurable results while building the foundation for broader GenAI integration, including AI agents embedded directly in procurement workflows.
How widely is generative AI actually adopted in procurement?
Adoption of intent is near-universal: Deloitte's survey of over 250 CPOs across roughly 40 countries found 92% assessing or planning GenAI capabilities, with about 37% piloting or deploying. Production-grade deployment delivering measurable financial results is far rarer, and that is the gap most procurement teams are currently sitting in.
Does the MIT "95% of AI pilots fail" statistic mean AI doesn't work in procurement?
No, and the figure is often misquoted. MIT's NANDA research found that around 95% of the 300 enterprise AI projects it analysed showed no measurable P&L impact. That's a statement about attributable financial results, not about whether the technology functioned or whether users benefited. A project can improve work meaningfully and still register nothing measurable if no baseline was captured beforehand.
What is the realistic return on GenAI investment in procurement?
Estimates vary widely and depend heavily on the source. IDC research sponsored by Microsoft reported an average $3.70 return per dollar invested, reaching roughly $10 among leading adopters. As a vendor-sponsored study of adopters, that sits at the optimistic end. Set against MIT's finding, the honest reading is that returns are real but highly concentrated among organisations that redesign processes rather than layer tools on top of them.
Why do procurement AI pilots stall before production?
Most commonly because of data foundations rather than model quality. Pilots run on hand-cleaned demo datasets while production has to handle inconsistent supplier records, contracts stored as email attachments, and spend data split across systems. Unclear ownership and the absence of a pre-pilot baseline compound the problem.
What should a procurement team do first?
Audit data quality, integration, and governance before funding new AI use cases; assign explicit ownership for each initiative; and capture baseline metrics (cycle times, error rates, spend under management) so improvement can be demonstrated. Tool selection matters considerably less than these foundations.
.png)
.png)
.png)
.png)
.png)