The procurement industry’s appetite for agentic AI has never been stronger. According to The Hackett Group’s Agentic AI in Procurement Adoption Index — 2026 (Inaugural Edition), 64% of procurement leaders believe agentic and generative AI will fundamentally reshape procurement workflows by 2030, and nearly half had already run pilots by the end ofw 2024. (Source: The Hackett Group Agentic AI in Procurement Adoption Index 2026, via Zycus)
Yet the gap between pilot enthusiasm and production-scale impact remains stubbornly wide. Gartner predicts that through 2026, organizations will abandon 60% of AI projects unsupported by AI-ready data — and its Q3 2024 survey of 248 data management leaders found that 63% of organizations either do not have, or are unsure whether they have, the right data management practices in place to support AI. (Source: Gartner, Lack of AI-Ready Data Puts AI Projects at Risk, February 2025)
Gartner’s Kaitlynn Sommers, Senior Director Analyst in Gartner’s Supply Chain Practice, has been direct on the procurement-specific dimension: “Fragmented and low-quality data across procurement systems can hinder accurate outputs, and integrating stand-alone GenAI solutions with existing platforms is often complex due to differing technical specifications.” (Gartner Press Release, July 2025)
Technology is not the primary failure point. The data underneath it is.
Why Agentic AI Is More Data-Dependent Than Previous Procurement Technology
Traditional procure-to-pay automation follows explicit rules — approve if invoice matches, escalate if spend exceeds threshold. These systems are brittle but forgiving of data imperfection: the consequences of bad data are localized to a single failed rule or manual exception.
Agentic AI operates differently. It sets goals, plans multi-step workflows, and executes autonomously across the source-to-pay lifecycle — identifying sourcing opportunities, evaluating supplier bids, drafting contract terms, and monitoring compliance without a human trigger at each step. When an agentic system reasons through a sourcing decision, it draws simultaneously on supplier master data, historical spend records, active contracts, market benchmarks, and category performance data. If any of these inputs are fragmented, duplicated, or inconsistently classified, the agent’s reasoning is compromised — not at one step, but at every downstream step that depends on the corrupted input. The error does not stay localized. It propagates.
The Specific Data Problems That Break Procurement AI Agents
Procurement data environments are structurally prone to exactly the failure conditions that agentic AI cannot tolerate.
Fragmented supplier master data is the most common and most damaging. In organizations that have grown through acquisition or operate across multiple regions, the same supplier may exist under different names across multiple ERP systems, each with different payment terms, risk ratings, and compliance records. An agentic system tasked with supplier consolidation or risk assessment cannot reason across these records reliably — it sees fragmentation where there should be a single entity.
Inconsistent spend classification creates a second failure mode. When category taxonomies differ between business units, or when manual classification has introduced systematic errors over years of operation, agentic systems built on top of that data inherit the errors as structural assumptions. The agent does not know the classification is wrong — it optimizes against a flawed picture of what is actually being spent and with whom.
Siloed contract repositories compound the compliance monitoring challenge. If active contracts are distributed across legal systems, business unit folders, and legacy CLM platforms, an agentic compliance agent cannot enforce what it cannot see. The Hackett Group’s 2026 Adoption Index identifies data fragmentation across systems as a structural obstacle to procurement AI scaling beyond pilot. (Source: The Hackett Group Agentic AI in Procurement Adoption Index 2026, via Zycus)
These are not edge cases. Deloitte’s 2024 Global CPO Survey found that while 92% of CPOs were beginning to envision the possibilities of agentic AI, only 37% were actually piloting or deploying it — with data quality cited as one of the most significant internal barriers. (Source: Deloitte, Data Standards and GenAI in Procurement, 2024) The IBM Institute for Business Value found that 49% of executives cited data inaccuracies and bias as a specific barrier to embracing agentic AI. (Source: IBM, Data Quality Issues, 2025)
The Compounding Risk of Deploying Agents on Dirty Data
When a rule-based RPA bot encounters bad data, it fails visibly — an exception is thrown, a process stalls, a human is notified. When an agentic system encounters bad data, it may not fail visibly at all. It reasons through the problem using the available data, produces an output that appears coherent, and executes autonomously on that output. The error surfaces later — in a sourcing decision that missed the preferred supplier list, a compliance flag that never fired, or a spend report that misrepresents category performance.
This is precisely the governance risk behind Gartner’s prediction that over 40% of agentic AI projects will be canceled by end of 2027 — not because agents stopped working, but because outputs could not be trusted or defended. (Source: Harvard Business Review, Why Agentic AI Projects Fail, 2025) Deloitte’s 2025 Global CPO Survey reinforces the organizational dimension: siloed working was identified as the top barrier to procurement AI value delivery, cited by 57% of CPOs. (Source: Art of Procurement, State of AI in Procurement, 2026, citing Deloitte) Siloed data and siloed organizational structures co-evolve — and both must be addressed for agentic AI to operate reliably at scale.
How to Build the Data Foundation That Makes Agentic AI Work
Organizations do not need data perfection before deploying agentic AI — they need data readiness for the specific use cases they are prioritizing. APQC’s research found that eight out of ten organizations implementing AI in procurement experienced improved data quality as a result, suggesting AI deployment and data improvement can be mutually reinforcing when sequenced correctly. (Source: Art of Procurement, State of AI in Procurement, 2026, citing APQC)
Establishing a single source of truth for supplier master data is the non-negotiable starting point. The concrete action is a governance decision — not a technology project: define where the authoritative supplier record lives, which system owns it, and how all others reference it. The outcome indicator is measurable and immediate: duplicate supplier entries across systems should be declining within 90 days. Zycus’ Supplier Management capability provides the centralized supplier data layer that the Merlin Agentic Platform draws on when executing sourcing, compliance, and risk workflows.
Standardizing spend classification is the second prerequisite. The concrete action is a category taxonomy audit — identify the top 20 categories by spend volume, establish a single agreed classification standard for each, and use that as the training baseline for AI-assisted classification. The outcome indicator is spend coverage: the percentage of total addressable spend accurately classified at category level, which should improve within the first quarter of deployment. Zycus’ Spend Analysis applies this logic continuously, keeping the spend data feeding agentic decisions consistent and governed.
Connecting contract data to purchasing execution is the third requirement. The concrete action is an integration audit — map which contracts are currently visible to the purchasing system at point of transaction, and identify the gap between contracts held in legal systems versus those actively enforced at point of purchase. Closing even 50% of that gap, prioritizing high-value categories first, creates a meaningful compliance monitoring surface. The outcome indicator is contract compliance rate by category, measurable within one to two sourcing cycles.
Defining governance before scaling autonomy is the fourth component. The concrete action is a governance map: for each agentic use case, define which decisions the agent executes autonomously, which it recommends for human approval, and which it escalates without acting — documented before deployment, not retrofitted after the first exception. The outcome indicator is audit trail completeness: the percentage of agentic actions that are logged, attributable, and reviewable. The Hackett Group’s 2026 Adoption Index is clear that organizations building agentic AI at scale treat governance architecture as a prerequisite, not an afterthought. (Source: The Hackett Group Agentic AI in Procurement Adoption Index 2026, via Zycus)
The Strategic Implication for Procurement Leaders
The data readiness imperative is not a reason to defer agentic AI investment. It is a reason to sequence it correctly. The Hackett Group’s 2026 research projects a 9–10% growth in procurement workloads against just 1% budget growth in 2025 — a gap that cannot be closed by headcount and cannot be closed by agentic AI deployed on fragmented data. (Source: The Hackett Group Agentic AI in Procurement Adoption Index 2026, via Zycus)
Gartner’s finding that 74% of procurement leaders say their data is not AI-ready is a measure of the gap, not a verdict on the technology’s potential. (Source: Art of Procurement, State of AI in Procurement, 2026, citing Gartner) The 26% who have done the foundational work are already deploying at scale — and the distance between their outcomes and those of the remaining 74% is widening each quarter.
The path to agentic AI that works in procurement runs through data. That is not a caveat. It is the design condition.
