A CPO asks her AI agent how much procurement saved last quarter. The answer is instant, confident and wrong. The agent counted negotiated savings, but finance books realised savings.
The model was not the problem. The missing context was.
The AI ROI gap is a context gap
BCG finds only 5% of companies are achieving AI value at scale. In procurement, the bar is clear: in Ardent Partners research co-sponsored by Zycus, 61% of CPOs said clear and proven ROI is what moved AI from pilot into production.
What separates the winners is rarely the model. Gartner predicts that by 2027, organizations that prioritize semantics in AI-ready data will increase their agentic AI accuracy by up to 80% and reduce costs by up to 60%. Semantics means business context. And in procurement, context is harder to get right than anywhere else.
Why procurement is the toughest test for enterprise AI?
Take a question as basic as “Who are our top 10 suppliers?” Finance means top 10 by spend. The category team means top 10 by strategic value. Risk means top 10 by exposure. A human analyst knows who is asking and adjusts. An AI agent, unless it has been told, simply picks one.
The same problem runs through every core procurement term:
- Savings can mean negotiated, realised or cost avoidance.
- Spend can be total or addressable, and dated by PO or by invoice.
- Supplier can mean a parent company or each of its dozen vendor IDs in the master.
- On-contract depends on which contracts, which price lists and which tolerances apply.
Procurement leaders already feel this. In Deloitte’s CPO research, data quality emerged as the second major internal barrier to AI in procurement. KPMG found that 64% of leaders cite data quality as a barrier to Gen AI strategy. And The Hackett Group’s Agentic AI in Procurement Adoption Index 2026, co-authored with Zycus, rates procurement’s readiness on data quality, governance and master data management as low.
Clean data matters. But clean data without shared meaning still produces confident wrong answers.
What “context” actually means in Source-to-Pay?
For AI in procurement, context has three layers.
- Knowledge: what things are: Your category taxonomy, supplier master, contract terms, price agreements, rebate thresholds and the certified definitions of every metric you report on. Knowledge answers “what does this mean here?”
- Expertise: how work gets done: Your sourcing playbooks, category strategies, the way exceptions actually get resolved, who to escalate to and when. Expertise answers “what would a good buyer do next?”
- Norms: what is allowed: Delegation of authority thresholds, preferred supplier policies, three-way match tolerances, sanctions screening, ESG requirements. Norms answer “am I permitted to do this?”

For decades, experienced buyers carried this context in their heads. AI agents cannot. What was implicit must now be made explicit and machine-readable.
That makes ownership critical. As Forrester puts it in Don’t Delegate AI, a study commissioned by Zycus, AI “encodes procurement judgement into systems.” If procurement does not define that judgement, someone else will. The same study found that 42% say procurement defines AI vision, while 32% still defer to IT.
Gartner frames it architecturally, naming three foundational components for reliable agentic AI: semantics, operational state, and provenance. In procurement, where agents increasingly move real money, provenance (knowing why every decision was made) is non-negotiable.
The agent sprawl trap: welcome to the Shadow Agent Estate
As agents get easier to build and buy, procurement organisations are accumulating them fast. An intake agent here, an invoice agent there, a sourcing assistant from one vendor and a contract reviewer from another.
Zycus and The Hackett Group have a name for what builds up: the Shadow Agent Estate. A Shadow Agent Estate is the population of AI agents running across your enterprise that no one has inventoried, owned, or governed. Its liability side is agent debt: the compounding cost of agents that each do their task but cannot share context, coordinate or be audited together.
Shadow estates do not form through recklessness. They form because isolation is the default output of every path to agent creation unless someone explicitly fights against it. Zycus learned this first-hand. A Shadow Agent Estate still formed inside our own organisation, not through one bad decision, but through a long run of reasonable ones made without a governing architecture to contain them.
The risk is widely felt, and the controls are lagging. The Hackett Group’s 2026 Future-Ready Purchase-to-Pay Quick Poll found that 69% of companies expected additional material risk from agentic AI deployment compared to traditional automation. Yet in the Hackett Adoption Index, only 24% have defined KPIs, and only 19% have governance infrastructure in place.
This is where context and sprawl collide. If your intake, sourcing and AP agents each carry their own definition of “supplier” or “on-contract”, you have not automated procurement. You have automated inconsistency at scale.
Gartner expects that more than 40% of agentic AI projects will be cancelled by the end of 2027 due to escalating costs, unclear business value or inadequate risk controls. Ardent Partners draws the architectural line: standalone agents often lack data breadth and cross-module orchestration, while embedded agents leverage common data models, unified UX, and coordinated release cycles for consistent execution at scale.
More agents is not the answer. A shared foundation of context that every agent draws from is.
Context built in, not bolted on: Merlin MAX and the Merlin Trusted Core
At Horizon US 2026, Zycus Founder and CEO Aatish Dedhia launched Merlin MAX (Merlin Autonomous AI eXecution) with a clear point of view: procurement teams do not need more AI agents. They need procurement outcomes, delivered autonomously and safely.
Aatish Dedhia, Founder & CEO, Zycus, at Horizon US 2026. Watch all Horizon US 2026 sessions
Merlin MAX organises autonomous procurement into three towers:
1. Intake (“from requests to action”),
2. Strategy (“from insights to initiatives”) and
3. Outcomes & Autonomy (“from objectives to outcomes”).
Underneath all of them sits the Merlin Trusted Core: the shared context layer every flow draws from. Every agent works from the same definitions, follows the same rules and leaves the same audit trail. It is the architectural answer to the Shadow Agent Estate.

|
What AI needs
|
Gartner’s term
|
Merlin Trusted Core
|
|
Knowledge (what things mean)
|
Semantics
|
Canonical Data, Data
|
|
Expertise (how work flows)
|
Operational state
|
Orchestration, Integrations
|
|
Norms (what is allowed)
|
Provenance and controls
|
Guardrails, Financial Controls, Identity & Access, Auditability
|
When context is bolted on, every new agent needs its own setup, rules and clean-up, and agent debt grows with each one. When context is built into the core, every new flow inherits it from day one. That is how you scale from one autonomous use case to twenty without multiplying risk.
Read A CaseStudy – How a Light-Manufacturing Enterprise Delivered $4.3M in Annual Benefit with Autonomous Negotiation
Seqwater, Queensland’s bulk water supply authority, went live on Zycus in March 2026 with sourcing, suppliers and purchasing on one platform.
“Moving to a single, integrated platform provides the visibility, governance, and structure required to manage procurement at scale and support the organization’s long-term objectives.”
André Harvey, Chief Procurement Officer, Seqwater
Read how Seqwater went live on Zycus
From copilots to autonomy: the feedback loop that compounds ROI
The organisations seeing real returns treat context as something that improves over time, not something they configure once.
Every override is a signal. An invoice exception a human corrects, a sourcing award a category manager changes, an approval a policy did not anticipate. Fed back into the context layer, each one makes the next decision better. Gartner recommends that organizations should establish formal processes to review AI outcomes regularly and use those insights to refine the rules, knowledge, and data supplied to AI agents over time.
That is how autonomy grows: from agents that handle half the work while humans override the rest, to 70%, then 90%.
The appetite is there. The readiness is not. In Forrester’s Don’t Delegate AI, 69% of procurement leaders are confident in their AI vision, but only 31% are confident in their people and skills to scale it. Tail spend shows the gap clearly: 52% are comfortable granting AI high autonomy, but only 18% have deployed it. Closing that gap takes better context and the governance that makes autonomy safe, not a better model.
Context is the one AI advantage you cannot buy
Your competitors can buy the same models, license the same suites and read the same analyst reports. What they cannot buy is your context: your category strategies, negotiation history, supplier relationships, policies and the judgement of your best people.
Encoded properly, that is what makes your AI behave like your organisation and not like everyone else’s. In the AI era, context is intellectual property.
A 90-day playbook for CPOs
- Pick two or three high-value use cases. Invoice exceptions, tail-spend sourcing or contract compliance. Avoid trying to boil the ocean.
- Inventory your agents before adding more. Map every agent running across procurement, its owner and its data. Start with the agent debt self-assessment.
- Certify your core definitions. One agreed definition each for savings, spend, supplier, maverick spend and on-contract, signed off by finance, procurement and the business.
- Make your policies machine-readable. Approval thresholds, preferred supplier rules and match tolerances an agent follows by design, not interpretation.
- Let AI draft the context, and let humans certify it. Extract rules from the contracts, policies and playbooks you already have. Experts review rather than write from scratch.
- Test with and without context, then centralise the platform and federate the content. Measure the accuracy difference before you scale. A central team owns the platform. Category leads, AP and compliance own their domain context.
The question to ask before your next AI investment
Before you approve the next pilot, agent or platform, ask one question: Does it know what we mean?
If the answer is no, no amount of model intelligence will close the gap. If the answer is yes, and the context is built in rather than bolted on, AI stops being an experiment and starts delivering outcomes.
The choice in front of every CPO is simple. More AI agent sprawl, or better outcomes delivered autonomously.
See Merlin Agentic Flows in action
Go deeper: Zycus research with Hackett, Forrester and Ardent
- The Shadow Agent Estate, a Zycus Point of View with expert advice from The Hackett Group. What agent debt is, its four ledgers, and a 10-question audit.
- Don’t Delegate AI, a Forrester Consulting Opportunity Snapshot commissioned by Zycus. Why CPOs must own AI strategy, from a survey of 261 procurement leaders.
- Agentic AI in Procurement Adoption Index 2026, The Hackett Group with Zycus. Where procurement stands, from pilots to scale.
- The Rise of Agentic in Procurement, Ardent Partners, co-sponsored by Zycus. Embedded vs standalone agents.
Frequently asked questions
Why do most AI projects in procurement fail to deliver ROI?
The most common reason is missing business context, not the AI model. Agents that do not understand your definitions, workflows and policies produce confident but inaccurate outputs, which erodes trust and stalls adoption.
What is a Shadow Agent Estate?
It is the population of AI agents running across an enterprise that no one has inventoried, owned or governed. Its liability side is agent debt: the compounding cost of agents that cannot share context, coordinate or be audited together.
How long does it take to build context for procurement AI?
Starting with two or three high-value use cases and certifying core definitions, procurement teams can typically show measurable value within about 90 days.

