Compliance is not an enforcement problem. It is a design problem. When the easy path and the compliant path are the same path, maverick behavior disappears, not because anyone is watching, but because there is nothing to deviate to. That principle does not stop at tail spend. It applies equally to supplier onboarding, contract management, and internal approvals. Every time an organization asks why people are not following the process, the honest answer is usually that the process was never the path of least resistance.
AI anxiety in procurement is rational, but most CPOs are managing the wrong risk. The bottleneck is not the technology. It is process clarity. Agentic AI can automate full procurement cycles, from supplier qualification to invoice matching, but only where decision logic is explicit and machine-readable.
Organizations that deploy agents on top of undocumented processes do not gain efficiency. They expose their gaps. The CPOs getting this right run a contained pilot and a process diagnostic in parallel, not in sequence. The result is a procurement function that is leaner, more strategic, and built to scale.
This article discusses:
- What agentic AI is, how it differs from tools that only advise, and why explicit decision logic is the one condition it cannot work without
- Which procurement decisions can credibly be automated and which cannot, tested against reversibility, stakeholder dependency, and regulatory explainability
- Why most implementations disappoint for organizational rather than technical reasons, and what that means for how a programme is framed internally
- How the procurement operating model changes as decision logic becomes explicit, and what CPOs should do in the first 90 days
The Conversations Have Changed
A year ago, the opening question was usually some version of “show me an AI use case.” It is no longer that. The question now tends to be one of three: Am I behind? What am I supposed to be doing first? Or, less often but more honestly, how much of my function survives this?
The AI anxiety in procurement is rational. It is also, in our experience, pointing in the wrong direction.
The risk most CPOs are actively managing – picking the wrong vendor, running a failed pilot, getting flamed at a town hall when the headcount math comes out – is not the risk that will matter most in 2027. The risk that will matter is harder to avoid: discovering, too late, that the function never knew its own processes well enough to put an agent on top of them in the first place.
What Agentic AI Is, And What Most People Get Wrong
An agentic AI is software that acts, not just advises. More precisely: it is goal-directed rather than task-prompted, takes tool-mediated actions across multiple systems, and operates inside a defined policy envelope without per-step human approval. A chatbot with a few plugins is not an agent. An agent is what happens when the model is allowed to act.¹ That single change is what is structurally new, and it is also what makes the regulatory floor matter², because the cost curve on the underlying models has dropped roughly 80% in eighteen months³, and what previously sat in pilot decks is now production-grade.
Think of a routine purchase order approval: a chatbot in Outlook can draft the email asking a category manager to sign off, but it still waits for someone to approve. An agent skips that step. It checks the requisition against policy in SAP Ariba, issues the purchase order itself, and only escalates to a human if a threshold is breached.
Applied to procurement more broadly, this means an agent can read an incoming requirement, qualify a supplier, run a sourcing event, evaluate offers, issue a purchase order, and reconcile the invoice – end to end, with no human in the loop, provided the rules are written down. The “provided” is the part that matters. The single condition for any of this to work is that the underlying decision logic has to be explicit and machine-readable. Not “we usually go with X.” Not “Maria handles those.” Actual rules, with thresholds, escalation triggers, and compliance boundaries.
¹Gartner, How Intelligent Agents in AI Can Work Alone, 2025.
²EU AI Act (Reg. 2024/1689); CSRD (Dir. 2022/2464); German LkSG (2023). European Commission / EUR-Lex / BMAS.
³Stanford HAI, AI Index Report 2025, Ch. 1; Epoch AI inference pricing data.
| Dimension | Traditional AI | Agentic AI |
|---|---|---|
| Output | Recommendation | Decision + executed action |
| Human role | Decision-maker | Decision supervisor |
| Logic | Task-based | Goal-directed |
| Failure mode | Ignored by users | Acts on bad policy |
Three things are widely misunderstood about this, and each one changes a CPO’s investment decision:
The technology question is largely solved: open weights (AI that can be deployed without vendor lock-in), orchestration frameworks (software that coordinates what the agent does and when), ERP and source-to-pay integration are all good enough. The bottleneck is not AI quality. It is process clarity. That means explicit thresholds, not judgment calls; escalation rules written down, not held in senior buyers’ heads; policies that exist as code rather than convention. Wherever decision logic is explicit, agentic AI compounds value. Wherever it is not, the technology does not save you, it exposes you.
A pilot shows what is missing: the first six months of any serious deployment surface every implicit rule, every undocumented escalation, every silent contradiction in the policy manual. Companies that go in expecting a technology test and get a process audit usually burn credibility with internal stakeholders – business units promised efficiency, leadership expecting a deployment, before the system, ever stabilizes.
They are where the headlines are. The cash is in the operational use cases nobody writes a keynote about: supplier onboarding, tail-spend execution, continuous risk monitoring.
//AI alone is not enough – 70% of a transformation’s success depends on people, the organization, and processes
Share of respondents’ key challenges1 (%)2
80%
Insufficient model accuracy and reliability
78%
Lack of access to high-quality data
65%
Inability to explain hallucinations in AI
76%
Integrating AI with systems, tools, and APIs
67%
AI-driven security risks
63%
Compliance and implementing responsible AI
61%
High AI scaling costs that are difficult to control
92%
No expertise to manage unstructured data
88%
Silos limiting cross-functional collaboration on AI
86%
People adoption of change and using AI daily
78%
Difficulty aligning AI initiatives with firm strategy
73%
Social resistance and labor concerns limiting scale
71%
Shortage of AI talent
71%
Missing clear AI metrics and ROI measurement
59%
Lack of leadership support or commitment for AI
Algorithms
Data science capabilities to develop and implement AI and GenAI algorithms
10%
Technology
Scalable and modernized stack that supports business needs
20%
People, organization,
and processes
Effective processes supported by talent and change-management practices
70%
Share of respondents’ key challenges1 (%)2
80%
Insufficient model accuracy and reliability
78%
Lack of access to high-quality data
65%
Inability to explain hallucinations in AI
76%
Integrating AI with systems, tools, and APIs
67%
AI-driven security risks
63%
Compliance and implementing responsible AI
61%
High AI scaling costs that are difficult to control
92%
No expertise to manage unstructured data
88%
Silos limiting cross-functional collaboration on AI
86%
People adoption of change and using AI daily
78%
Difficulty aligning AI initiatives with firm strategy
73%
Social resistance and labor concerns limiting scale
71%
Shortage of AI talent
71%
Missing clear AI metrics and ROI measurement
59%
Lack of leadership support or commitment for AI
¹ Which of these challenges hinder adoption and scale of AI in your company?
² Percentage of respondents who agree and strongly agree.
Note: API = application programming interface; ROI = return on investment.
Source: BCG Build for the Future 2025 Global Study (n = 1.250).
Where Agentic AI Actually Pays Off
The wrong question is “where can we apply agentic AI?” Asked that way, every workshop ends with the same fifteen use cases on a whiteboard. The right question is: which decisions can we credibly take out of human hands, and which ones can we not? Three filters get you most of the way there: reversibility (can the decision be undone cheaply if wrong?), stakeholder asymmetry (does the outcome depend on a relationship the agent cannot hold?), and regulatory explainability (does the law require a human in the loop?). Applied honestly, the picture clarifies quickly.
| Will be automated | Will remain human |
|---|---|
| RFQ creation and supplier shortlisting | Strategic supplier negotiations |
| Catalog-based ordering | Make-or-buy decisions |
| Document validation and compliance checks | Critical supplier development |
| Invoice matching and payment release | High-stakes contract redesign |
| Tail-spend sourcing | Crisis management and escalation |
Three use cases stand out by this test:
An agent contacts the supplier, requests documents through a structured channel, validates against defined rules, files them in the system of record, and re-screens on a defined cadence. Typical impact: onboarding from weeks to roughly two days; audit exposure measurably reduced. The hard part is not the agent; it is defining what “qualified” actually means in machine-readable terms.⁴
Autonomous Purchase Request Screening – From Policy to Action
An agent screens every purchase request above a defined threshold before commitment. Specialist checks run automatically: compliance, price reasonability, contract coverage, savings levers, and return each request as either validated or challenged, with a quantified alternative attached. Challenged requests route to a human with the full reasoning already prepared.
A recent client ran more than 700 purchase requests through this process in a single year. Roughly half were challenged. The result: €1.2M unlocked and around 6% on addressed spend. The engine itself, built on our Spend Guard tool, was not the hard part. Encoding the client’s own thresholds, policies, and escalation logic into it was.
That distinction matters. The technology is a vessel. What you pour into it, your rules, your risk tolerances, your definition of a good deal, determines whether it creates value or just creates activity; client’s own thresholds and policies into it was the deliverable and the hard part.
⁴ Gartner Newsroom, 2025. Impact figures based on Inverto engagement experience.
An agent runs the full cycle: classification, supplier selection from approved catalogues, RFQ where needed, awarding by rules, PO issuance, three-way matching, payment release within tolerance. Typical impact: process cost down 50–70%, cycle time from days to hours, maverick spend down by roughly 40%.
Autonomous Contract Review – From Unstructured Documents to Auditable Intelligence
An agent reads unstructured supplier contracts and produces a structured, auditable output: every value grounded in a verbatim quote with page reference. Compliance checks, liability caps, missing termination notices, indemnity gaps: flagged automatically, at scale.
A recent client faced 480 supplier contracts across four jurisdictions to review before signing a healthcare acquisition.
The process completed in three working days against a six-week manual baseline. 78% of risk clauses were surfaced automatically, built on our Document Intelligence tool, which ships with twelve extraction profiles out of the box. Getting the client’s own extraction schema right was the deliverable.
That pattern holds across every contract-intensive moment: acquisitions, supplier consolidations, regulatory audits. The agent does not replace legal judgment. It ensures legal judgment is applied where it actually matters.
An agent ingests signals across financial, geopolitical, cyber, and ESG dimensions, monitors the supplier base against an integrated risk taxonomy, and when a threshold triggers, stages a response: alternative suppliers pre-qualified, mitigation playbook pulled up, draft communications prepared. Typical impact: response time to disruption from days to hours. The lever is preparation, not detection.
AI-Assisted Competitive Sourcing: From Scattered Bids to an Optimized Award (RFX Suite)
An agent structures the bid sheets, ingests and consolidates supplier responses across rounds, scores them against a weighted model, and computes multiple awarding scenarios under real constraints, volume splits, capacity limits, single-source rules – surfacing the optimum for a human to sign off rather than assemble by hand.
A recent Inverto engagement ran a pharmaceutical tender covering roughly 300 SKUs through this process, consolidating what would otherwise have been more than 500 Excel tabs per round into a single, refreshable model.
Built on our RFX Suite, the analysis was ready before the responses even landed, and the tender cycle compressed by up to 50% while holding 100% data consistency across every bidding round. The engine handled the mathematics; the deliverable was encoding the client’s own awarding logic, constraints, and scoring weights into it, the part that determines whether an “optimal” award is actually the one the business wants.
The pattern holds across any high-volume, rules-heavy tender: MRO, indirect categories, clinical-trial sourcing. The agent does not replace the awarding decision. It ensures the decision is made against every scenario, not just the two or three a human had time to build.
Why Do Most Implementations Disappoint?
Because most failures are organizational, not technical. The technology is not the problem; the implementation context is. Three predictable failure modes account for the majority of disappointing outcomes, and all three are avoidable.⁵
- Expecting autonomy on day one: Heavy human oversight is the norm well into production, not a temporary phase. In the first large-scale study of production AI agents, 74% of deployed systems still depended primarily on human evaluation and 68% were constrained to at most ten steps before requiring human intervention.⁶ Programs pitched internally as deployments rather than supervised learning periods burn credibility before they earn it.
- Treating agentic AI as a substitute for process work: It is the opposite. Agentic AI requires explicit decision criteria, escalation rules, and compliance boundaries to function. Programs built on weak process foundations do not just underperform; they surface, often publicly, that the documented process and the real process were never the same thing.
- Confusing visibility with impact: AI-supported negotiation and AI-driven spend analytics are the most visible investments. Neither is where the cash is. Negotiation depends on power dynamics and accountability that cannot be delegated to software. Analytics improves transparency, but transparency is no longer most procurement functions’ bottleneck.
⁵ Gartner Newsroom, 2025. ⁶ Pan et al., Measuring Agents in Production (MAP), UC Berkeley, arXiv:2512.04123, 2025.
Where To Put The Money
The axis that matters is no longer visibility but readiness: how far a vendor has already productized a use case versus how much a CPO has to build internally.
▲
High
Low
EBITDA impact
High impact · build deliberately
Supplier onboarding
High impact · low process maturityAI-led negotiation
High impact · needs operating modelHigh impact · start now
Start here ↑
Tail-spend automation
High impact · largely productizedSupplier risk monitoring
High impact · largely productizedNot described
Low impact · easy to buy
Spend analytics
Easy to greenlight · modest ROI← the seduction
▶
Must build internally
Productized · buy now
Vendor readiness
The cases that move EBITDA split in a telling way. Tail-spend automation and continuous supplier risk monitoring sit top-right, high impact and largely productized, so a CPO can start almost immediately. Supplier onboarding and AI-led strategic negotiation sit top-left, the same order of impact, but low process maturity, so they have to be built deliberately into the operating model rather than bought.
The seduction is the bottom-right: cases like AI-driven spend analytics are easy to greenlight because a vendor is actively selling them with a polished demo and a reference list, yet their EBITDA impact is modest. The pattern is not accidental. What sits on the right is whatever a vendor has already packaged; the highest-leverage work on the left is precisely the part no vendor can hand over, because it depends on internal process clarity. That clarity, not the technology, is the bottleneck.
The Organizational Reality
Agentic AI is not an IT project, it is an operating model redesign. Three consequences tend to be underestimated:
Every threshold, every escalation rule, every implicit preference becomes explicit code. Some power structures inside procurement will not survive the daylight. Senior buyers whose value has rested on irreplicable judgment will find that “judgment nobody else can replicate” reads, in policy language, as “logic we don’t have.” This has to be navigated openly, not denied, and the conversation belongs to the CPO and CHRO together, not to IT. It needs to happen before deployment, not after the headcount impact is visible. The framing matters more than most programs acknowledge: “your judgment becomes policy” lands very differently than “your role is at risk.” One invites senior buyers into the process as authors. The other puts them on the defensive before the first agent goes live.
Operational buyers move toward agent supervision and exception handling. Strategic buyers and category managers move deeper into work that genuinely requires judgment. The middle compresses; in functions with mature implementation, mid-tier transactional headcount has reduced by 30–50% over 24–36 months.⁷
⁷ Inverto engagement experience. See also: OECD, Employment Outlook 2023; OECD, Using AI in the Workplace, 2024.
When an agent issues a non-compliant PO, the person responsible is whoever wrote the policy, because nobody pressed send. Under the EU AI Act,⁸ certain procurement use cases may qualify as high-risk, triggering documentation, oversight, and liability obligations that most CPOs have not yet modeled. A number will discover, too late, that they have quietly inherited risk that previously sat distributed across hundreds of individual decisions.
⁸ EU AI Act, Annex III; Articles 9-15.
Conclusion:
Whichever path is chosen, the question being asked in 2027 will not be whether procurement uses agentic AI. It will be why a human is still doing this. The anxiety in the conversation today is the asset; directed at process clarity rather than at technology, it is the most useful tool a CPO has.
Our Digital & AI Experts
Christoph Schenter
Principal Digital & AI
Johannes Hofberger
Principal Digital & AI
Moritz Hagemann
Project Manager
Simon Rühle
Project Manager
Further Digital & AI Insights
