The autonomy gap | Why procurement AI stalls without decision architecture

“Agentic Procurement”, “Autonomous Sourcing”, “No-touch Contracting”

Heard any of these terms recently? If you work in procurement, chances are high you have.

They refer to agentic AI systems designed to independently handle parts of the sourcing process, reducing the need for human involvement from the initial request through to evaluating supplier inputs. In theory, an agent identifies a sourcing need, runs an RFP, negotiates terms, and awards. In practice, we are not there … yet.

If our last article (read the piece here) focused on choosing vendors that survive the current AI market shakeout, this part is about ensuring your organisation survives its own ambition. Now the conversation moves from vendors to applications. What does AI in procurement actually mean today? And how should organisations approach it?

Right now, the race for AI feels like a sprint. Every week brings new model announcements, and claims of “100% automation”. However, AI Autonomy is not just another plug and play solution. It requires a new delegation model. Thus, the organisations that identify the right business cases, and roll out in controlled stages will make more progress than those chasing speed.

Where AI already delivers

Before diving into where full autonomy reaches its limits, it’s worth taking a look at what is already working well.

Take a €2m marketing services RFP: until recently, a large part of the effort sat in preparation and analysis, from pulling old templates to building slides for the negotiation round. If used well, AI tools can take the weight off the repetitive parts of the process:

1. The first draft of the RFP is structured and aligned with previous events
2. Supplier proposals are summarised into comparable formats instead of being read line by line
3. Contract deviations are flagged against the standard position
4. The negotiation brief is prepared with benchmarks and key gaps already highlighted

At the end, the decision still sits with the buyer and commercial judgement still requires experience, but the time spent on manual consolidation significantly drops. This example is a form of augmentation. From a maturity perspective, current AI implementation activity falls into three progressive layers: Automation (task optimisation), Augmentation (decision support), and Orchestration (agentic AI).

1. Automation: task optimisation through machine learning with AI
enhancement

This is the most mature layer, and many organisations have deployed some of the examples below. Here, a mix of traditional machine learning models and increasingly GenAI optimises high-volume, rules-based activities.

Examples in this area include:

• Invoice matching
• Spend classification
• Contract clause extraction and metadata tagging

These systems operate in contained subprocesses. The decision logic is predefined.

2. Augmentation: LLM enabled decision support

The second layer introduces large language models into professional workflows:

• Drafting support for sourcing events
• Supplier proposal summarisation
• Contract risk flagging
• Negotiation brief preparation

While these solutions speed up tasks and work as helpers, the scope here remains controlled. They do not generate savings on their own, but they improve execution efficiency.

3. Orchestration: building the foundations for agentic AI

The most advanced organisations are preparing the foundations for agentic
AI. This involves a number of steps, including:

A. Defining the economic focus (e.g. tail spend sourcing)

B. Structuring the core data (e.g. deduplicated master data, standardised categories)

C. Codifying decision architecture (e.g. defined approval thresholds & decision rights)

D. Standardising workflows (e.g. intake -> sourcing -> contracting)

E. Building the orchestration layer (e.g. system integrations, workflow routing logic, system handovers)

F. Introducing AI in controlled scope (e.g. if a request sits under tail spend copilots for
RFx drafting)

Among these, codifying decision architecture is the most critical and most overlooked step. It defines the rules of the process: what an agent can do, when it must escalate, and where it is constrained. This is also where the transition from augmentation to orchestration typically breaks down.

Procurement platforms such as Coupa, SAP Ariba, Jaggaer and Ivalua have introduced capabilities that span all three layers of AI maturity, typically layered onto existing systems rather than built as native functionality. At the automation level, they optimise high-volume tasks such as invoice processing and spend classification. At the augmentation level, they introduce copilots and predictive insights that support sourcing decisions and contract analysis.

At the orchestration level, some platforms including SAP Ariba and Ivalua increasingly build the backbone for agentic workflows through integrated data models and orchestration logic. However, the maturity of capabilities varies significantly.

Overall, the market is moving faster than organisational readiness and the constraint to successful implementation is the absence of defined decision rules, ownership, and process structure. Those that begin preparing for operational autonomy now will be in a position to scale operational autonomy and build a lasting advantage. To understand how this gap plays out in practice, let’s look at where full autonomy currently breaks down in procurement.

Where full autonomy currently breaks down in procurement

Imagine a €5m IT services sourcing event: An agent 1) runs the RFP, 2) scores the proposals and 3) selects supplier A based on a weighted criterion.

On paper, optimal. But let’s imagine further: 1) the supplier master data contains duplicate records 2) finance has a preferred supplier relationship strategy the model cannot see, and 3) the contract contains pricing indexation risk that was not captured in training data.

The example demonstrates how the environment the agent operates in is not structured for autonomous decision-making. Without clearly defined decision rules, escalation logic, and ownership, autonomy breaks down.

Autonomy requires decision architecture

Agentic AI primarily tests organisational limits, rather than technical ones. How can an organisation run a successful race when it has not defined the rules? After all, autonomous or semi-autonomous workflows require explicit decision architecture, which many procurement functions have never codified.

• Who owns which decision?
• At what financial threshold does escalation occur?
• What risk exposure is acceptable without human review?
• What negotiation boundaries are fixed, and which are discretionary?

For decades, much of this has lived in experience and informal judgement. Senior buyers “know” when to escalate. Category managers “sense” risk. Humans compensate for ambiguity, agents cannot. The moment tasks get delegated to agentic, probabilistic systems to make decisions and flag problems, all the grey zones become blockers.

So how does it work? An agent operates through an internal logic layer, which includes its instructions (system prompt), access to memory and pre-defined behavioural rules. Together these form the decision architecture, defining what the agent is allowed to do, and when escalation is required. Orchestration ensures that these rules are executed across systems, approvals are triggered correctly, and that checks happen in the right sequence. It is orchestration that enables the agent to interact with systems such as SAP Ariba or Ivalua to retrieve data and execute actions. These platforms provide the underlying infrastructure and data layer.

To make this more tangible, let’s return to our example of the IT sourcing event.

1. Input: a business user submits a sourcing request via Ariba. For example: “We need a new IT services vendor “, with a budget of €5M, scope and timeline defined.
2. Action & access to relevant tools: based on this request, the agent retrieves relevant data from connected systems, such as previous sourcing events, approved supplier lists, contract repositories and supplier risk profiles. Through its internal logic layer, it
   structures the request into a draft RFQ, suggests evaluation criteria (e.g. price, delivery capability, experience), and proposes an initial supplier longlist.
3. Output: the output is a structured RFQ, a defined evaluation approach and a proposed supplier list. The buyer reviews, refines, and takes the event live.

In reality, enabling a workflow like this depends on a highly fragmented technology landscape. ERP cores, sourcing suites, CLM systems and risk tools operate on different data structures and governance models. Procurement platforms such as SAP Ariba, and Coupa are beginning to bridge this fragmentation by introducing AI-driven capabilities and orchestration logic that support agentic workflows. This provides the technical foundation for automation across systems. However, these platforms should not be confused with agents that independently execute complete E2E workflows. To layer agents onto this
infrastructure, organisations must define decision architecture, an organisational transformation that will not happen overnight.

Much of it is not new. Organisations have long known the importance of clean data and clear accountability. What is changing are the costs of not fixing them. Building decision architecture and clean data foundations takes years. Organisations that don’t start now will not be able to catch up later. When the pressure to scale agentic AI comes, and it will come, they will realise the foundations aren’t there and others will be racing ahead.

The marathon ahead

In practice, the pace of procurement autonomy will depend less on model capability and more on how quickly organisations put the right structure in place.

In the near term, we are likely to see:

• Constrained agents operating within clearly defined financial and risk boundaries
• Embedded human oversight by design, not exception
• Transparent audit trails for every system-triggered action
• Gradual expansion of automation scope as trust increases

This is part of how the process needs to be structured, and it depends on having clear decision rules in place.

In the longer term, who knows, procurement itself may be restructured around intelligent workflows rather than functional silos. But that shift will only happen once decision-making can be reliably delegated.

Today, while the market is pushing autonomy, most organisations are still building
structure. Procurement AI largely operates in three layers: task automation, decision support and workflow orchestration. These improve efficiency and consistency. They do not yet transform unit economics. Hard savings still depend on human strategy and negotiation.

Co-pilots improve speed, but to achieve structural cost improvements we need to delegate decisions within defined boundaries. That requires changes to how procurement is set up. Organisations that treat AI as part of their operating model will move beyond pilots. That is where the real progress sits.