How AI orchestration resolves Source-to-Pay exceptions at industrial scale

The trap of simple automation: RPA versus agentic orchestration

Robotic process automation follows a script: open the invoice, read field twelve, compare it with the order, post or reject. It works until the supplier changes its invoice layout, splits a delivery, adds a surcharge line or references two orders on one document. Then the robot stops, and the exception lands on a buyer’s desk with no context. Most Procure-to-Pay automation programmes plateau here: the easy eighty per cent is touchless, the remaining twenty per cent costs more than before.

Unlike RPA, AI orchestration does not only automate linear tasks; it handles contextual drift and takes decisions under conditions set by the organisation. An agent reads the document as a whole, understands that the surcharge is a freight line covered by the contract, that the two references belong to one framework order, that the quantity deviation is within the tolerance Finance agreed, and resolves the case. What it cannot resolve within its rules, it routes to a person with the evidence already assembled.

This is an operating-model change inside procurement and supply chain, not a technology practice beside them. The buyer keeps the judgement; the agent removes the typing, the searching and the waiting.

The architecture of a Procure-to-Pay process run by AI agents

The orchestration layer sits between the documents that arrive, the Source-to-Pay platform or ERP already in place, and the people who own the decisions. Six steps, in order.

• Document reading: contextual extraction of the data on invoices, purchase orders, delivery notes and contracts, checked against item and supplier master data rather than copied from fixed positions.
• Matching and exception detection: three-way match between order, receipt and invoice; price, quantity, delivery-point and contract-term deviations detected and classified by type.
• Exception triage: minor anomalies resolved automatically within thresholds, for example a price variance below a defined tolerance or a quantity within the contracted range, with the rule applied and the evidence recorded.
• Human arbitration: a buyer, a planner or the accounts-payable owner notified only above the tolerance threshold or on a pattern the rules have not seen, with the context prepared so that the decision takes minutes.
• Audit record: every automated action logged with its inputs, the rule applied, the output and any human override, readable by internal audit and by the quality system where one exists.
• Learning loop: exception patterns reviewed weekly with Finance and procurement, rules and thresholds adjusted, the share of touchless cases tracked as a KPI.

The same architecture upstream and across the supply chain

The flow does not start at the invoice. Where requisitions are raised by hand from maintenance plans, production schedules and bills of materials, the same layer reads operational demand, checks contract coverage, stock and delivery point, creates the order and routes it by value and risk. Exceptions, and only exceptions, go to a person.

In the supply chain the pattern is identical: replenishment exceptions, ageing stock actions and forecast overrides proposed by agents inside the policies set in the operating model, with a planner deciding above the threshold. Procurement and supply chain are two linked disciplines, and AI orchestration is applied to both with the same discipline of owner, threshold and audit trail.

Governance, audit trails and ISO/IEC 42001

An agent that resolves exceptions touches supplier data, contracts and payments. It is therefore governed like any other control in Source-to-Pay. Decision rights are written down: what the agent may resolve alone, what a person must approve, how escalations work. Thresholds are agreed with Finance, and with Quality in regulated industries, before the first automated action. Each use case is entered in a register with its purpose, owner, inputs, outputs, risk classification and the controls that make it acceptable.

Monitoring follows: logging of inputs, outputs and overrides, drift indicators on extraction quality and exception rates, incident handling. These are the elements of an AI management system as ISO/IEC 42001 describes it, and they are built so that the obligations of the EU AI Act for the organisation’s risk classes are covered by the same registers and controls. The practical test: if an auditor asked which AI touched a payment, who approved the rule and how a deviation was caught, the answer is in the log.

Practical case and measured impact

A documented OREDJA engagement, anonymised. Outcomes for this programme are stated qualitatively; no quantitative result is published.

• Sector: pharmaceutical group with several regulated production sites, Europe.
• Scope: requisition-to-order for maintenance, production and laboratory materials across all sites, on the Source-to-Pay platform already in place.
• Approach: process-mining diagnostic of requisition and order flows; an orchestration layer between maintenance, production planning and the platform; a four-week pilot on one site and one category with every automated order validated by a person, then sampling; scale site by site over two quarters.
• Key result: requisitions created, checked and routed without manual entry across the group’s production sites; orders leave the platform within minutes of a work order or production run being released; people handle exceptions, and only exceptions.
• Impact: part-number, quantity and delivery-point errors caught as exceptions before ordering; buyer time redeployed from transaction processing to suppliers, contracts and critical materials; every automated action traceable in the quality system.

A second data point: contract deviations

At a European water utility, the orchestration layer was applied to the contract base rather than to invoices: contracts read and structured by an AI layer with buyer validation, price revisions recomputed before acceptance, deviations from agreed terms flagged to the category manager. Measured on the utility’s own baseline and validated by Finance: 8% savings on the categories in scope in the first sourcing wave, 100% of the contract base structured, sourcing preparation time halved. The reading layer stayed in place as a standing control.