Since it emerged at the heart of the industrial revolution of the nineteenth century, the procurement function has never stopped changing in nature. Long confined to an administrative role (placing orders, paying suppliers), it has gradually established itself as a strategic lever: defending competitive advantage, managing supplier risk, watching the market, contributing to innovation. This rise has a flip side. More data to process, more criteria to analyse, more regulatory requirements to absorb. And the tools available today to procurement, finance and accounting departments have not kept pace.

This evolution is far from over. Manutan, in its article “How has the procurement function evolved since the 1850s?”, points out that the function still needs to become more responsive on two fronts: performance, and the automation of operational tasks. That is exactly where the tools in place reach their limits.

Today’s tools and their limits

The procurement function relies on a base of tools that have become indispensable: the ERP, e-procurement platforms, analytics tools. They have digitised processes, improved the traceability of flows, given a better view of spend and simplified part of the administrative work. Nobody is proposing to do without them.

But they have three limits that are becoming hard to ignore.

First: they work on a descriptive logic. They record what was entered and return it as reports. Decisions (assessing a supplier, preparing a negotiation, identifying a risk) remain entirely on the teams, who have to go and find the information elsewhere.

Second: they rely on predefined rules. An approval workflow, a threshold, a tolerance. Faced with a situation outside the frame (a supply disruption, an unexpected price movement, a geopolitical crisis), the rule does not know what to do, and the case falls back on a person.

Third: the quality of the output depends entirely on the quality of the data entered. A miscoded reference, a rounded rebate, a price that was never updated, and the analysis produced is wrong without anything flagging it. The tool does not check what it is given; it processes it.

These tools carried a major transformation. They are not obsolete, and nobody should plan to replace them. But on their own they cannot make procurement more predictive, more agile or more strategic, because they were never designed to read a supplier document, question it and compare it with the agreement that governs it. That is the context in which artificial intelligence stops being optional and becomes the layer that was missing between the documents and the system of record.

The explosion of supplier data

While the scope of procurement was widening, the volume and complexity of supplier data grew faster still. Larger supplier panels, more contractual references (framework contracts, unit price schedules, priced bills of quantities, technical specifications), pressure on sourcing costs, ESG requirements, regulatory obligations such as the CSRD or the CS3D: buyers now work in an unprecedented density of information.

The ERP, e-procurement and analytics tools centralise part of this information. But exploiting it remains hard, and the problem is not just one of quantity. It is qualitative and structural.

  • Data is scattered. ERP, procurement tool, Excel files, mailboxes: consolidating it takes time, making it consistent takes more.
  • Its reliability varies. Entry errors, outdated information, duplicates. An imperfect base produces imperfect analyses, and nobody knows to what extent.
  • Evaluation criteria have multiplied. Assessing a supplier is no longer a matter of comparing a price and a lead time. It now includes regulatory compliance, environmental performance, supply chain risk, the partner’s financial strength.

This complexity exceeds what traditional tools and human analysis can absorb. Buyers face a mass of information they cannot use, spend their time on processing tasks, and decisions slow down. Time spent reading, re-keying and checking is time taken from the missions that justify their role.

Managing supplier data therefore becomes a subject in its own right, for finance and accounting as much as for procurement. It requires tools capable not only of centralising information, but of structuring it, making it reliable and confronting it with what was negotiated: contract by contract, line by line, before the payment goes out rather than months later at audit.

What agentic AI brings

Faced with these limits, agentic artificial intelligence is not one more tool to add to the stack. It takes on what descriptive tools do not do: read the documents, extract the data, compare it and flag what does not add up.

Concretely, an AI can analyse large volumes of data continuously to identify trends, detect anomalies and spot discrepancies between documents that no team could reread in full. Its contributions to procurement, finance and accounting fall into three families.

Taking over low-value tasks. Contract analysis, offer comparison, invoice checking, verification of compliance with negotiated prices: these repetitive, time-consuming tasks can be entrusted to specialised AI agents, each on one type of document and one type of check. The human approves; the agent prepares the case.

Augmented decision support. By crossing internal data (purchase history, supplier performance, contractual terms) with external data (market movements, ESG indicators), AI produces more objective, faster recommendations. It makes it possible to simulate the impact of a renegotiation scenario, or to identify alternative suppliers from the actual history.

Anticipated risk management. Rather than suffering the unexpected, an equipped procurement department can see it coming: a price drifting gradually, a supplier failure taking shape, an alert as soon as the case appears. The function moves from a reactive stance to a proactive one.

Precision and traceability: the conditions of performance

Integrating these technologies sets its own demands. Two prerequisites come up in every project: data quality and traceability of actions. And companies must support their teams through the change, otherwise the tool goes unused.

An AI fed with incomplete or unreliable data produces insufficient results. That is why data governance must come before any automation project: knowing where the documents are, in what format, who is responsible for them and which ones are authoritative when two of them disagree.

Traceability, for its part, guarantees that every AI-assisted decision is documented, verifiable and compliant with regulatory requirements such as the GDPR and the CSRD. In a context where reporting obligations are tightening, an alert without proof is useless: a controller cannot dispute an invoice on the strength of a score, and an auditor will not accept one. Every discrepancy flagged must be traceable to the document, page and line it comes from, and the clause it was compared with.

Towards an augmented procurement function

One point is worth repeating: AI does not replace the buyer, the controller or the accountant. It augments them with more powerful decision-support tools. By taking on transactional and repetitive tasks, it gives teams back the time for what constitutes their real value: strategy, negotiation, the supplier relationship.

That is how procurement gradually becomes more proactive, more agile and more strategic, and how finance and accounting stop discovering at audit what could have been seen before payment. An augmented function, not a replaced one.

At Zylio, we chose a precise scope: verifying purchasing documents. Our agents read quotes, contracts, orders and invoices, and confront every line with what was negotiated. They detect invoicing errors, check the compliance of invoiced prices with negotiated prices and put quotes on a comparable basis. Zylio does not touch the ERP, contacts no supplier and commits nothing without approval: the agent prepares the case, the team decides.

On the scopes we analyse, this represents 15 to 45% of time given back to teams, per FTE, depending on the scope and the maturity of the data: time that returns to high-value work. To find out what it would give on your own documents, a call with an expert is enough to frame a first scope.