AI in Procurement: Detect Off-Contract Spend with Your ERP

AI in procurement can help detect orders that do not comply with agreed-upon prices, suppliers, or terms without reviewing every document manually. The key is to connect ERP data with active contracts, compare equivalent terms, and deliver evidence-backed alerts so procurement and finance can decide what to do.

On October 26, 2026, Oracle announced new applications with AI agents, including Procurement Spend, designed to analyze purchases, spending, and supplier agreements to detect deviations and recommend actions subject to review and approval, according to its official announcement. These are vendor-announced capabilities, not proven savings. For a company in Mexico, the useful question is how to apply this approach to its operations without assuming it needs to switch ERPs.

1. What AI in procurement can detect

An off-contract purchase does not always indicate misconduct. It may result from a pending update, an authorized exception, or a data entry error. That is why the system should identify potential discrepancies, not draw automatic conclusions.

The first step is to define what constitutes noncompliance with a term at your company:

Potential deviation	Required comparison	What the team should validate
Price above the agreed rate	Order unit price against the active contract	Discounts, volume, taxes, and freight
Unauthorized supplier	Order supplier against the approved supplier list	Exception or pending supplier approval
Different payment terms	Order terms against the signed agreement	Applicable contract amendment
Purchase after expiration	Transaction date against the contract term	Documented renewal or extension
Missing discount	Cumulative volume against the negotiated discount tiers	Period and products covered

For example, a construction company could detect an equipment rental rate above the agreed amount. Before confirming the discrepancy, it would need to check whether the rate includes transportation, different equipment capacity, or a service extension. Comparing amounts without that context would produce unhelpful alerts.

2. What data you need to cross-reference with your ERP

The pilot requires a limited but complete sample. Extracting every purchase is of little use if the corresponding contracts are missing.

Gather these sources:

• Purchase orders and their line items: supplier, product or service, quantity, unit, currency, price, date, cost center, and terms.
• Contracts and attachments: prices, effective dates, discount tiers, exclusions, renewals, and authorized amendments.
• Supplier master data: internal identifier, RFC tax ID where applicable, status, and approved categories.
• Approvals and exceptions: who authorized different terms, for which transaction, and for what period.

If the scope includes verifying what was received or invoiced, also include receiving records and invoices. These are not essential for comparing an order against a contract, but they are needed to review differences between what was agreed, purchased, and billed.

Data preparation must resolve equivalencies: the same supplier may appear under different names, and a material may be purchased by the piece or by the box. Without standardizing identities, units, and currencies, the system will confuse formatting differences with noncompliance.

When evidence is missing, the correct result is “insufficient information,” not “compliant purchase.”

3. How AI integration with an ERP works

It is best to separate AI tasks from deterministic checks. The model helps interpret documents; rules and calculations verify explicit terms.

An initial workflow can work as follows:

1. Access the ERP in read-only mode. Use APIs, connectors, or controlled exports, depending on the capabilities available.
2. Extract contract terms. Identify prices, dates, deadlines, and exceptions while retaining references to the source document and section.
3. Link records. Match each line item with the supplier, product, and contract version applicable on the purchase date.
4. Compare terms. Run calculations and business rules on normalized data.
5. Prepare an explainable alert. Show the difference, supporting evidence, and suggested action for review.

For example, AI can locate a discount clause written in natural language. The application should then calculate whether the required volume was reached, rather than leaving the arithmetic to a generative response.

A custom AI integration lets you evaluate this workflow using existing systems. Feasibility will depend on data access, permissions, and document quality—not just the model selected.

4. How to plan a focused pilot

Start with a category that has relatively clear terms, recurring purchases, and a designated owner available to validate results. Avoid rolling it out across all business units at once.

Define the scope and baseline

Select suppliers, a historical period, and discrepancy types. Document how those transactions are reviewed today and how long they take to resolve, including the search for supporting evidence.

Next, prepare a team-reviewed sample that includes compliant purchases, known discrepancies, and authorized exceptions. This baseline will help verify whether the system distinguishes relevant cases.

Run without changing operations

Process historical purchases first. Then test new transactions, but without blocking orders, modifying suppliers, or changing payment terms.

Classify each alert as confirmed, dismissed, or awaiting information. Also record the reason: wrong contract, incorrect unit, valid exception, or an actual difference.

Before expanding the scope, agree on acceptance criteria: alert quality, review time, document coverage, and the absence of unauthorized changes. AI process automation should build on those results, not just a working demo.

5. What a person should validate

A useful alert should support a decision without requiring someone to reconstruct the entire case file. Include:

• Affected order and line item.
• Agreed-upon terms and recorded terms.
• Calculated difference and method used.
• Contract, version, and excerpt supporting the comparison.
• Missing data, assigned owner, and review status.

Procurement can validate prices, suppliers, and commercial exceptions. Finance can review the financial impact and how payment should be handled. Operations can clarify quantities, deliveries, or scope changes. Ambiguous clauses may require legal review.

During the pilot, do not automatically delegate payment holds, supplier claims, or contract amendments to the system. Its recommendation must remain separate from authorization to act.

You also need role-based permissions, access logs, and traceable decisions. Contracts contain sensitive information: each user and service should access only what is necessary for its role.

6. How to measure spend control with AI

Counting alerts does not demonstrate value. A system that flags everything can increase the workload. Track metrics that connect detection, review, and resolution:

Metric	How to measure it	Purpose
Review time	Median minutes per case, before and during the pilot	Determine whether manual effort decreases
Confirmed alerts	Confirmed discrepancies as a share of resolved alerts	Evaluate the usefulness of detections
Missed discrepancies	Cases found through human review that the system did not flag	Identify blind spots
Verifiable coverage	Line items with sufficient evidence as a share of line items analyzed	Distinguish missing data from compliance
Resolution time	Time from alert creation to closure	Identify operational bottlenecks

To find missed discrepancies, also review a sample of purchases with no alerts. Compare equivalent transactions and break down results by category or deviation type.

Also distinguish between flagged amounts, confirmed discrepancies, and amounts actually corrected. A flagged difference does not equal savings: it may be valid, duplicated, or unrecoverable.

7. Conclusion: start with a verifiable decision

AI in procurement adds value when it connects a transaction to the terms it should have met and supports an evidence-based decision. The starting point is a pilot with a limited scope, read-only ERP access, identifiable contracts, and designated owners responsible for validating alerts.

If you want to evaluate this approach, gather a sample of orders, contracts, and authorized exceptions. Contact Slashweb to discuss a custom integration and define which discrepancies to detect first, which connections your ERP needs, and how to measure results before scaling.