AI Invoice Approval: CFDIs, Purchase Orders, and Receipts

AI invoice approval is useful when a CFDI arrives in order but does not match what was purchased or received. For administration and finance teams, the challenge is deciding whether the discrepancy allows processing to continue, requires clarification, or should block the invoice file before payment.

On November 5, 2026, Sage presented its vision for AI Workspaces for Intacct: an AP Agent that compares invoices with purchase orders, identifies exceptions, and routes approvals. The Sage announcement published on FT outlines a limited rollout to selected customers during November, not general availability; it also does not demonstrate CFDI compatibility and notes variations by feature and market. The takeaway for a Mexican company is not to adopt that product, but to define the controls it needs before automating decisions.

1. What to compare across the CFDI, purchase order, and receiving record

Matching CFDIs to purchase orders should incorporate a third element: the goods receipt or service acceptance record. Each document answers a question:

• CFDI: What is the supplier billing for?
• Approved purchase order: What was agreed to be purchased, and under what terms?
• Receipt or acceptance: What was delivered or accepted as complete?

Compare line items, not just totals. Two documents can have the same total but different quantities, descriptions, or unit prices. You also need to account for partial receipts and quantities already associated with other invoices to avoid reusing a delivery as supporting evidence.

Before comparing, standardize units of measure, currency, discounts, and tax treatment. A box does not necessarily equal an individual item; a net amount is not directly comparable to a total that includes taxes, either.

Tax compliance review and commercial approval are separate controls. Passing the validations defined by accounting does not prove that the goods arrived or that the expense is authorized.

2. AI invoice approval: separate three responsibilities

An architecture that supports effective controls separates interpretation, rules, and authorization. Mixing them makes it difficult to explain why an invoice moved forward.

AI interprets and explains

It can suggest matches between different descriptions, summarize receiving notes, and draft an explanation: “The invoiced quantity exceeds the recorded receipt; a delivery is pending.” Each statement must reference specific documents or records.

When evidence is missing, AI must flag it. A likely match between item descriptions does not equal a confirmed receipt or authorize changes to the ERP.

Rules calculate and enforce controls

Calculations and blocks must run through verifiable logic: line-item discrepancies, tolerances, required documents, potential duplicates, and approval limits. AI can explain the result, but it cannot invent an exception or change a policy.

A person resolves issues that require judgment

Procurement confirms negotiated terms; the warehouse verifies deliveries; the requesting department accepts services; finance authorizes exceptions within its authority. Recording a receipt should not, by itself, give someone permission to approve any discrepancy.

An enterprise AI integration must preserve this separation. Approving an invoice must not automatically trigger payment: treasury retains its controls and authorizations.

3. Define tolerances, owners, and evidence

Avoid a universal tolerance based on the total. Define limits by discrepancy type, purchase category, and risk level. If you use both a percentage and a maximum amount, document whether both must be met; make clear how each is calculated.

The following matrix is a starting point, not a production-ready policy:

Exception	Control rule	Responsible party	Evidence needed to resolve
Price discrepancy	Compare the net price per line item against the current purchase order	Procurement; finance if the exception exceeds procurement's authority	Revised purchase order or documented authorization
Partial receipt	Hold full approval if the invoiced quantity exceeds the received quantity	Warehouse or requester	Additional receiving record or clarification from the supplier
Missing document	Prevent the file from moving forward when required supporting documentation is missing	Document owner	Document linked to the invoice file
Potential duplicate	Block repeated use of the same UUID and review other matches	Accounts payable	Relationship between documents and recorded resolution
Tax discrepancy	Route to accounting review without treating it as a commercial tolerance	Accounting	Validation and, when applicable, corrected documentation

Tolerances must have a version, an effective date, and an owner. Do not allow positive and negative discrepancies to offset each other in the total and hide significant line-item exceptions.

For each case, add a response deadline and an escalation path. If someone does not respond, the system must escalate rather than interpret silence as approval.

4. An illustrative workflow: detect, explain, and record

Suppose a purchase order is for 100 units, a receiving record shows 80, and a CFDI bills for 100. This is an operational example, not a client case study.

1. Detect. The system finds a 20-unit discrepancy between the invoiced and received quantities. It checks for other linked receiving records and whether they have already been used to support previous invoices.
2. Explain. AI presents the discrepancy and links to the purchase order, the CFDI line item, and the receiving record. It does not claim the supplier failed to deliver: a delivery may simply not have been recorded.
3. Assign. The warehouse verifies the delivery. If it received only 80 units, procurement follows up with the supplier; if it received 100, the warehouse completes the record with supporting evidence.
4. Resolve. An authorized person determines whether to wait, request a correction, or apply partial processing permitted by policy and the ERP.
5. Record. The invoice file retains the decision, its rationale, the approver, and the documents considered.

If a receiving record or purchase order changes after approval, the system must rerun the matching process and determine whether the review should be reopened. An approval based on earlier data cannot carry over to a modified file without review.

5. How to organize an ERP-connected pilot

Start with a purchase category that has well-documented purchase orders and receiving records. Include files with no discrepancies as well as known exceptions; testing only simple cases does not allow you to evaluate the controls.

The pilot needs an explicit scope:

• Data: supplier identifiers, UUIDs, line items, purchase orders, receiving records, statuses, and existing approvals.
• Permissions: read-only access initially; any subsequent write access must be limited to authorized operations and users.
• Statuses: pending matching, under review, blocked, approved, or rejected, with defined transitions.
• Traceability: references to evidence, rule version, generated explanation, and final decision.
• Contingency: when data is incomplete, failures occur, or concurrent changes arise, stop processing and route the file for review.

Start in observation mode: the system makes recommendations while the team maintains its usual process. Compare its results against files reviewed by people before enabling automatic approvals for routine cases that meet all the rules.

AI process automation must include a way to pause the workflow, revert statuses when appropriate, and prevent duplicate records when retrying operations.

6. Metrics for deciding whether to expand

Establish a baseline for the current process and compare equivalent cases. Do not confuse higher processing volume with better control.

• Resolution time: from detection to decision; distinguish active review from time spent waiting for documents.
• Complete files: the proportion with all required supporting documentation at approval.
• Correct assignment: the proportion of exceptions routed to the right person on the first attempt.
• Reopened cases: decisions revisited because of new evidence, errors, or ERP changes.
• Control errors: files that moved forward improperly and unjustified blocks identified through review.

Break down results by exception and responsible department. Before the pilot, agree on which results support expansion, which require adjustments, and which incidents require stopping it. There is no universal threshold that works for every company.

7. Conclusion: automate controls, not uncertainty

AI invoice approval adds value when it turns discrepancies into clear tasks with an owner, supporting evidence, and a traceable decision. Rules set the boundaries; AI helps people understand the file; authorized people resolve exceptions.

If your team spends too much time clarifying discrepancies before payment, contact Slashweb to evaluate a pilot. The starting point is to review your files, policies, and ERP to define what to automate and what must remain subject to human approval.