AI Corporate Card Reconciliation: CFDI and ERP

AI corporate card reconciliation aims to solve a specific problem: the employee has already paid, the charge appears on the card, but Finance still needs to find the CFDI invoice, identify the expense, and verify its entry in the ERP. Automating that matching process takes more than reading receipts: it requires connections, rules, and designated people to resolve discrepancies.

On November 3, 2026, Expensify announced a referral agreement with Visa that includes transaction syncing, AI categorization, receipt matching, and accounting integrations for eligible customers and programs. This is a commercial agreement, not a new jointly developed integration; the announcement does not confirm coverage for Mexican cards or CFDI support.

For a Mexican company, the practical question is what it needs to connect to reconcile expenses using its current cards and systems, which decisions it can automate, and which should remain under human review.

1. What This Process Reconciles—and What It Does Not

The starting point is an expense already paid with a corporate card, not a vendor invoice awaiting payment authorization. The goal here is not to approve orders or schedule transfers, but to connect three pieces of evidence:

• Card transaction: amount, currency, merchant, dates, and transaction reference.
• Supporting documentation: CFDI invoice in XML format, along with receipts or supplementary documents when applicable.
• ERP record: expense, general ledger account, cost center, responsible person, and journal entry reference.

It is useful to separate two statuses: “charge matched to supporting documentation” and “entry confirmed in ERP.” Finding a CFDI invoice does not prove that the expense has been recorded correctly.

This workflow should not be confused with paying the card balance, either. If the expense is already recorded, the integration should link to it—not create another expense or generate a new payment obligation to the merchant.

2. Connections for Linking Transactions, CFDI Invoices, and ERP Records

Before choosing technology, review these four connections. Not all of them need to operate in real time for the pilot to be useful.

Card Transactions

Confirm what access the bank or card issuer provides: an API, periodic files, or exports. Verify program coverage, update frequency, identifiers, and availability of refund data. Distinguish pending authorizations from posted charges to avoid processing the same purchase twice.

Keep both the transaction date and the posting date: they may differ. For foreign currency transactions, store the original amount separately from the amount charged in the account's currency, when available.

Supporting Documents and Responsible Employees

Centralize XML files and supplementary documents through an authorized mailbox, portal, or repository. The XML should be the source of structured CFDI data; applying OCR to its PDF representation adds an unnecessary extraction step.

Link each card to its responsible employee, company, and cost center. If the card is shared, request evidence of who made each purchase.

ERP Master Data and Records

Integrating CFDI invoices with an ERP requires access to existing expenses and current master data. Check whether the integration can attach evidence, update references, and receive confirmation that data was written. If it requires files, define how you will track the outcome of each import.

An AI integration should start with what is actually available, rather than assuming that every bank or ERP has a ready-made connector.

3. AI Corporate Card Reconciliation: Rules First

AI adds value when information is ambiguous: an abbreviated bank description, a merchant whose name differs from its legal business name, or an expense description that is difficult to categorize. Exact validations should remain governed by verifiable rules.

Task	Recommended Approach	Required Control
Extract UUID, RFC tax ID, total, and currency from XML	Structured parsing	Validate format and required fields
Detect duplicate files and transactions	Identifiers and rules	Prevent duplicates during retries
Match the merchant to the CFDI issuer	AI and a mapping table	Show evidence supporting the proposed match
Suggest an account or cost center	AI using approved master data	Review ambiguous cases
Confirm amounts and currencies	Explicit rules	Document any discrepancy
Write to the ERP	Controlled workflow	Authorization and confirmation of the result

Matching amounts and dates alone is not enough: identical purchases can occur. The system should consider the responsible employee, merchant, currency, timing, and available documents.

A model's confidence score is not a substitute for evidence. Automatic reconciliation should be limited to combinations tested against reviewed data. Everything else goes into an exception queue as part of AI process automation with defined controls.

4. How to Handle Missing Documents, Refunds, and Discrepancies

The goal is not to force everything to balance, but to separate clear-cut cases from those that require a decision.

Missing Supporting Documentation

Keep the charge marked as awaiting documentation, assign it to the responsible employee, and send a request that includes the purchase reference. A receipt can help identify the expense, but it should not be artificially converted into a CFDI invoice. Finance determines how to handle cases that reach period-end close without sufficient documentation.

Split Charges or Multiple Supporting Documents

A purchase may generate multiple charges or be supported by several documents. Allow one-to-many and many-to-one relationships, recording how much of each transaction and supporting document is allocated. Controls should prevent the same amount from being reused to substantiate another expense.

Refunds and Cancellations

Link the credit to the original charge without deleting the history. A partial refund leaves a different balance than a full refund. The financial transaction and the CFDI status must also be reviewed separately: one does not automatically establish what happened to the other.

Amount Discrepancies

Tips, currency conversion, or additional charges may explain discrepancies, but should not be assumed without evidence. Define acceptable reasons, documentation requirements, and who authorizes adjustments. Any tolerance must be explicit; it should not hide discrepancies or determine tax treatment on its own.

5. Human Review and Safe ERP Writes

The reviewer needs a screen showing the charge, the proposed CFDI match, the accounting record, and the reason for the suggestion. They should be able to confirm, reject, split the match, or request information without reconstructing the case across emails and spreadsheets.

Before writing to the ERP, apply these controls:

• Document validation: XML structure, expected recipient RFC tax ID, and a CFDI status check through an authorized mechanism.
• Duplicate prevention: check existing records and use a stable key for each transaction.
• Least-privilege permissions: separate read access, proposed changes, and write authorization.
• Audit trail: retain evidence, changes, the responsible user, and the ERP response.
• Data protection: limit what is sent to the model and avoid full card numbers or other unnecessary data.

Document validation does not determine tax deductibility. Tax decisions belong to the responsible department. When existing systems do not provide the required workflow, custom software can provide the connection and review queue without replacing the ERP.

6. A Focused, Measurable Pilot

Start with one entity, one card program, a group of employees, and a closed period. Select representative expenses and include exceptions; testing only straightforward matches would produce an incomplete evaluation.

The pilot can be organized into four steps:

1. Build a validated reference set. Finance reviews the sample and establishes the correct matches.
2. Run in shadow mode. The system proposes links without changing accounting records.
3. Analyze errors and unresolved cases. Adjust rules, mappings, and documentation requirements.
4. Enable supervised writes. Move forward only when agreed-upon criteria are met and a correction procedure is in place.

Measure these separately:

• Correct reconciliations: matches confirmed against the reviewed sample; report incorrect matches as well.
• Coverage: correctly reconciled transactions as a share of the total included.
• Unresolved cases: number, age, and cause, such as missing documentation, ambiguity, or an integration error.
• Review time: actual minutes per case, compared with the previous process using equivalent samples.
• Confirmed records: transactions accepted by the ERP without duplicates.

This lets you distinguish a suggestion that appears accurate from a reconciliation that has been verified and recorded.

7. Conclusion: Connect Evidence Before Automating Decisions

AI corporate card reconciliation works best when it combines structured data, accounting rules, and exception review. AI helps suggest matches and classifications; controls determine when to accept them.

At Slashweb, we can help you design a pilot to connect transactions, CFDI invoices, and ERP records, with a defined scope and metrics. Let's discuss your reconciliation process to identify what to automate first and what evidence you need before scaling.