Every filing season someone in your firm opens a chatbot and pastes in an XML. I understand why. CFDI volume never drops, the team is looking to save hours wherever it can, and the tool answers in seconds. My job is not to tell you to avoid AI. It is to tell you where it genuinely helps and where it creates problems that cost more than the hours you saved.
AI in accounting is not one single thing. Classifying invoices is a different problem from preparing a tax return, and treating them the same is the most expensive mistake I see. One is mechanical work you review at the end. The other carries your signature and your client's responsibility before the SAT. Let's take them in order.
Where it helps: classifying and sorting CFDI
CFDI 4.0 has been the only valid invoice format since April 1, 2023, and every document carries structured fields — issuer and receiver RFC, CFDI usage, product or service key, payment method and form, tax breakdown. That repetitive work of reading, grouping, and tagging hundreds of XML files is exactly where a well-configured model earns its keep. Sorting expenses by category, catching duplicate invoices, flagging the ones whose CFDI usage doesn't match the operation. It doesn't replace your judgment; it takes away the part that bores you.
The rule I use is simple: AI proposes, you validate. A model can sort 800 invoices in minutes and get 30 of them wrong. If your process assumes you review those 30, you saved real hours. If it assumes the model doesn't make mistakes, you already lost — you just don't know it yet.
Reconciliations and report drafts
Reconciliation is the other place AI behaves well. Cross-checking issued CFDI against recorded income, comparing what your books say against what the client captured, pointing out where a discrepancy sits. AI is good at flagging the discrepancy. It is not the one that decides what that discrepancy means or how it gets fixed. That is still your read.
Same with reports. A draft of the monthly report for the client, a summary of movements, the first version of an explanatory note. There AI gives you a starting point you edit, not a finished document you send unread. Think of it as a fast intern with good writing who occasionally invents things with total confidence. Useful, always reviewed.
Where it is risk: the filing and fiscal judgment
Here my tone changes. A tax filing is not a draft. The SAT no longer reviews by hand — in its 2018 to 2024 management report it announced that its next steps include predictive models and artificial intelligence in the development of its systems. It cross-checks CFDI against filings in near real time. And if it detects inconsistencies between your invoices and what was declared, it can cancel the Certificado de Sello Digital, meaning it can leave your client unable to invoice at all. A badly built automation doesn't cost you an afternoon; it shuts down the operation.
The responsibility is always yours. Neither the PAC nor the software nor the model carries the error. The PAC validates that the CFDI meets the format and stamps it, but the obligated party is the taxpayer, and you are the one confirming the numbers and the judgment are right. A language model doesn't know your client's particular case, hasn't read this year's Miscelánea, and will sound convincing even when it is wrong. Interpreting a rule, deciding how to treat an operation, fiscal judgment itself — that is not automated, it is backed by your name.
SAT data does not go into public tools
An RFC, a name, a tax domicile, a client's amounts are personal and fiscal data. Pasting them into the public version of a chatbot is a transfer of data to a third party, with everything that implies. INAI's cloud computing guide is clear about what a provider must meet: not claim ownership of the information, maintain confidentiality, and allow deletion of the data when you're done. The free version of most chatbots gives you none of that in writing.
- For client work, use a tool on enterprise or API terms with a commitment not to train the model on your data. Never the personal account someone set up at home.
- When you can, anonymize. Often the model doesn't need the real RFC or name — it needs the shape of the problem.
- The most sensitive material, protected by professional secrecy or that the client asked you to keep confidential, does not enter a public model. That line gets written down and respected.
What actually changes the outcome
The difference between a firm that uses AI well and one that gets into trouble is not the tool — it is the team's judgment. Knowing, without a second thought, which task a model can touch and which it can't. Where AI proposes and where you decide. What data can leave the firm and what never does. That doesn't come with the software; it comes with training.
When I train a firm I don't start with the trendy tool, I start with those lines. A team that has them clear classifies hundreds of invoices in minutes, delivers drafts faster, and doesn't put anyone's digital seal at risk. AI in accounting is neither magic nor threat. It is a fast assistant that needs someone with judgment deciding what to trust it with. That judgment is teachable, and that is where the effort is worth putting.