evidoria

← Back to browse

Good practice Imported

SAT's AI-Driven Tax Fraud Detection and OrientaSAT Chatbot — Mexico

Mexico · Mexico City · See the Mexico profile · See the Mexico City profile

Evidence: Observational / pre–post Top 74% 47/100 · Ask Evidence Copilot about this practice

Mexico's tax authority uses AI to flag fraudulent e-invoicing and shell companies, and runs OrientaSAT, a chatbot serving taxpayers — but an academic review flags bias risk, opacity and a 22% staffing cut alongside the rollout.

~1,200 companies
Shell companies identified (2020)
~3,500 transactions
Fraudulent transactions identified (2020)
47.5% increase to 246,309 interactions
OrientaSAT interactions growth (2023)
25.4% increase
Audit efficiency increase (Q1 2024)
201.6 pesos recovered per peso invested (vs 160.8 a year earlier)
Enforcement recovery ratio (Q1 2024)
-22% 7,004 positions
SAT headcount reduction (2022)

Details

Maturity
Established
Promoter
Servicio de Administración Tributaria (SAT)
Period
2017–present
Keywords
taxation, fraud detection, e-invoicing, chatbot, public finance

Context

Since October 2017, Mexico's Servicio de Administración Tributaria (SAT) has applied machine-learning models to its mandatory nationwide electronic-invoicing system (CFDI), comparing declared income and VAT bases against issued invoices to flag simulated transactions and shell companies ("empresas fantasma").

Objectives

The programme aims to detect fraudulent e-invoicing and shell-company activity across the whole tax base, and since March 2021 also to assist taxpayers directly through OrientaSAT, a free AI chatbot available during filing season.

Activities

SAT's models compare declared income and VAT bases against issued invoices nationwide, while OrientaSAT answers taxpayer queries; both feed into the agency's broader 2024 "Plan Maestro" for AI-assisted risk classification and enforcement.

Results

In 2020 the models were credited with identifying roughly 1,200 shell companies and about 3,500 fraudulent transactions, and OrientaSAT's interactions grew 47.5% to 246,309 in 2023. By Q1 2024 SAT reported a 25.4% increase in audit efficiency and a recovery ratio of 201.6 pesos per peso invested in enforcement, versus 160.8 pesos a year earlier.

Conclusions

A peer-reviewed UNAM-affiliated study raises concerns about algorithmic bias against specific taxpayer groups, non-disclosure of the risk-scoring methodology and unclear legal accountability, and notes that SAT's headcount fell by 7,004 positions (-22%) in 2022 alongside the automation push; SAT's own causal claims about AI-driven revenue gains have not been independently audited.

Implementation
Indicative cost
High (€500k–€5M) — No public cost figures were found; conservatively banded high given the system's nationwide scale across Mexico's entire e-invoicing base and continuous operation since 2017.
Time to results
Long (> 3 years) — Tax-fraud detection models have run continuously since October 2017; OrientaSAT added March 2021; classified as long-running given nearly a decade of continuous operation.
Staffing & skills
SAT tax administration and IT/data-science staff running CFDI risk models, OrientaSAT chatbot support team

Conditions for success

  • Mandatory nationwide e-invoicing (CFDI) data already in place to train and run risk models
  • Continuous system operation since 2017 giving multi-year data for model refinement

Common failure modes

  • A peer-reviewed UNAM study documents undisclosed risk-scoring methodology and potential algorithmic bias against specific taxpayer groups, with unclear legal accountability
  • SAT's headcount fell 22% (7,004 positions) in 2022 alongside the automation push, raising questions about capacity and oversight
  • SAT's own revenue and efficiency gains attributed to AI have not been independently audited

Where it fits

Governance type
national tax authority (centralised federal agency)
Scale
national, entire mandatory e-invoicing base
Income level
upper-middle-income

Commonly funded by

National / regional programmes

Indicative funding routes for practices of this type — always check each programme's current calls and eligibility rules.

Do you run this practice? Claim it — verified implementers get a public contact pathway and can propose corrections.

Data sources

Where this practice's information was retrieved from, and when.

Similar practices you may find useful