PUBLIC SECTOR & FRAUD
Tax fraud detection: audit the right returns first
Tax fraud detection with TIMi estimates, for each declaration, the probability of an irregularity, to prioritize audits, and shows the factors behind each score.
Risk score
Each declaration gets a score, to decide which ones to check first.
Sovereign by design
Runs on a standard server, with no external cloud.
Documented per file
Each score comes with the factors that justify it.
HOW IT WORKS
How tax fraud detection works.
Three steps, in one engine, on a standard server with no cluster.
1. Prepare the data
Anatella gathers the declarative data, cross-checks with other available sources (registers, third-party declarations) and the history of audits already carried out, then structures them into variables.
2. Build the model
Modeler compares the profile of declarations that led to a confirmed adjustment with declarations in progress, to estimate a probability of irregularity.
3. Score and explain
Each declaration gets a risk score with its factors, such as a gap between the declaration and external cross-checks, or a profile close to confirmed adjustments.
Anatella
ModelerGETTING IT RIGHT
Choices that shape an audit model.
The model helps to prioritize. People decide.
Prioritize, do not decide
The model ranks the files to look at first. The decision to audit and to adjust stays a human one, documented by the factors the model identified.
Learn from confirmed adjustments
A history of confirmed cases is needed. The more representative it is of the situations you meet, the more reliable the model is on new cases.
Keep the data sovereign
The model runs on a standard server, with no cluster and no migration to an external cloud, which matters for public administrations handling sensitive data.
EXPLAINABILITY
A score says audit. The factors say what to check.
A risk score alone says a declaration should be checked, not why. A white-box model shows the factors behind the score for this file, so auditors focus on the lead identified instead of starting the file over.
For audit teams
Handling time per priority file goes down, because the triggering factor is already identified.
For contestable decisions
A model that justifies its scores variable by variable helps to document an audit decision, which matters for an administration subject to transparency requirements toward taxpayers. Whether it meets the legal requirements stays with your legal team.
For model owners
Fraud schemes change over time. Modeler’s built-in Drift Report shows when a model loses accuracy, so you update it when needed instead of on a fixed calendar. It is advisable to review the model after each filing campaign.
DATA
What data a tax audit model uses.
Anatella connects to more than 80 sources, so this data can be gathered without prior custom development.
Declarative data
Income, expenses, applicable regime.
External cross-checks
Registers and third-party declarations, where available.
Audit history
Audits and adjustments already carried out, as a learning base.
Context
Sector or geographic data.
PROOF
Closest references.
We do not yet have a published customer story on tax fraud. These two public sector deployments show the same constraints: sensitive national data, a standard server and sovereignty.
30 TB processed on national infrastructure
Colombia’s Ministry of Health handles 400 million new records a year on a sovereign infrastructure, with no external dependency.
A sensitive social budget modeled on one computer
A French département forecast a budget line on sensitive social data with no external dependency and no additional infrastructure.
FAQ
Common questions.
Does the model replace the judgment of an auditor?
No, it prioritizes the files to examine first. The decision to audit and to adjust stays a human decision, documented by the factors the model identified.
Do we need a history of confirmed adjustments?
Yes. A history of confirmed cases is needed. The more representative it is of the situations encountered, the more reliable the model will be on new cases.
Can it handle volumes at the scale of a country?
TIMi’s architecture is designed to process large volumes on a standard server, with no cluster and no migration to an external cloud, which addresses the sovereignty concerns of public administrations.
Is there a worked example?
The TIMi Quick Guide includes a didactic example on a public census income dataset, where people declaring a low income but with a high predicted probability of high income are flagged as suspicious. It illustrates the method. It is not a customer case.
SEE FOR YOURSELF
Test it on your own data.
Run TIMi on your own data with a free trial, or talk to us about your case.
