A bank wants to use AI to detect fraud. A government agency wants to analyze census data with an AI model. In both cases the same question comes up, and it often stalls the project before it starts: how do you train that model while keeping sensitive data under the organization's control?
The dilemma of sensitive data and AI
Training an AI model normally means the model —and whoever administers it— can see the data it learns from. For personal, financial or health data, that adds a risk many organizations prefer to avoid, even when the technology provider is trustworthy. The usual alternative, keeping that data out of AI's reach, leaves a significant part of the value the technology could deliver on the table.
What is homomorphic encryption, in simple terms?
Homomorphic encryption makes it possible to perform computations on data that stays encrypted throughout the entire process. In simple terms: the AI model learns from the information, which remains encrypted at all times for anyone who processes it.
This is different from encrypting data that is stored but decrypted in order to be processed. Here the processing itself happens on the encrypted data, from start to finish.
What accuracy is gained, and what stays the same?
The logical question is whether this method affects the quality of the resulting model. In products like Xcapit's Privacy, the accuracy of a model trained on encrypted data is equivalent to that of one trained on plaintext data: the difference lies in the process, and the final result holds up.
Where is it already being applied?
Banks and fintechs use it to collaborate on fraud detection across institutions, keeping customer data under each one's own control. Government agencies use it to analyze census or tax data while keeping control of the information within their own systems.
Does your organization handle sensitive data and is it considering adopting AI today? This is one of the architecture decisions worth settling before choosing a provider.
José Trajtenberg
CEO & Co-Founder
Lawyer and international business entrepreneur with over 15 years of experience. Distinguished speaker and strategic leader driving technology companies to global impact.
Building regulated financial products?
From digital wallets to tokenized collateral and secure platforms, we build fintech systems designed for production.
You Might Also Like
Digital transformation for utilities: modernizing energy without replacing the core
A practical guide for utilities and energy companies: how to integrate SCADA, IoT, AI, tokenization, and cybersecurity to modernize operations without replacing critical systems.
Why Vulnerability Scanners Don't Replace a Pentest — and How AI Changes the Equation
Scanners like Nuclei and ZAP detect known CVEs but miss the vulnerabilities that actually get companies breached: IDOR, privilege escalation, race conditions, and business logic flaws. This article explains why, shows benchmark data (47 findings vs 0), and introduces a third option: AI-driven pentesting that reasons like a human attacker at scanner speed and cost.
OpenClaw Security Anatomy: What AiSec's 35 Agents Found in the World's Most Popular AI Agent
We ran AiSec — our open-source AI security framework with 35 specialized agents — against OpenClaw, the most popular AI agent on GitHub (191K stars). In 4 minutes and 12 seconds, it found 63 vulnerabilities mapped to 8 security frameworks. Here is the full technical breakdown.