How to Anonymize Personal Data Before Sending It to AI

The short answer

Anonymization means replacing or removing anything that identifies a person before the text reaches an external AI service. The working pattern: detect personal data → replace it with labels or pseudonyms → send the request → if needed, restore the real values in the response on your own side.

What counts as personal data

PII (personally identifiable information) usually includes:

• full name, date of birth, photo;
• phone number, email, address;
• passport data, PINFL, tax ID;
• card and account numbers;
• health and income details;
• indirect attributes: a job title in a small company, a rare diagnosis, an exact workplace.

Indirect attributes are dangerous because individually they reveal nothing, but together they point to a specific person.

Step 1. Detection

Use several layers:

• Regular expressions — for formatted data: phones, emails, document and card numbers.
• NER models (named entity recognition) — for names, addresses, organizations. Quality on Russian and Uzbek text must be tested separately.
• Dictionaries — lists of your clients, employees, partners.
• Structured fields — if data comes from a CRM, simply do not send unnecessary fields.

import re

PHONE = re.compile(r"\+?998[\s-]?\d{2}[\s-]?\d{3}[\s-]?\d{2}[\s-]?\d{2}")
EMAIL = re.compile(r"[\w.+-]+@[\w-]+\.[\w.]+")

def mask(text: str) -> str:
    text = PHONE.sub("[PHONE]", text)
    return EMAIL.sub("[EMAIL]", text)

Step 2. Choosing a technique

With pseudonymization, the mapping table lives only on your server. The AI sees "Client_1", and your system puts the real name back after the response.

Re-identification risks

A person can be recognized even without a name:

• through a combination of attributes: district, age, profession, event date;
• through free text where customers mention details about themselves;
• through unique events — a rare deal, a well-known incident;
• when pseudonyms are identical across requests and can be linked.

Reduce the risk: send the minimum of fields, generalize exact values, rotate pseudonyms between sessions when requests do not need to be linked.

Requirements in Uzbekistan

Uzbekistan has a Law "On Personal Data". Among other things, it requires personal data of Uzbek citizens processed with information technologies to be stored in databases physically located in the country. Practical takeaways:

• keep the original personal data on servers in Uzbekistan;
• send only anonymized text to external AI services;
• obtain the data subject's consent and record the purposes;
• review the AI provider's contract and data retention policy.

Agree on the exact setup with a lawyer: requirements and their interpretation change.

Common mistakes

• Relying on regular expressions alone — they miss names and addresses.
• Sending the mapping table to the AI along with the request.
• Writing raw data to logs before masking.
• Never testing anonymization on real samples.