Personal data in AI work

When a Mind or a builder works with a column that holds personal data, the platform replaces the real values with masked ones before the AI reads them. This page collects the rules the AI follows with those values, and what changes when your company blocks personal data from AI requests.

Masked values

A masked value is a stand-in for one personal value: a label names the kind of data, then a ten-character code, as in {{EMAIL_a1b2c3d4e5}}. You may also see them called platform tokens.

• The AI copies a masked value exactly as it is, both double braces and the whole code included; a single changed character breaks the mapping.
• Masking is stable per column: a value keeps the same code in its own column, and carries a different code in every other table and column.
• The values you see in the product stay the real ones. The platform puts them back into what it shows you, and the AI answers naturally rather than commenting on the masking.

What the AI can do with a masked value

• Filter, count and group. The AI treats a masked value as the value itself: it counts and groups by the column as usual, and a code copied from that column's own results works as an exact-match filter on that column, or in a filter that lists several values.
• Codes belong to their column. A code used anywhere else, such as one copied from chat text or from another table, matches nothing.
• No joins. The AI never joins two tables on a personal-data column, because the same value carries a different code on each side.
• No pattern filters, no sorting. Pattern filters, such as a contains or starts-with check, and sorting, read the code rather than the hidden value, so their results mean nothing. A masked column supports exact match only.

Reading and writing limits

• A builder's reads that include personal data return at most 50 masked rows per read.
• A masked value cannot be stored: writing one into a saved object, such as a measure, a name or a description, is refused, and the refusal names the field, never the value.

When your company blocks personal data

Your company can block personal data from AI requests altogether. The Company card in Settings carries Prohibit PII transmission, which only an Administrator can change. With it on, the platform checks every request made with your company's API key before it reaches a model, including the AI features of DataMind OS.

• A request that carries personal data is not forwarded: it fails with status 400 and code content_filter, and it uses no tokens.
• The error names the kinds of data it found, never the values.
• The check looks for email addresses, Georgian mobile numbers, payment card numbers, IBANs and Georgian personal ID numbers.

The option is described in full on Prohibit PII transmission; the error is listed in Errors and retries.

Questions, answered

Can the AI filter by a masked value?

Yes, when the code comes from that column's own results: it works as an exact-match filter on the same column, and in a filter that lists several values. Grouping by the column works too. A code from chat text or another table matches nothing.

Can the AI join two tables on a personal-data column?

No. Each table and column has its own codes, so the same value carries a different code on either side of a join, and the AI never joins on personal-data columns.

Do counts change because of masking?

No. The masked form stands in for the real value, so counting and grouping produce the same results as on the real data.

Why can't the AI search inside a masked column?

A masked column supports exact match only. Pattern filters and sorting read the code rather than the hidden value, so their results mean nothing.

Can the AI reveal the real value behind a mask?

No. The AI cannot decode or reconstruct a masked value: it copies the code exactly and uses it as the value for the work. In the product, the values you see stay the real ones; the platform puts them back into what it shows you.

Can the AI save a masked value into a model or a description?

No. Writing a masked value into a saved object, such as a measure, a name or a description, is refused; the refusal names the field, never the value.

How many personal-data rows can a builder read at once?

At most 50. A builder's reads that include personal data return up to 50 masked rows per read.

What does the content_filter error mean?

Your company blocks personal data from AI requests, and the request contained some. The request was not forwarded. Remove the personal data and send it again. See Prohibit PII transmission and the error list on Errors and retries.

Does the block cover DataMind OS as well as the API?

Yes. The check applies to every request made with your company's API key, including the AI features of DataMind OS.