How accurate is Verho?

We test Verho on texts written to look like the messages people send at work, to a colleague, a client or an AI service. They are typed the way people type in a hurry, with typos, missing capitals and not much punctuation. Every name and number in them is made up.

The table shows how much of each kind of personal data Verho found. A value counts as found if Verho marked it, even when the mark was a word too long or too short. The small number under each share is how many there were in the texts.

Measured on our server in Helsinki, with the model Verho uses today, on 9 November 2026.

Kind of data	Finnish	Swedish	English
Names	98.6 %of 360	97.8 %of 271	100.0 %of 236
Personal identity codes	100.0 %of 23	100.0 %of 16	100.0 %of 12
Addresses	100.0 %of 44	100.0 %of 39	100.0 %of 45
Health information	92.9 %of 113	90.8 %of 131	87.6 %of 105
Salary information	96.9 %of 64	97.8 %of 91	98.7 %of 80
Bank accounts	100.0 %of 13	100.0 %of 13	100.0 %of 12
Organisations	100.0 %of 68	98.5 %of 67	96.7 %of 61
Email addresses	100.0 %of 42	100.0 %of 44	100.0 %of 51
Phone numbers	96.9 %of 32	100.0 %of 33	100.0 %of 37

No tool finds everything. That is why Verho shows you what it marked before you send, so you can add anything it missed.