What Mesh removes before the AI writes your notes

A meeting reaches Mesh as audio or as a transcript. Speech-to-text turns the audio into text first. From there, the text goes through two changes before the AI steps that write your summary, action items and compliance flags read it.

Mesh keeps the transcript in two forms. The full text is stored encrypted. A redacted copy, with sensitive numbers replaced, is the one those AI steps read. Before the AI runs, Mesh also swaps the full names it knows, and most first names, for placeholders. When the AI answers, Mesh puts the real names back and stores the summary encrypted.

Redaction looks for five kinds of numbers and replaces each one with a marker, such as [ACCOUNT_REDACTED]. It is pattern matching, with checks that cut false alarms: a card number must pass the Luhn test, and a routing number must pass the ABA checksum.

On those AI steps, the same check runs again just before Mesh sends the text to the model. If a pattern still matches, the call stops. It does not go out with the number in it.

Some numbers stay on purpose. A summary needs the amounts, the percentages and the dates. Leaving phone numbers alone keeps one from being mistaken for a Social Security number.

Names are not numbers, so redaction leaves them alone. A second step replaces them. Before the AI runs, Mesh lists the names it knows for that meeting: yours, the client’s, and those of the other participants and speakers. Each full name maps to one placeholder: PERSON_1, PERSON_2, and so on. For most two-part names, the first name maps to the same placeholder. A surname by itself does not.

That list lives in memory for one run and is never saved. When the AI answers, Mesh turns the placeholders back into real names before it stores anything, so your summary reads like a summary.

The text Mesh sends out to build your search index gets a stricter step: each full name Mesh knows becomes the same word, SPEAKER, with no way back. The search index that Mesh stores keeps the redacted text, with the real names in it.

Redaction protects what the AI reads. Field encryption protects what Mesh stores. Mesh encrypts named fields one at a time, before they reach the database, so a copy of the database on its own shows only ciphertext for those fields.

Each data key is bound to your workspace in AWS KMS. KMS itself refuses to unwrap that key for a request that names a different workspace. Mesh holds the one KMS key that protects every data key, so your firm has no key to manage.

Every protection here has an edge. These are the ones to know.

• Speech-to-text hears the whole call. Redaction works on text, so the services that record and transcribe a call receive the audio, numbers and names included.
• Patterns miss some numbers. An account number counts only when a word like “account” comes right before it, as the second line of Fig. 3 shows. A number spelled out in words does not match at all.
• One step reads the full transcript. To fill in a client’s date of birth and mailing address, a narrow AI step reads the full transcript, with numbers and names as said. It returns only those two fields, and Mesh stores both encrypted.
• Placeholders cover full names and most first names that Mesh knows. A surname on its own, as in “Mr. Caldwell,” stays as said, and so does a name that comes up only in conversation. When two people share a first name, Mesh replaces only their full names.
• Not every field is encrypted. Client names, meeting titles and times are stored as plain fields. So is the redacted copy of the transcript, including the pieces of it kept for search, with names and speaker labels as said. A few coarse profile values, such as risk tolerance, time horizon, income range and employment status, are plain fields. Uploaded files sit in private file storage, but Mesh does not encrypt them field by field, and their file names are plain fields.

Mesh does not train AI models on your data. To report a security issue, write to security@meshfp.com.

Request early access to try Mesh on your own client meetings. Nothing goes to a client until you approve it. Notes and tasks go to Redtail when you push them, or when a workflow you set up runs.