Skip to content

How customer data is handled ​

Overdesk reads conversations from your help desk, drafts replies with a language model, and writes the drafts back to your help desk unsent. This page follows a conversation along that path and says what is protected at each step, including where the protection stops.

Where your data goes ​

DestinationWhat it receivesWhose account
AnthropicConversation text with recognisable details replaced, for drafts, triage, follow-up nudges, reports, analyses, Ask Overdesk, and the issue text you ask it to writeYours, through your Anthropic API key
Voyage AIKnowledge-base text and search queries, to power knowledge-base search. Search queries, past conversations, saved replies and connected sources have recognisable details replaced first; your documentation site and help-center articles are sent as publishedOverdesk's own account
Your help deskThe finished draft, unsent, with the real details put backYours

Apart from the providers that run Overdesk itself, such as hosting, the database, email delivery and payments, other services receive data only once you connect them. A connected store, billing or licensing system is sent your customer's email address so Overdesk can look up their account; see Customer context. Reports and alerts you send to Slack or ClickUp show "(customer)" in place of the customer names Overdesk recognised, and contact details stay replaced. The Privacy Policy lists every sub-processor and where each one runs.

What your Anthropic account does with a request is governed by your agreement with Anthropic, because the request is made with your key.

What is replaced before the model sees it ​

Before any conversation text is sent to the model, Overdesk replaces recognisable personal details with stand-ins. It finds them by their shape:

  • email addresses, and phone numbers, international (written with a + and a country code, or with a leading zero and a space, dash or dot between the groups of digits) and North American (the area code in brackets or the groups of digits split by spaces, dashes, dots, en dashes or non-breaking spaces), also when the next word or an extension is written straight after them or a separate number, such as a year, is one space away
  • payment card numbers, IBANs and cryptocurrency addresses
  • social security numbers written with dashes or spaces, or as a run of 9 digits in text that mentions an SSN, and MAC addresses
  • licence keys, API keys and similar codes, cloud storage secret keys written beside their access key ID, and passwords written beside their label or their username, as in login: admin / Hunter2x9! or user bob, pass Hunter2pass (when the value is one unbroken string that mixes letters and digits, a run of six or more digits, or a string of 12 or more different characters, fairly evenly spread, that mixes upper and lower case with a symbol from ! @ # $ % ^ & *), a WordPress application password written without a label (six groups of four letters and digits), including a key in a WordPress define() line or a quoted setting, and a Freemius key with symbols such as # or @ in it
  • every other copy, in the same text, of a key or of a password written beside its label or its username, written on its own or after its label, even a label run into the word before it
  • private keys pasted with their BEGIN and END lines: the key between those lines is replaced, and the two lines stay
  • access tokens written after "Bearer", and the credentials in an "Authorization: Basic" header
  • the secret in a link that logs in to a site, such as a temporary login link's token, a WordPress reset key, a magic login link's token, a session id or a signature. The name before the = stays, so you can still tell what kind of link it was. On a login page, such as wp-login.php, a reset or magic login address, or a link whose address asks for a reset, and on an admin page under /wp-admin/, any value of eight or more characters under a name such as key, token, code or session is replaced, and so is a run of 16 or more hexadecimal characters mixing letters and digits under any name. Elsewhere, such a value is replaced when it reads as a token: letters and digits mixed, not words joined by dashes, so a slug such as summer-sale-2024 stays. A customer who posts a temporary login link, a reset link or a magic login link has their reply held for your team (see Escalation)
  • PO boxes, and street addresses introduced by a phrase such as "address:" or "lives at"

Names are replaced too when they come from a structured field: the customer who opened the conversation, anyone else on the customer's side who wrote one of its messages, and the name a connected store holds for them.

When the model replies, Overdesk puts the real values back, so the draft your agent reads is complete. A key the customer sent, a password or code written beside its label or its username, or a WordPress application password written without a label, is the exception: if the reply repeats one, it stays a placeholder and the reply waits for your team, so Overdesk never writes it back to the customer. Redaction is on by default for every workspace.

What it does not catch ​

The replacement works from patterns and structured fields, so it has limits. Treat it as a strong safety net, not as a substitute for your team's judgement about what belongs in a ticket.

  • Names written only in the text. Overdesk runs no name-recognition model, so "my colleague Dave can't log in either" reaches the model with the name in it.
  • A phone number typed as one bare run of digits with no spaces or symbols. It is left as written, because in a live ticket it is far more often an order number. So is a number split by slashes, and so are a few edge forms: a country code in brackets, a single digit after the leading zero, a first group of seven digits followed by one group of three, a number starting with zero that a word is written straight before, and two numbers written one space apart with nothing between them.
  • A password written without a label or a username beside it, which reads like any other word. Writing it after "login" with no colon counts as no label, and so does joining it to its username with a bare slash after "login:" (login: admin/Hunter2), which reads the same as a path, or writing it as a word and its number after a way to sign in, such as Login: Google / Office365. A password written beside its username but joined to it by a comma, a dash, a tab or a line break, a WordPress application password written without a label but split across lines, and a copy run into the word after it are left as written too.
  • A password written beside its label that is a number under six digits, an ordinary word, a run of words joined by dashes such as sunset-harbor-lamp-quarry, or a passphrase with spaces. Only an unbroken string that mixes letters and digits, a run of six or more digits, or a string of 12 or more different characters, fairly evenly spread, that mixes upper and lower case with a symbol from ! @ # $ % ^ & *, is replaced. One with a symbol such as ? or - instead, or with fewer than 12 different characters, or one that leans on a few of them such as CorrectHorse!Battery, is left as written.
  • A key's vendor prefix. The secret part of a key is replaced, but a prefix such as "FWG-" or "sk-proj-" can stay readable.
  • A short key with no label beside it. One under about twenty characters, such as A1B2C3D4E5F6G7H8I9J, can be left as written; beside a label such as "key" or "licence" it is replaced.
  • A password written beside a label in another language, such as "mot de passe:" or "contraseña:". Only labels such as the English pwd, pass, pw, passwd, passcode and passphrase, the German Passwort and Kennwort, the Dutch wachtwoord, and a prefixed name such as a mail or database setting that ends in one of them are read. A value under four characters is left as written.
  • A WordPress application password written without a label in which four or more of its six groups read as words, which leaves fewer than 1 in 1,000 as written.
  • A code written beside its label that holds an "@", which is read as an email address, or is a valid card number: it is replaced, but a reply that repeats it gets the real value back. A run of digits beside a label that is not a valid card number is left as written.
  • A cloud storage secret key, such as an AWS secret access key, with no label and no access key ID beside it. Written after a label such as "secret key:", or beside its access key ID the way the AWS console lists them, it is replaced whole. On its own it reaches the model, and Overdesk does not send a reply that repeats it (see What is never sent).
  • A key in configuration whose label Overdesk does not read, such as a serialized PHP option, an XML element, a table cell, or a value on the line below its label behind a quote marker, and a password written without a label Overdesk reads in the same places. Either is still replaced when the value looks like a key on its own, such as one starting with a known prefix like sk_, or a long run of letters and digits mixed together.
  • A PuTTY key file, which has no BEGIN and END lines. Long parts of it are replaced; the rest can reach the model.
  • A value written straight after "Authorization:" with no "Bearer" or "Basic" in front of it. Payment providers write their authorization IDs that way, so it is left intact.
  • Your own team's names. They go to the model as written, so it can follow who said what on a conversation several agents have answered.
  • Links, IP addresses, postcodes and street addresses with no lead-in. Their patterns collide with version numbers, ticket numbers and ordinary words, so they are left intact. Treat a reset or invite link pasted into a ticket as something the model will see.
  • Pictures and PDFs. Redaction reads text and cannot read inside an image, so an attachment reaches the model exactly as your customer sent it while a draft is written. You can switch this off, and turn off reading the pages customers link to, on the Customize AI page. See Pictures and links.

Stand-ins are unique to your workspace ​

The stand-in for a value is derived from a secret that belongs to your workspace. The same email address produces a different stand-in in every workspace, so stand-ins cannot be matched across workspaces.

What Overdesk stores ​

  • Your knowledge base. Past conversations are redacted before they are stored, and the customer's email address is replaced with a stand-in that stays the same each time, so their conversations still group together. Name replacement from structured fields runs on text headed for the model, not on the stored copy, so a name written in a ticket's text is stored as written.
  • Drafts. Each draft is stored with the real details put back, because it is the reply your team reads on the Agent drafts page and in your help desk.
  • A record of each reply Overdesk sent: the conversation and message it went to, and a fingerprint of the reply, never its text. It is kept once the draft ages out, so Overdesk never mistakes its own reply for your team's.
  • Test site runs. The record of what a draft tried on a test site and what the test site showed, with contact details replaced. Screenshots of a test site are shown to the model and not stored.
  • Your product zip files, the ones you add on the Sandbox tab. They hold your code, not customer data.
  • Reports and analyses. The saved copy shows the customer names the model mentioned, and contact details stay replaced. The copy sent to Slack or ClickUp shows (customer) instead.
  • Findings, the issues an analysis recorded. They show (a customer) wherever the model named someone, in every copy, including what the finding webhook sends, and point to their conversations by number rather than quoting them. An investigation brief for a finding's issue is not stored: it goes into the issue form, and to your tracker only when you open the issue.

How long each of these is kept is on Data retention. To remove one customer's data, see Erasure and deletion requests.