Privacy notice

How ChatterLab handles data

Who is responsible

The controller for the data described here is SRI Group BV, Korte Gasthuisstraat 18 bus 301, 2000 Antwerpen, Belgium, enterprise number 1037.052.239, which runs ChatterLab. For any question or request about your data, write to management@sri-group.eu.

Two kinds of people

Customers create an account, build agents and pay for a plan. Visitors talk to an agent on a customer’s website. Customers are the controllers of what their visitors say; ChatterLab processes it on the customer’s behalf.

On chatterlab.ai and in the dashboard, the chat is an agent of our own, built with ChatterLab, and we are its customer: what you write to it, and the details you leave in its contact form or a ticket, are kept as a visitor’s are, below, with us as their controller.

What we store

  • Account. Your name, email address and sign-in identity, held by our authentication provider, and the workspace and agents you create.
  • Where you signed up. Which public page of this site you were on when you opened the sign-up page, as your browser tells us with that request. It reaches us through our authentication provider and we keep it with your workspace, in our legitimate interest in seeing which of our pages lead people to an account. Nothing is stored on your device for this, and we keep nothing about a site you came from before ours.
  • Getting started. In the first month after you sign up, we send the address on your account up to five mails that help you set up your agent, in the language of the page you signed up on. Which one you get depends on how far you got, so we read whether you made an agent, whether it learned from a source, whether you tested it and whether it is on a site; never what anyone asked it. The last one, on the Free plan only and to accounts made since the sign-up page says so, says what the paid plans add. We send them to carry out the service you signed up for and, for that last one, in our legitimate interest in telling a customer about our own, similar service, as we said when you signed up. The link in each mail stops them all, for good. We keep which mail went when, with your workspace.
  • Product news. Now and then we mail the address on your account what we improved in the product, in the language of your workspace, with your first name if your account has one. Only to accounts made since the sign-up page says we mail new accounts and that each mail can stop them, and in our legitimate interest in telling a customer about our own, similar service. The link in each mail stops them, and the getting-started mails with them, for good. We keep which edition went to which workspace, with your workspace.
  • Team members. Where a workspace invites colleagues, the address each invitation is sent to, the role it offers, who sent it and when, and the language it is sent in; once a colleague joins, their address and name as their account shows them, brought up to date as they use the workspace, and when they last used it. Colleagues read the workspace’s conversations, contact details and tickets as its owner does; what else they may do depends on their role. The link at the foot of every invitation mail stops all invitation mails to that address; we then keep a key derived from the address, not the address itself, so as never to mail it one again.
  • Installation steps. When a customer asks us to send the steps for putting an agent on their site to an address they type, such as their web developer’s, we send that address one mail through Resend, with a reply going to the customer. We do not keep the address: the audit trail keeps only its domain, and the count that limits these mails to it is a rate-limit hash.
  • Knowledge. The files, pages and text you give an agent, the passages they are split into, and a numerical representation of each passage used to find relevant ones.
  • Conversations. Every message a visitor sends and every answer an agent gives, with the model used, the passages it drew on, a confidence score and any rating the visitor left. A random identifier, made in the visitor’s browser when they first send a message or a file and kept only until the tab is closed, holds one visit’s messages and files together; it is not linked to a person unless the visitor leaves their details in the contact form or opens a ticket, below. On a site that signs its visitors in, the conversation is kept with that site’s own reference for the visitor.
  • The chat button. On a site that passes its visitors’ consent, and only while the customer has switched on its greeting or second message, the button keeps a small record of the visit in the browser until the tab is closed: when the visit began, how many pages were seen and which message was shown. It times the second message and is never sent to us. How often the chat is opened is counted per hour, with nothing that identifies a visitor, and each conversation records whether the button or one of its messages opened it.
  • Files visitors send. Where a customer allows it, a document a visitor attaches to a question, and the text read from it, kept for 30 days. After that only its name stays with the conversation.
  • Contact details visitors leave. Where a customer switches on the chat’s contact form, what the visitor enters in it (a name, an email address, a phone number or a company, as the customer chooses), whether they want news and offers, the words they were shown and the page they were on. They are e-mailed to the addresses the customer names and kept for 3 years.
  • Tickets. Where a customer switches on tickets, the e-mail address, name and question a visitor leaves for the customer’s team, and the page they were on. The visitor is sent one confirmation and the team’s answers by e-mail. The visitor’s replies by e-mail are received and added to the ticket, as is what the team writes, including notes only the team sees. A received mail is kept as its text; its attachments are not.
  • Mail to an agent. Where a customer lets an agent answer e-mail, the mail their customers send to the customer’s own address and forward to the agent: who sent it, its subject and its text, kept as a conversation, as one in the chat is, with the agent’s answer or its draft when none was sent. Attachments are not read or kept. Mail forwarded while this is switched off is not kept.
  • WhatsApp. Being tested with our own workspace and not yet offered to customers. Where it is switched on, a message someone sends to the business’s WhatsApp number reaches us through Meta’s WhatsApp Business Platform. We keep its text and the sender’s WhatsApp identifier or phone number as a conversation, as one in the chat is, with the agent’s answer, which goes back the same way. The name on the sender’s WhatsApp profile is not kept. Pictures, voice messages and other files are not read or kept; the sender is told that the agent reads text only. Where the business also uses the number in the WhatsApp Business app, what somebody at the business writes there to a person the agent is talking with is kept in that conversation, so that the agent leaves that chat to them; for a picture or a file only a line saying one was sent. What they write to anybody else, such as a private chat on the same number, is not kept. Nor are the contacts and the earlier chats Meta sends us when such a number is connected. The permission Meta gives for the number is kept encrypted.
  • Webhooks. Where a customer adds a webhook, each new lead and ticket, with the details in it, is sent to the address the customer names, which is theirs to choose and to answer for. What is sent is kept for 7 days, to try again while that address does not answer.
  • Helpdesks. Where a customer connects Freshdesk, Zoho Desk or Zendesk, a ticket is made there instead of here: the visitor’s address and name, the question, the page and the conversation so far go to the customer’s own helpdesk, which they chose and answer for. The key or the permission for it is kept encrypted.
  • Usage and billing. Message counts, tokens and credits per month, and a reference to your subscription. Card details never reach us; they are handled by our payment provider.
  • Audit trail. A record of significant events, such as an agent created, a source added or a plan changed, kept for 24 months, as the EU AI Act requires at least six.
  • Abuse protection. The public chat endpoints are rate-limited per visitor address. Only a fingerprint of the address is stored, for one day. It is made with a key that changes every day and is deleted minutes after its day ends; without that key the fingerprint cannot be turned back into the address. Like anything we delete, the key remains for a while in our database host’s backups.

Who processes it

  • Hosting: Vercel, in Frankfurt. Every request to the application passes through it, with the address it came from.
  • Database and files: Neon (Postgres) and Cloudflare R2, both in EU regions with the EU jurisdiction setting on storage.
  • Answers: models running in Europe, reached through OpenRouter, a US company (OpenRouter Inc.). The passages an answer draws on, the recent conversation, any file the visitor sent with the question and, on a site that signs its visitors in, what that site says about them (a name, an e-mail address, other details) are sent for each reply to its European address, where they are processed in Europe (the EU/EEA, the UK and Switzerland) or the request is refused. Its logging of them and its use of them are both off for our account. By its own documentation it otherwise keeps them only when content is flagged as possible child sexual abuse material, or on endpoints it calls Safety Retention, such as those for Google’s models, where it or the provider may keep them to prevent abuse. OpenRouter’s account, billing, authentication and security records, abuse-monitoring metadata and other operational records may be held in the United States. Its data processing agreement includes the EU Standard Contractual Clauses and an EU-only processing addendum, which we have elected.
  • Reading scanned pages: Cortecs, an Austrian company, through which a scan a visitor sends is read by Mistral’s OCR, within the EU.
  • Passage representations: OpenAI’s embedding service, in the United States, under its data-processing terms; API data is not used to train its models. The text of each passage is sent to it when a source is added or retrained, and the text of each question when it is asked, since a question has to be represented the same way to be matched. With e-mail, below, it is one of the two steps that run in the United States.
  • E-mail: Resend, which sends the reports and the contact details a visitor leaves to the addresses a customer enters, sends the mail about tickets, an agent’s answers to mail, the invitations to join a workspace, the installation steps a customer sends to an address and the mails that help a new account get started or warn that its credits are running out, and receives the replies to it and the mail forwarded to an agent. Resend processes mail in the United States, under the EU-US Data Privacy Framework and standard contractual clauses; the mail itself leaves from, and arrives at, servers in the EU (Ireland).
  • WhatsApp: Meta (Meta Platforms Ireland), whose WhatsApp Business Platform carries the messages between the sender and us, while it is being tested as described above.
  • Sign-in: Clerk, which also receives an invited address when its owner creates an account from the invitation’s mail. Payments: Stripe, including EU VAT handling. A VAT number a customer enters is checked with the European Commission’s VIES register before it is used, and again before each invoice, as the law on reverse charging asks.

Statistics on this website

The public pages of this site (the home page, the pricing and other information pages, the documentation and this notice) use Google Analytics to count visits: which pages are read, where visitors come from and roughly where they are. It runs only after a visitor says yes in the question at the foot of the page, and nothing of Google’s is loaded before that. A no is kept as well, and either choice is asked again after about six months.

With a yes, Google sets its own cookies (_ga and _ga_ followed by the property’s id) and receives the page address, the referring page, the browser and device, and the IP address, which Google says it does not store. Google may process this in the United States, under the EU-US Data Privacy Framework. The tag is told to use no advertising storage and no ad personalisation, so nothing is used for ads. The legal basis is consent.

The choice can be changed at any time with Cookie settings at the foot of every public page; taking a yes back removes Google’s cookies from this site. The dashboard, the chat widget on customers’ sites and the pages a token opens do not use Google Analytics, so visitors of a customer’s chat are never counted by it.

How long

Knowledge and conversations stay until the customer deletes the source or the agent, or asks us to close the account; deleting an agent removes its conversations and every uploaded file. Usage records stay with the account for billing. Files visitors send, and the text read from them, are deleted after 30 days, or sooner with their conversation. Contact details visitors leave are deleted after 3 years, or sooner with their conversation or when the customer deletes them. A ticket is deleted 2 years after it was solved, or sooner with its conversation or when the customer deletes it; one that is never solved stays until the customer deletes it. Mail to an agent stays as its conversation does. What a webhook sent is deleted after 7 days. An invitation that expired, was taken back or was declined, and a team member’s place they left or were removed from, are deleted 30 days later, each with the address it carried. The key that stops invitation mails to an address is kept for as long as we send them. The audit trail is deleted after 24 months. Rate-limit hashes are deleted after a day.

Your rights

Under the GDPR you can ask what we hold about you, have it corrected or deleted, receive a copy, and object to or restrict its processing. Visitors should ask the website they were talking to, which controls that data; we help that customer answer. Anyone can also complain to their national data-protection authority.

Changes

This notice is updated when the product changes what it stores or who processes it. The date on the footer is the last update.