Where your data goes

Is my content used to train an AI model?

No. We do not use customer content, conversation transcripts, or lead submissions to train third-party AI models. Conversations and the excerpts we retrieve from your knowledge base are sent to Groq to generate the answer in that conversation, and Groq has confirmed it does not train on API inputs.

Finding the right passage in the first place uses a second provider. Your knowledge-base text is sent to OpenAI once when a document is indexed, and the search query — the visitor’s message plus the preceding exchange, each part capped separately at 1,000, 400 and 300 characters, so about 1,700 in total — is sent when a question is asked. OpenAI turns text into numbers so we can compare meaning; it writes no part of the reply. OpenAI states that API data is not used to train its models and that inputs are kept for up to 30 days for abuse monitoring, then deleted.

Two other jobs send visitor questions to Groq as well, outside the conversation they were asked in, and we would rather name them than leave them inside “providing the service”. A nightly job groups the questions your agent answered badly or could not answer at all, within your workspace, which is what fills the suggestions on your Training page. And we sample recorded turns — question, reply, and the passages the reply came from — back through a model to measure how often the answers were actually right. Neither trains anything, and neither leaves the providers named here.

Where is it stored?

The application and database run on a single server in DigitalOcean’s London region. Your account data, your connected documents, your conversations and your lead captures are all held there.

Several of our sub-processors are headquartered in the United States and may process data there. Where personal data leaves the UK or EEA we rely on the UK International Data Transfer Addendum or the EU Standard Contractual Clauses, alongside the technical measures below.

Who else can see it?

Every company that touches your data is named in the privacy policy, with what it is used for and exactly what is shared. In short:

• Groq — generates the replies. Sees conversation messages and the excerpts we retrieve as context.
• OpenAI — finds which passage to answer from, by meaning rather than by matching words. Sees your knowledge-base text and the search query. Writes none of the reply.
• DigitalOcean — hosts the application and database, in London.
• Stripe — subscription billing. Card details go directly to Stripe; we never see or store them.
• Postmark — delivers verification, reset, invitation and receipt emails.
• Google and Microsoft — only when you connect a Drive or Microsoft 365 source, and only for the files you select.
• Google Analytics — usage analytics on our own site, and only for visitors who accept analytics cookies.
• Cloudflare — the bot check on our sign-in form, which sees the IP of the person signing in, and the DNS and content-delivery layer in front of serochat.ai, which sees request metadata as it passes through. The widget never loads the bot check onto your site.

We do not sell or rent personal data, and we do not share it with advertisers or data brokers. We will give existing customers reasonable notice before adding a sub-processor that materially affects their data.

Who is the controller?

It depends whose data it is, and the distinction matters if you are filling in a supplier questionnaire. We are the controller for data about you as an account holder — sign-up details, billing, dashboard usage. We are a processor for the visitor conversations, lead captures and knowledge-base content inside your workspace: you are the controller of that, and we process it on your instructions.

How long is it kept?

• Account data — while your account is active. Requesting deletion starts a 7-day grace period (signing back in cancels it), after which the account, its workspaces, members, conversations, documents and sessions are hard-deleted from the primary database and drop out of our backups as those rotate. Backups are retained for up to 14 days, so no copy survives more than 21 days from the request.
• Billing records — 7 years after the end of the relationship, as UK accounting and tax law requires, limited to what HMRC needs.
• Conversation transcripts — deleted automatically once a conversation has been idle for 90 days, and deletable by a workspace admin at any time before that. Captured leads outlive the transcript, with the link to it cleared.
• Server logs — up to 30 days.
• Widget event log — up to 90 days, and the setting we currently run is 30. Visitor IPs are stored only as a SHA-256 hash, never in plaintext.
• Support correspondence — up to 3 years after the last interaction.

What protects it

• TLS on all traffic to the dashboard, the widget and the API.
• Passwords stored as bcrypt hashes. Verification, password-reset and invitation tokens stored as SHA-256 hashes only, single-use, and expiring on their own clocks: 24 hours for a verification link, 30 minutes for a password reset, 7 days for a team invitation.
• Integration OAuth tokens and API keys encrypted at rest, with the application key kept outside the database.
• Parameterised queries throughout, and rate-limited authentication endpoints to slow credential stuffing.
• Least privilege on internal access: production access is restricted to named engineers and logged.

Can the agent be tricked into taking an action?

No, because it cannot take actions at all. Most AI support products now ship an actions layer — the agent can search the web, call your APIs, process a refund on a visitor’s behalf. That is real power, and it creates a specific risk: an agent with tools can be manipulated by a crafted message into using them wrongly, an attack known as prompt injection. Serochat’s agent has no tools. It composes an answer from the passages retrieved from your knowledge base, or hands the conversation to a human, and those are the only two things it can do.

The rules that keep those answers honest are enforced the same way: our server checks each reply after the model produces it — the citation must match a passage that was actually retrieved, and an answer the sources do not support is escalated to a person instead of shown — rather than relying on instructions we hope the model follows. The worst a hostile visitor can get out of your agent is an answer.

What we do not claim

We are a small UK company and we are not going to imply otherwise. Serochat does not hold SOC 2 or ISO 27001 certification, we do not publish penetration-test reports, and we do not offer a contractual uptime SLA. If your procurement process requires any of those, we would rather tell you now than at the end of a questionnaire.

No system is perfectly secure. In the event of a personal-data breach affecting you, we will notify you and, where required, the Information Commissioner’s Office without undue delay and in any case within 72 hours of becoming aware.

Reporting a vulnerability

Please email [email protected] and we will respond as quickly as we can. We will not pursue anyone who reports a genuine issue in good faith and gives us a reasonable chance to fix it.

Every statement on this page is drawn from the privacy policy, which is reviewed against the live database schema and application code. Where the two could ever disagree, the privacy policy is the binding version.