What your security team will ask about.
This product holds your customer conversations, your support history and your revenue numbers. Below is what we enforce and where the enforcement actually happens. Where we fall short of what an enterprise expects, that is on the page too.
Isolated at the database level.
One workspace's records cannot be read from another workspace's session. The rule lives in the database, so a mistake in application code does not open a path around it.
PostgreSQL row-level security is enabled and forced on every table that holds your data. Each carries an isolation policy tied to the workspace on the session, and no role can be granted an exception. We hold identity and credentials in a separate Keycloak realm for each workspace, with no cross-realm access.
Within a workspace, records are filtered by ownership and role: a CSM sees their own accounts and meetings. The database roles that can bypass those policies are reserved for sign-in and run on a separate connection pool.
Encrypted in transit. Stored credentials (integration keys and OAuth tokens) are encrypted with AES-256-GCM before they reach the database, and the API returns only a last-four hint, never the value.
Recording requires attested consent, follows a workspace privacy policy, and writes to an audit table. Our own recorder is the only source. The Teams and Zoom connectors contribute metadata and transcripts but cannot pull provider recording media. Transcription runs on a speech model we host inside our own AWS environment, so no third party receives the audio. For surveys, the privacy mode set at build time decides whether a respondent can be named.
Configuration exports as portable packages, translations as CSV. A vendor in this category recently shut down with twenty-six days' notice and no migration path.
Checked on the server for every request.
A hidden button is not access control. When a request arrives that the user is not entitled to make, the server rejects it, including requests that never came through our interface.
Model providers never see your data.
Your data never reaches the companies that make the AI models we run, so they cannot train on it. We never use your data to train AI models. Within your own workspace, we calibrate our scoring to your data, and that calibration never leaves your workspace. Across customers, we use only aggregated, de-identified statistics, to improve the product and to show benchmarks inside it. None of it can reasonably be used to identify a customer, an account or a person, and we do not attempt to re-identify it.
Model inference runs through Amazon Bedrock inside our own AWS environment, in United States regions. We configure Bedrock for zero data retention, so it keeps nothing once a request completes, and AWS neither shares inputs or outputs with the model’s provider nor uses them to train foundation models. We do store your requests and results, in your workspace, because that is the product.
Every model call writes an audit row: workspace, user, feature, provider, model, token counts, latency, whether a fallback was used, and which data categories were involved.
A follow-up email is drafted, then stops in front of a person. Nothing reaches your customer from the draft itself. Delivery is a separate step, separately configured, that someone triggers. Field changes heading back to your CRM queue the same way.
The assistant reads through the same role and field-level access as the person asking. Fields hidden from that role never reach the prompt.
Scores show their components and predictions name their model and confidence.
Where it runs, and who else touches it.
The platform runs on AWS in United States regions. The AI uses models we select, served through Amazon Bedrock inside our AWS environment. Everyone who touches your data is named below.
Amazon Web Services is the only company that processes your data for us. We run the rest ourselves inside our AWS environment: transcription, real-time audio and video, meeting capture, and sign-in. Anything marked tenant-selected is off until an administrator in your workspace turns it on and supplies the credentials, and we give customers at least 30 days’ notice before adding anyone to this list.
We are not certified yet.
ferent.ai is a new company and we do not hold SOC 2 or ISO 27001. Implying otherwise with careful wording would only delay the moment you find out.
An audit report is a third party confirming our controls exist. Until we have one, we will show you the controls themselves, in a live session, with your security team asking the questions.
Row-level isolation, permission enforcement, the AI audit trail and the rule that nothing goes out unless a person sends it can be shown running against a live workspace.
Send the questionnaire you use. We answer every question in writing, including the ones where the answer is no.
We will take your security team through the data model, the isolation boundary and the AI request path, and answer your engineers directly.
Mask sensitive values before processing. It reduces what you have to trust us with.
The commitments on this page go into the agreement, including the limits on what your data may train and that nothing reaches your customer unless a person sends it.
Ask where certification sits in our plans and you get a date, or an honest "not scheduled yet".
If certification is a hard gate for you today, tell us early and we will say so.
Send us your questionnaire.
Better to run your security review in week one than discover it in week six. Both routes below open our contact form. Choose "a security review or documentation pack" and it reaches the right people. To report a vulnerability, email support@ferent.ai; security.txt says the same.