← Back to home

Sieve · privacy

Privacy Policy

Last updated · July 22, 2026

Sieve classifies incoming Gmail and Outlook messages as Solicitor, Malicious, or leaves them alone. This policy explains what data we access, how we process it, and what we store.

Data we access and what we store

To classify a message we read its content (the body and subject) and a few lightweight mailbox signals, such as whether you have written to the sender before. We determine that by searching your Sent folder for their address, which tells us only that such a message exists; we do not read the contents of your sent mail. This access is transient: incoming message content is processed in memory to produce a verdict and then discarded. We may inspect attachment metadata such as filename, type, and estimated size for security signals, but we do not store attachment files.

What we store is mostly metadata: message and thread IDs, sender domain, the label we applied, and the signal that drove the decision. We never store message bodies or snippets. For messages we flag as Solicitor or Malicious we also store the subject line, encrypted at rest, so you can review what we filed; those subjects are deleted automatically after 30 days. Subjects of mail we leave alone are never stored, and a message you explicitly correct is retained as described under Corrections below.

How we process your data

Classification is fully automated. Deterministic checks run first; only messages that do not short-circuit are sent to a vetted AI model provider under its standard API terms: your data is never used to train the provider's models, and the provider retains classification inputs and outputs for up to 30 days for safety and abuse review before deletion. No human at Snowfish reads your email. When the AI runs, it receives the extracted body, URLs, and relevant attachment metadata, emits a categorical verdict, and we discard that transient content from our own memory.

Sensitive data

Email content may incidentally contain sensitive personal information (health, financial, legal, or other protected categories). By connecting your inbox, you acknowledge and consent to the transient processing of such data solely for the purpose of classification.

This processing is fully automated and occurs transiently on our servers. Message bodies, snippets, rationales, and attachment files are not stored. Only messages that require model-based classification are transferred to a vetted AI model provider; classification inputs and outputs are not used to train the provider's models, and the provider retains them for up to 30 days for safety and abuse review before deletion.

What we store

We do not store message bodies, snippets, or contact graphs. We store the subject line only for mail we flag as Solicitor or Malicious (encrypted, deleted after 30 days) and for messages you correct. Raw sender addresses are stored, encrypted, only for messages you correct (see Corrections). Attachment metadata may be inspected transiently; attachment files are never stored.

Corrections

When you correct a classification (by moving a message back to your inbox or changing its label, category, or folder in Gmail or Outlook, or marking it in the app), we store that message's subject and the sender's email address, encrypted at rest (AES-256-GCM), so the Corrections view can show you which message you corrected and improve future decisions.

A raw sender address is stored only in this correction case. Subjects are also kept, encrypted, for mail we flag as Solicitor or Malicious, as described above. These correction records are removed when you delete your account, and are purged automatically after 90 days.

Improving Sieve

To make Sieve more accurate, your corrections adjust your own filter. When you move a message back to your inbox or relabel it, that signal refines how we classify future mail for you. We may also use aggregate, domain-level signals (which domains are commonly flagged, with no tie to the content of your mail) to improve detection across the product, and we may use and combine limited signals derived from users' activity, including activity involving other Sieve users you communicate with (and, on team plans, your teammates), to provide, personalize, and improve our filtering services. You can turn this off in your settings; on team plans your admin controls it.

We do not send the contents of your mail to any third party to improve or train a product, and we do not train machine-learning models on your data. Our AI model provider's terms likewise prohibit training on customer data and require deletion within 30 days. Any human access to stored data is gated and audit-logged, and happens only to investigate abuse or security issues, fix problems, or comply with applicable law, never to read your mail to improve the product.

If we ever introduce a feature that uses more of your mail for deeper analysis or learning, it will be a clearly labeled choice that is off by default, and we will update this policy before turning it on.

Team plans

On a team plan, your connected inbox can be linked to your organization. Two team features affect how mail in a linked inbox is classified. Both are described here and surfaced in your team settings, and neither applies to personal (non-team) inboxes.

Admin filtering rules. A team admin can create allow or block rules for a sender address or domain that apply to every inbox linked to the team. Your own allow and block decisions always take precedence over a team rule. A team block never overrides our malicious-mail detection, and a team allow never bypasses it. Rule changes are recorded in the team audit log.

Teammate trust sharing. When the team-level setting is on, the fact that a teammate has deliberately trusted a sender (for example, by adding them to their allow list) may be used as an advisory signal when classifying your mail from that same sender. No message content, contact lists, or correspondence history is shared between teammates; matching uses hashed sender identifiers, and the classifier receives only a one-sentence conclusion about the single sender being evaluated. A teammate's block decisions are never shared, the contributing teammate is never identified to you, and the signal can never override malicious-mail detection. Your admin controls the team-level setting, you can opt out for yourself at any time in settings, and invitations to a team with sharing enabled disclose it before you join.

What team admins can see. Team admins can view aggregate activity for linked inboxes (counts of classified and filtered mail), sender domains flagged as threats across the team, and the team audit log. These views never expose the contents or subject lines of your mail.

Data retention

Classification and correction metadata is automatically purged after 90 days. Encrypted subjects on flagged classifications are removed after 30 days. Deleting your account removes all stored data immediately. Disconnecting your inbox immediately revokes our access to Gmail or Outlook; records already stored are then removed on the schedule above, or immediately if you delete your account.

Third parties

We never sell your personal data. We never use Gmail, Outlook, Google, or Microsoft mailbox data for advertising or share it with advertising or analytics platforms. We may share limited website browsing data (never your message content) with analytics and advertising providers solely to measure and improve our own marketing of Sieve, as described under Cookies, analytics, and pixels below.

Google API Services User Data Policy

Sieve's use and transfer of information received from Google APIs to any other app will adhere to the Google API Services User Data Policy, including the Limited Use requirements.

We request the gmail.modify OAuth scope to apply our Solicitor and Malicious labels to incoming messages and remove the INBOXlabel (the archive action). This is the core user-facing function of the product. The same permission also searches your Sent folder to tell whether you have written to a sender before, and creates drafts you ask for, such as the heads-up note offered when a correspondent's email authentication is misconfigured. We never send mail from your mailbox. gmail.modify is the only Gmail scope the Sieve web application requests: the minimum that lets us both read a message to classify it and apply or remove a label. The Sieve Gmail add-on is a separate application with its own, read-only scopes, described under The Sieve Gmail add-onbelow. Neither application requests send, settings, or full-mailbox scopes. We use Google user data only to provide and protect the service: transient classification, Gmail label actions, 90-day operational metadata, and encrypted 30-day subjects for mail we flag. We do not use Google user data for advertising, and no human reads it except in the limited cases the Google API Services User Data Policy permits (with your consent, for security or abuse investigation, to comply with applicable law, or where needed to operate or fix the service), with access to stored data gated and audit-logged. Google user data is never used to train or improve AI or machine-learning models, whether ours or our subprocessors'. You can revoke our access at any time by disconnecting your inbox, and delete all stored Google user data immediately by deleting your account; we also purge it automatically on the retention schedule described below.

AI processor disclosure. Most incoming mail is handled by deterministic pre-filter rules (whitelist hit, known correspondent, Gmail-categorized, etc.) and never leaves our backend. When a message does notmatch any pre-filter rule, the message itself is sent to a vetted AI model provider for classification: the sender address, subject line, a truncated plaintext excerpt of the body, and derived technical signals about that message (authentication results, header shape, and whether names mentioned in the message match your correspondence history, computed on our servers; your contact list itself is never transmitted to the model provider or any other third party). This is processed under the provider's standard API terms: the data is not used to train its models, and the provider retains it for up to 30 days for safety and abuse review before deletion. Separately, URLs found in message bodies may be sent to Google Web Risk for reputation checks, and the account report emails we deliver through Resend may include the subject lines of mail we flagged, so you can review what was filtered. No other third party receives Google user data.

The Sieve Gmail add-on

The Sieve Gmail add-on is an optional side panel inside Gmail. It shows how Sieve treated the message you have open, explains the decision, and lets you correct it. It is a separate application from the Sieve web app, and it requests no access to your email of its own. Its three scopes are gmail.addons.execute, which lets the add-on run and draw its panel; userinfo.email, which we use to match the open mailbox to your Sieve account; and userinfo.profile, which supplies your first name for the panel greeting.

Almost everything the panel shows already lives in your Sieve account: the verdict, the reasons behind it, and your own filter rules. The one thing it reads live is the open message's headers, and it reads them using the Gmail access you already granted Sieve at getsieve.io — not a separate grant of its own. That is why the panel asks you to reconnect if your Sieve connection has expired.

From the open message the add-on reads seven headers and nothing else: From, Subject, Authentication-Results, List-Unsubscribe, List-Unsubscribe-Post, References, and Message-ID. It does not request or receive message bodies, snippets, or attachments. Header values are not stored. They exist for the lifetime of the request that drew the panel and are discarded with the response.

Two things are stored, and only when you tap a button yourself. When you filter or trust a sender, we store that sender's address or domain, encrypted, as your filter rule or allow-list entry. When you dismiss a sender security notice, we store that sender's domain and which authentication check failed, so we do not raise it again. Both are visible and removable in your dashboard.

Changes the panel makes to your mailbox, meaning applying or removing a label, archiving, restoring, or creating a draft, run on the gmail.modifypermission you already granted the Sieve web app when you connected your mailbox. The same permission answers whether you have written to a sender before, by searching your Sent folder for messages addressed to them in the last 12 months. That search reads no message content. The add-on's own permission covers only the message you have open and cannot change anything.

The add-on never sends email. Where it offers to notify a correspondent about a broken authentication setup, it creates a draft in your mailbox and stops; you review it and decide whether to send.

Your rights

Security

Cookies, analytics, and pixels

We use strictly necessary cookies for sign-in, session, security (CSRF), and other essential functions. We may also use analytics and marketing technologies, such as PostHog, Google Tag Manager, and Meta Pixel, on our website and app to understand how visitors use Sieve and to measure and improve our own marketing. These collect website and usage data such as pages visited, IP address, browser type, and interaction events.

These analytics and marketing technologies never receive your Gmail or other Google user data, your message content, or anything we access through Google OAuth; they apply only to your use of our website and app. Where required by law, we ask for your consent before setting non-essential cookies, and you can manage cookies through your browser settings.

Data residency and transfers

Sieve is operated from the United States, and the limited data we store is processed and stored in the United States. If you use the Service from outside the United States, you understand that your data will be transferred to and processed in the United States, where data protection laws may differ from those where you live.

Changes to this policy

We may update this Privacy Policy from time to time. For material changes we will give notice by email or through the Service before they take effect, and the “Last updated” date above always reflects the current version. Continued use of the Service after a change takes effect means you accept the updated policy.

Contact

For privacy questions or data deletion requests, email support@getsieve.io.