Privacy Policy

Version 1.2 · 2026-11-17

1. Who we are

Attify is operated by Dor Tagger, trading as Attify ("Attify", "we", "us"). Attify independently audits the outcomes that AI agents claim, under published standards: Standard #1 measures customer-support resolution, Standard #2 meetings booked by AI sales agents, and Standard #3 appointments booked by AI receptionists.

Questions about this policy or your data: privacy@attify.app.

This policy covers the Attify website and web app, the Attify API, the MCP server and the emails we send. It explains what we collect, why, where it goes, how long we keep it and what you can do about it.

2. Two kinds of data, two roles

Attify handles two different kinds of personal data, and our role differs for each.

  • Account and service data (your name, email, organization, billing status, logs). We decide why and how it is used, so we are its controller. This policy describes it.
  • Customer data (the conversations, meetings and bookings in an export you upload or a system you connect). The organization that owns that system is the controller. We process the data only on that organization's instructions, under our Data Processing Agreement. If you wrote to a company's support team, booked with a company, or were invited to a meeting by one, and want to know what happens to your data, contact that company first; we will help it answer you.

Witnessed data is conversation data that reaches us another way: a vendor streams its AI agent's interactions with a customer's end users to us as they happen (section 3.7). The vendor sends it. We process it on the vendor's instructions for the vendor's private report and, once the customer consents in Attify, also on the customer's instructions for verification, under the DPA (section 2.4 there). If you talked to a company's AI agent, the same applies: contact that company first.

3. What we collect

3.1 When you sign in

Sign-in is handled by Clerk. We receive and store:

  • your name and email address;
  • the organization you act for, its name and kind (vendor, company, investor), and your role in it (owner or member);
  • when you were last seen.

3.2 When you upload an export (a check)

You can upload a help-desk export (CSV or JSON) for a free check, without an account. We store:

  • the file, temporarily (see section 6);
  • the column mapping, the results and the PDF reports computed from it;
  • your email address, if you ask for the full audit: we email you a one-time link to confirm it.

An uploaded result is labeled self-reported. Whoever uploads a file must have the right to share it with us (see the Terms).

3.3 When an organization connects a system

A help desk (today Intercom, Zendesk, HubSpot, Salesforce Service Cloud, Gorgias or Freshdesk) can be connected with read-only access, through OAuth or an access token or API key: by its owner for a connected audit of its own, or, in a vendor's verification program, by the customer organization's owner after agreeing to the scope. We then read:

  • every conversation of the last 90 days, and each day's new or updated ones after that;
  • for each conversation, only the fields the standard needs: when it started and was updated, the channel, the help desk's own customer and conversation ids, each message's text, time and whether the customer, the AI agent or a person wrote it, reopens, the AI agent's own resolution flag, and the customer's answer where the help desk asks whether the issue was resolved, or a satisfaction rating;
  • the email addresses of the help desk's admins, as the help desk lists them, and when the help desk was created where it says. In a verification program they show the customer is a company separate from the vendor, and we may email one of these admins once to ask whether the company uses the vendor's product (the answer, and when, is kept with the program).

A CRM, calendar or booking tool (HubSpot, Salesforce, Google Calendar, Microsoft 365, Calendly or Cal.com) can be connected the same way, for the meetings and bookings standards. For each meeting or appointment we read: when it was booked and for when, its status, who cancelled or rescheduled it and when, who or what created it (a person, an invitee through a link, or an integration), the organizer's and attendees' email addresses and their responses, attendance where the platform records it (a video call's participants and minutes present, an outcome or a no-show mark), what was booked (service, duration, price, provider, location, party size, title) and its ids in other systems. From a CRM we also read the contacts, companies and deals linked to those meetings: email addresses, company domains, lifecycle stage, when they became customers, and a deal's status. Section 3.9 says what we do with Google Calendar data.

An agent platform or observability account (Vapi, Retell AI, ElevenLabs, Bland, Voiceflow, LangSmith, Langfuse) is connected the same way, with a key its owner makes: by its owner for an audit of its own, by a customer in a vendor's program, or by a vendor for a deployment (witnessed data). We read each finished call, conversation or traced session in the same fields, plus the tools the agent called and what they answered, and the people of the account where the platform lists them, to tell whose account it is. The end user's phone number or id is kept only as a keyed digest, never as itself. When a vendor connects its own account, we process what we read on the vendor's instructions, as for witnessed data.

The connector credentials are encrypted before they are stored (section 8).

3.4 Vendor verification and traction reports

  • Consent records. When a customer agrees to a vendor's verification program, we record who agreed (name, email, organization), when, the exact words of the scope shown, and the standard and rules versions.
  • Invitations. The vendor gives us the customer's name (as the vendor calls it) and an invitation email address.
  • Investor requests. An investor gives us a company name, a founder's email address, optional claims and an optional note.
  • Share links and the view log. When a company shares a traction report, we record each opening of the link (page, PDF or JSON), with the time and, if the viewer is signed in, the viewer's email and organization. An optional allowlist holds the email addresses allowed to open the link.

3.5 Buyers' dashboards and the Attify Index

A company that pays for an AI agent can see its agents' audited results on a monthly dashboard. It can choose to contribute its anonymous audited outcomes to the Attify Index, which is what keeps its free plan free. We record that consent (who gave it, when, and the scope version) and its withdrawal. The Index pools contributing companies' outcomes by job type and month. It names no company, vendor or agent, and a job type's edition publishes only when at least 5 companies contributed and no company holds more than half of its claims. Withdrawing stops future contributions; an edition already published stays as published, since it identifies no one.

3.6 Billing

Payments are handled by Polar, the merchant of record. Polar collects your payment details; we never see or store card numbers. We store Polar's customer id for your organization, your subscriptions and orders (plan, status, dates) and the webhook events Polar sends us.

3.7 When a vendor streams witnessed data

A vendor whose AI agent runs on its own platform can stream each interaction the agent has at one of its customers to us as it happens, through our Ingest API, an SDK or OpenTelemetry. Before it can, the vendor attests that it has the right to share that customer's interaction data with us. We then receive and store:

  • for each interaction: its messages and their times, and whether the end user, the AI agent, a person on the customer's side, the system or a tool wrote each one; the channel; the end user's id and the interaction's reference in the customer's own systems, when the vendor sends them; handoffs; the outcome the vendor claims; and feedback (a score, with optional text);
  • for each event: its id, type, when it happened and when it arrived, its sequence number, a digest of its content and how the request was authenticated. Events hold no text;
  • for the deployment: its name, what it reports running (agent, version, model), its keys (the secret stored sealed), and the vendor's attestation (who attested, when, the terms version and the words attested).

Message and feedback text is redacted when it arrives (section 5 lists what redaction replaces). The text of system and tool messages is not kept. When the customer consents, its consent is recorded as section 3.4 says. When it reconciles a month, we store the total it states and where that figure came from, and, for each sampled interaction, whether it matches the customer's own systems and whether the customer or its connected help desk answered.

3.8 Other data

  • Waitlist. If you ask to join the verification program: company, role, email, job type and an optional note.
  • Spend ledger. For each full audit, the email that confirmed it, the check id, the model, the number of judge calls and the cost, so daily spending caps hold.
  • Email ledger. A keyed hash of each mailbox a sign-in link was sent to, to cap emails per address. Rows older than two days are deleted.
  • IP addresses. Used to rate-limit public endpoints. Counts are held in memory and are not written to the database.
  • Error reports. When enabled, errors are reported to Sentry. They carry the error's type and where in the code it happened, never message text. Not enabled at the time of writing.
  • Audit log. An append-only log records consent, publication, revocation, sharing, billing and plan changes, with who did it and when.
  • Mail you send us. Mail to our addresses at attify.app reaches us through a forwarding service (section 7) and is kept in our own inbox for as long as we need it to answer you.

3.9 Google user data

This section covers the data Attify receives from Google APIs, which today means Google Calendar.

What we access. When an organization connects Google Calendar, Attify asks Google for read-only access to who the Google account is (openid and email) and to the events on calendars the account owns (calendar.events.owned.readonly). We read only the account's primary calendar. We never create, change or delete an event, and we ask for no other Google data: no Gmail, Drive, Contacts or Google Meet.

What we read and keep. The account's email address and the primary calendar's name. For each event of the last 90 days and every later one, then each new or changed event: its ids; when it was created and last updated; its start and end, and the original start of a moved occurrence of a recurring event; its status (confirmed, tentative or cancelled); whether a person or an app created it, and the creator's email address; the organizer's and attendees' email addresses and responses (meeting rooms are left out); its title; and its location. Birthdays, focus time, out-of-office and working-location entries are skipped. We do not keep an event's description, attachments or video-call details.

How we use it. Only to provide the audit the organization asked for, in features it sees in Attify. We compare the meetings and appointments an AI agent claims to have booked with the calendar's own record and test them under Standard #2 (meetings booked by AI sales agents) or Standard #3 (appointments booked by AI receptionists): whether each one exists, who created it, and whether it was cancelled or moved. The account's email address shows which company the calendar belongs to. In a vendor's verification program, it is how we check that the customer is a company separate from the vendor, and we may email that address once to ask whether the company uses the vendor's product.

The LLM judge. For a sample of at most 200 meetings or appointments per audit, the judge (Anthropic's Claude models, through Anthropic's API; section 5) reads each sampled event's booking time, start, title and, for an appointment, its location, beside the conversation that booked it. Links, email addresses, phone numbers and order numbers in them are replaced with placeholders first, as in the conversation. Anthropic does not use it to train its models.

Who sees it.

  • People in the organization that connected the calendar see its audit.
  • In a vendor's verification program the customer agreed to, the vendor sees the customer's results (rates and counts), never its events. Published scores pool every customer and name none. Attify's reviewers may read the judge's redacted sample, exactly as the judge read it, to check the judge.
  • A company that contributes to the Attify Index adds its audited outcomes (counts, never events) to figures that name no one (section 3.5).
  • The subprocessors that run the service (section 7): Render and Supabase store it in Frankfurt, Vercel serves the web app that shows it to you, and Anthropic receives the judge's sample.
  • Otherwise, Attify's people read Google user data only with the organization's permission (for example, to answer its support request), when needed for security, or to comply with the law.

We do not sell Google user data, use it for advertising, transfer it to data brokers or other information resellers, or use it to determine creditworthiness or for lending.

How we protect it. It travels over TLS. Google's access and refresh tokens are encrypted at rest (section 8) and used only by Attify's API, to read the calendar. The data is stored in the EU (Frankfurt), access is role-based, and the audit log records sensitive actions.

How long we keep it, and how to delete it.

  • Event titles are deleted 30 days after the event's last update.
  • The rest of each event (times, status, email addresses, location and ids) is kept while the connection exists, because the checks read it across periods: a meeting cancelled or moved later, or a person booked again.
  • "Disconnect and delete everything", on the connection's page in Attify, deletes everything read from the calendar, with the scores computed from it, and revokes Attify's access at Google. Deleting the organization does the same.
  • Removing Attify in your Google Account (myaccount.google.com/permissions) stops our access at once. To delete what we already read, disconnect in Attify or write to privacy@attify.app.
  • A published score or Index edition stays as published: it pools companies, names no one and holds no event.

Limited Use. Attify's use and transfer to any other app of information received from Google APIs will adhere to the Google API Services User Data Policy, including the Limited Use requirements, and to the Google Workspace API User Data and Developer Policy. We do not use Google user data to develop, improve or train generalized artificial intelligence or machine learning models.

4. Why we use it

PurposeDataLegal basis
Run your account and sign you inSection 3.1Contract
Run a check, a connected audit or a buyer's dashboard you asked forSections 3.2, 3.3, 3.5Contract with the controller; we act as processor
Run a verification program and build traction reportsSections 3.3, 3.4Contract; the customer's recorded consent to the scope
Build a vendor's private report on its deployment and, once the customer consents, verify itSection 3.7Contract with the vendor; the customer's recorded consent to the scope; we act as processor (section 2)
Publish the Attify IndexSection 3.5The contributing company's recorded consent; the Index identifies no one
Send the emails the service needs (links, invitations, requests, shared reports, monthly verdicts)Email addressesContract
Charge for paid plansSection 3.6Contract; legal obligations (handled by Polar)
Keep spending, abuse and email limitsSection 3.8Legitimate interest in keeping the service safe and affordable
Keep an audit trail of what was agreed and doneConsent records, audit logLegitimate interest; legal claims
Fix errorsError reportsLegitimate interest

We do not sell personal data. We do not use it for advertising. We do not train AI models on it.

5. The LLM judge

Part of every judged audit is a check by an LLM judge (Anthropic's Claude models, through Anthropic's API), using the prompt published in the standard.

  • Only a sample is sent: at most 200 records per subject per audit, plus small samples for contact reasons.
  • Every transcript is redacted before it leaves our systems: links, email addresses, @handles, order, account and ticket numbers, phone numbers, long digit runs and first names in direct address are replaced with placeholders.
  • Redaction is automatic and pattern-based. It removes common identifiers; it cannot guarantee that no personal detail remains in free text.
  • The judge's answers are cached by us to avoid paying twice. The cache of a connection is deleted when its text is purged; the cache of a check is deleted with its upload.
  • Anthropic receives the sample under its commercial API terms: it does not use the data to train its models, and it keeps it only for the limited period those terms set, unless the law requires otherwise.

Human review

To check the judge, Attify reviewers may read a redacted sample, exactly as the judge read it. That sample's text is deleted after the same retention period as connected text.

6. How long we keep it

DataKept
An uploaded fileDeleted within 30 minutes of upload, or when the full audit finishes
A check's results and PDFs30 days, then the check is deleted
A full-audit email linkExpires after 20 minutes; deleted with its check
Message text and titles from a connected system30 days after the record's last update, then deleted
Records after the text is deletedTimestamps, ids, who replied and the AI flag stay while the connection exists, as contacts for the 7-day repeat-contact check
Scores and per-record outcomesKept while the connection exists. An outcome holds: eligible, claimed, audited, start time, standard version, and the system's own record id; no text and no customer id
Everything read from a connectionDeleted when the connection is disconnected, the admin address a direct confirmation was sent to included (the answer stays with the program)
Message and feedback text in witnessed dataRedacted when it arrives; deleted at most 30 days after it arrives, whether or not the customer consented
Witnessed records after the text is deletedInteraction and event records without text (ids, times, sequence numbers, a digest of each event, how the request was authenticated) and the scores computed from the text, while the deployment exists
Everything a deployment sentDeleted when the deployment is deleted, or with the vendor's organization. The vendor's attestation stays in the audit log
Published scores, attestations, Index editions and the public changelogKept as the public record of what was published
Traction reportsKept until the company's organization is deleted; share links expire (30 days by default, at most 90)
Investor request linksExpire after 30 days
Consent records and the audit logKept after deletion, as evidence of what was agreed and done. They hold no message text
Spend ledgerKept as the accounting record of judge spending; it holds an email address, a check id and costs, no message text
Account dataUntil the organization or user is deleted

7. Who we share it with

We share personal data only with the subprocessors that run the service, each for its own purpose. The current list, with what each receives and where, is in Annex III of the DPA. In short: Anthropic (the judge), Render (the API, EU, Frankfurt), Supabase (the database, EU, Frankfurt), Vercel (the web app), Clerk (sign-in), Polar (payments), Resend (email), ImprovMX (forwarding mail sent to our addresses), and, only when enabled, Sentry (error monitoring) and PostHog (cookieless page analytics). Witnessed data adds no subprocessor: it is stored on the same hosting, and only redacted samples of it reach the judge, as for connected data.

Within the service:

  • A vendor sees each customer's results (rates and counts) and consent records, never its conversations, meetings or bookings. Its reports quote no record.
  • Public surfaces (the registry, vendor profiles, badges, the Attify Index, the public files) show pooled results only, with no customer or brand named.
  • For witnessed data, the vendor already holds what it sent. It sees its private report and, once a customer consents, that customer's results and reconciliation outcome. The customer that consented sees the redacted text of the interactions drawn for its monthly check.
  • Investors see a traction report only when the company shares it. Customers appear as "Customer 1" to "Customer N" unless a customer opts in to being named; only the customer can opt in.

We may disclose data when the law requires it, and we will tell the affected organization unless we are not allowed to.

8. Security

  • Traffic is encrypted in transit (TLS).
  • Connector credentials are encrypted at rest with envelope encryption (AES-256-GCM, a fresh data key per secret, wrapped by a key that never touches the database).
  • Links we send (share links, invitations, requests, check links) are stored only as hashes.
  • Witnessed data is accepted only with a key we issue for that deployment, stored sealed and revocable. We timestamp every event when it arrives and sign a receipt of every batch (what arrived and when) with our Ed25519 attestation key.
  • Access is role-based: owners, members, and Attify's own reviewers and admins.
  • The audit log records sensitive actions (section 3.8).

The Security page explains each measure. No system is perfectly secure. If a breach affects your data, we will tell you and, where required, the authorities.

9. International transfers

Our API and database run in the EU (Frankfurt). Attify is operated from a country that the European Commission, the United Kingdom and Switzerland recognize as giving adequate protection to personal data. Some subprocessors process data outside the EEA, including in the United States. For those transfers we rely on the EU-US Data Privacy Framework where the subprocessor is certified under it, and otherwise on the EU Standard Contractual Clauses, with the UK Addendum and the Swiss amendments where they apply.

10. Your rights

Depending on where you live, you may have the right to access, correct, delete, restrict or object to the use of your personal data, to receive it in a portable format, and to withdraw consent.

  • Organization owners can export their organization's data as JSON and delete their organization from the app. Deletion removes the organization's data; consent records and audit-log entries stay as the evidence of what was agreed, and hold no message text.
  • A connected customer can disconnect at any time. We revoke our access at the platform where its API allows it (otherwise we show where to revoke it) and delete everything we read from that connection.
  • A customer that consented to a vendor's witnessed data can withdraw at any time. The vendor's interactions then count only toward the vendor's private report, and their text is deleted on the usual schedule.
  • A company contributing to the Attify Index can withdraw at any time from its dashboard.
  • People whose data is in a company's system, or who talked to a company's AI agent, should contact that company. We will support it (and, for witnessed data, its vendor) in answering.
  • Anyone else: write to privacy@attify.app. We answer within one month.

You can also complain to the data protection authority where you live or work. Attify has no establishment in the EU, so no EU authority leads on it.

11. Cookies

The web app sets only the cookies it needs to work:

  • an access cookie, set only for Attify's own preview of draft pages;
  • Clerk's session cookies, when you sign in.

We use no analytics or advertising cookies.

We measure the site without cookies and store nothing in your browser for it:

  • Page analytics, once switched on: page views and clicks go to PostHog (EU) in its cookieless mode, which tells visits apart by a hash it computes from your IP address and browser with a salt that changes every day. Addresses, ids and anything you type are never sent.
  • Tests of the marketing pages: while we compare versions of a page, we count which version each visitor saw and whether they went on to act (a click, a sign-up, a completed check). A visitor is a keyed hash of the IP address and browser, kept for that day only; afterwards only the counts remain.

12. Children

Attify is a business service. It is not meant for children, and we do not knowingly collect their data.

13. Changes

We will post changes here with a new version number and date. If a change materially affects how we use personal data, we will tell organization owners by email at least 30 days before it applies.