Privacy policy

Effective 4 September 2026, updated 26 October 2026. Written to be read, not to be survived — this is the actual policy, in plain language on purpose. If anything is unclear, ask us.

Who we are

hiy.ai (“hiy”, “we”) operates this website and service. We are the data controller for the personal data described here. Contact: hello@hiy.ai.

What we store

If you build an agent: your account email, the material you add as sources, the profile you write, your settings, the answers you give when you set your agent up (what you’re building, its main job, where your knowledge lives, roughly who talks to it and in what language, and whether there’s a team — a question you skip is simply not stored), and the conversations your agent has.

If you talk to an agent: the messages in that conversation, and — only if you choose to leave them — the contact details you submit. If the owner has set up an action that asks for a few specific things — an order number, a plan, which environment you’re on — the values you give it are stored on the same record as those contact details: a short list the owner chose to ask for, never more than six things in one action, and only the ones you actually gave. When rules during the chat are on for a workspace, each question records the path of the page the chat began on, never the query.

Bounced and complained addresses (added 10 October 2026): when our email provider tells us an address permanently bounced or that its owner marked a message as spam, we keep that address, and the reason, so we stop sending to it. It is removed with the deletion of the account that held it only if you ask; otherwise it is kept for as long as sending to it would do harm.

Masking (added 10 October 2026): an owner can turn on masking of personal data in conversations. When they do, the kinds they chose (card, IBAN, national-ID and passport numbers, and optionally phone numbers and emails) are replaced before the conversation is saved, so the stored conversation, the Inbox, exports and the transcript sent to a webhook or Zapier hold the masked text. It applies to messages saved after it is turned on, and it is off unless the owner turns it on. The contact details a visitor submits through a form are kept as given. A message sent from Slack is also held by Slack, outside hiy.

If you store a credential for an action: a token or password you paste in so one of your actions can reach the system it calls is encrypted before it is stored, and the encrypted value is all we hold — there is no screen anywhere that shows it back to you. Beside it we keep the name you gave it, its last few characters so you can tell two apart, who added it and when, and a log of every time it was stored, replaced, revoked, read by an action, or refused — which action asked, and when. Removing it destroys the encrypted value and keeps that log.

If you sign in with Google: we receive your name, email address, and profile picture from Google, and use them only to create and identify your account. We don’t request access to anything else — no contacts, no Drive, no Gmail. Our use of information received from Google APIs adheres to the Google API Services User Data Policy, including its Limited Use requirements.

We also keep basic technical data needed to run the service, such as rate-limiting counters, and first-party product events (for example “a question was asked”, “a link in an answer was clicked”) so owners can see whether their agent is working. If you build an agent, that also includes which section of your own dashboard you opened — so the app can put the parts you use first. It records the section name and nothing else. The same goes for which of the five setup questions you answered: the question’s name and the kind of agent, never the answer itself. If you press Tell me when it’s live on a page marked Preview, we record which feature it was, your workspace’s plan and how many agents it has, so we can tell you when it ships. These events are ours alone — no third-party analytics, no advertising pixels, no cross-site tracking.

In your own browser: a copy of the conversation you’re currently having is kept in that tab so a refresh doesn’t lose it, and is discarded when you close the tab. Where a business’s chat widget shows a Messages tab, it also keeps your last 20 conversations with that business’s agent on your device (questions, answers, titles and times, including anything you typed into the chat; never a sign-in), until you press Clear in Messages or clear the site’s data. It’s not a cookie.

If you asked for a person (added 25 October 2026): where the business replies in the chat, that conversation also keeps a read-only key on your device for 7 days, so you can reopen it from Messages and see their replies. The key can only show replies, and it stops working after 7 days. The conversation’s id is kept beside it for those 7 days, so that device can also keep writing in the conversation. Clear in Messages removes both.

Cookies

Every cookie we set is first-party and does a job you asked for. There are no advertising cookies, no third-party tracking cookies, and no cross-site tracking — so there’s no cookie banner, because there is nothing to consent to beyond using the service. In full, that is two cookies:

  • Sign-in — keeps you signed in to your account. Lasts as long as your session does.
  • hiy_ref — set when you arrive through someone’s referral link, so the person who referred you is credited if you sign up. It holds only their public address and a signature — no identifier for you, and nothing that follows you to any other site. It expires after 30 days.

hiy_ref is signed so it can’t be forged, and it’s HttpOnly, so page scripts can’t read it.

What we do with it

Your material is used to answer questions as your agent, and nothing else. Your content trains your agent alone — never a shared model, and never another owner’s agent. We don’t sell data.

Conversations are visible to the owner whose agent was involved, so they can see what people ask and improve it. If you leave contact details, they go to that owner so they can reply. If you left an email and a person from the business replies in the chat while you’re away, a reply you haven’t read there within 3 minutes is emailed to you, a few at a time. A reply you’ve already read in the chat you asked from isn’t emailed; one you only see after reopening the conversation from Messages may still be. To stop these emails, ask the business to remove your details, or write to us.

An owner can invite colleagues into their organisation, and an invited colleague can read the same conversations, contact details and collected values for the agents the owner names in the invite (added 10 October 2026: an invite names at least one agent and opens only those; people invited before then keep every agent in the organisation until the owner limits them, and an invite is accepted only once the invited address has been confirmed; we email the invited address the invitation, and again each time the owner presses Resend) — the people replying to you are often a team, not one person. They cannot change the agent, publish it, or manage the account. One thing they can change: material in the organisation’s Knowledge Hub that was added there rather than on a particular agent — an invited colleague can add to that pool and take things out of it, because it belongs to the organisation and not to one person. They cannot mark material confidential, and material added on a specific agent stays that agent owner’s to change. Team Agents are the exception in the other direction: nobody, including invited colleagues, gets a per-person record of what a colleague asked one.

We process this data because it’s necessary to provide the service you asked for; where we rely on anything else (like sending you a product email), we ask first or make it easy to turn off.

Third parties we rely on

  • Model providers process your material twice: once to write the summary your agent answers from, when a source is added or changed, and again to generate the actual answer when someone asks a question. Anthropic and OpenAI are called directly. Some models — DeepSeek and Google today — are reached through Vercel AI Gateway, which routes each request only to hosting providers that hold a zero-data-retention agreement with Vercel; an open-weight model is never sent to its vendor’s own API.
  • OpenAI also turns every source into the search data that makes it findable, at the moment it’s added — this runs on OpenAI regardless of which provider is handling the summary or the answer.
  • Supabase hosts the database and authentication, in the EU (eu-west-1).
  • Vercel hosts and serves the application.
  • Cloudflare opens a page in its browser service when a page you add builds its text with JavaScript and the plain read comes back empty. It receives the page’s address and returns the page; nothing you have taught your agent is sent.
  • Resend delivers notification emails, when notifications are enabled. When an owner forwards their support address to an agent (added 12 October 2026, in early access), Resend also receives that mail for us, and sends the agent’s replies: the message, its sender and its headers. We read it to answer, keep it as a conversation like any other, and keep only the names of any attachments. It also sends a business’s chat replies to a visitor who left an email and hasn’t read them in the chat, as described above.
  • ElevenLabs carries a voice call, when a workspace in the voice pilot has turned calls on and a visitor presses Call or rings the business’s hiy phone number: it turns the caller’s speech into text and the agent’s answer into speech. On a phone call it also receives the caller’s number and the number they called; hiy keeps only a one-way code made from the caller’s number, to count calls, never the number itself, unless the caller gives it as a number to be called back, which is kept like any contact detail. hiy keeps the call’s transcript in the conversation, and ElevenLabs’ short summary of the call beside it, never its audio, and each call’s ElevenLabs agent is set to save no audio and to keep its own copy of the transcript for one day. On the website the visitor is told, before the microphone opens, that the call is with an AI, recorded and transcribed, and on every call the agent’s first words say so. ElevenLabs may process a call outside the EU.
  • Telnyx is the phone carrier for a business’s hiy phone number, in the voice pilot. When someone rings that number, Telnyx receives the caller’s number, the number they called and the call’s audio, and passes the call to ElevenLabs. To register a number, Telnyx also receives the business details the owner gives: the business’s name, company number and address, and a contact name and email; hiy keeps them with the line, and, when the owner forwards their own number, that number and its carrier. Telnyx may process a call outside the EU.
  • Meta (the WhatsApp Business Platform) delivers WhatsApp messages, when an owner connects their WhatsApp Business number to an agent (added 12 October 2026, in test). Meta sends us each message someone writes to that number, with their WhatsApp number and profile name, and sends the agent’s replies. The owner’s WhatsApp Business account is theirs, under Meta’s own terms.
  • Stripe processes payments, if you subscribe. Card details go straight to Stripe and never touch hiy — we store their customer and subscription identifiers so we know what you’re entitled to, and nothing else about your card.

Each processes data only to provide their service to us. An owner can also point one of their agent’s actions at a system of their own choosing, which is not on this list and is covered next.

Endpoints an owner chooses

An owner can give their agent an action that makes one request to a URL they name. When it runs, that request carries the name the owner gave the action, a short factual summary — written by the model — of what the person asking needs, and, where the action carries them, the facts the owner’s own site attested about that person. A lookup that only reads can also carry a short value the person typed, such as an order number, in its address.

The system it writes into is the owner’s processor, chosen by them under their own agreement with it. hiy does not select it, does not contract with it, and does not process what it holds on its own behalf: hiy makes the call the action describes and records that it happened, and it does not become a party to what happens at the other end. It is not added to hiy’s own list above.

A fact an owner’s site attested about the person asking is that owner’s data about their own user, and the same answer follows: they are its controller, hiy never asks for it and never keeps it, and it exists here only for as long as the answer it was sent for.

An owner can also give their agent a webhook: one address that hiy POSTs to the moment someone leaves contact details or a conversation is handed over. That request carries the contact details the person gave and, for a handover, the last few messages of that conversation and a short model-written summary of them. The same rule applies: the system at that address is the owner’s processor, not hiy’s.

hiy keeps what each webhook delivery sent, exactly as it was sent, for 30 days (added 11 October 2026), so the owner can see what went out and send it again; it is then deleted. Only the agent’s owner can read it, in the product, and not a colleague invited to other agents. An owner can also choose to send a webhook every message of a conversation, or a note when one is closed (off unless they tick it); each is kept in the same way. A delivery to a Zapier hook is not kept past its send beyond a record that it happened.

If an owner connects Zendesk or Intercom, the same handover can also open a ticket there. The ticket carries the last few messages of the conversation, a short model-written summary of them, a link to the conversation in the owner’s account, and the email the person gave (and, in Zendesk, their name), so the owner’s team can reply from their helpdesk, which may add the person as a contact. The same rule again: it is the owner’s processor, and hiy makes the call and records that it happened.

If you talk to an agent in Slack

If an owner connects Slack, messages sent to their agent there — in a direct message or an @mention — are processed and stored the same way as a conversation on their website: the same models answer, and the same conversation record is kept, with Slack as the channel it came in on rather than a separate policy.

One thing is held apart from that record, briefly (added 12 October 2026). So that a message is answered once and in order, and a reply is not lost if a server restarts, the text of each Slack message waits in a short-lived queue until your agent has answered it, and the agent’s reply waits in another until it is posted. Both hold the text as Slack sent it, before any masking of personal details. The message’s copy is emptied as soon as it has been answered, and the reply’s as soon as it has been sent or given up on (a reply that could not be posted within ten minutes is dropped, never posted late). Any row that is somehow left over is deleted after 30 days.

If you message a business on WhatsApp

If an owner connects their WhatsApp Business number (added 12 October 2026, in test), what you write to that number is processed and stored the same way as a conversation on their website, with WhatsApp as the channel. We keep your WhatsApp number and what you wrote, in the conversation; photos, voice notes and files are not stored. If you write stop, the business sends you nothing first again. If you ask for a person, your number and name go to the business with your question.

To have this deleted, ask the business you wrote to, or email hello@hiy.ai with the number you wrote from and the business’s name. We delete the conversations from that number and its record with that business, and confirm by email. Messages already on your phone and with Meta stay there; Meta’s own deletion is under Meta’s terms.

If you bring in an assistant’s memory

If you paste in what ChatGPT, Claude or Gemini remembers about you — see the import guide — a few things about that material are different from everything else you add, and worth stating on their own.

It can be special-category data. An assistant’s memory of you can include health details, sex life or sexual orientation, political opinions, religious beliefs, finances, or details about your relationships. Under GDPR that’s Article 9 “special category” data, and we treat it that way regardless of what your particular import contains.

The basis is your consent, not necessity. Everything else in this policy runs on necessity — it’s needed to provide the service you asked for. A memory import runs on the consent you give on the import screen instead, which names those categories in words before you can add anything. Withdraw that consent by deleting the import, and we honour it in full: the source and everything indexed from it are deleted together, not hidden behind a flag.

Who sees it. The same processors listed above see this text too, nothing additional — the model provider that writes its summary and later answers from it, and OpenAI, which generates its search data.

How long we keep it. The same as any other source — for as long as it’s on your agent. See below.

Deleting it here doesn’t touch what the assistant itself holds. Removing the import removes hiy’s copy in full. ChatGPT, Claude or Gemini keeps whatever it already remembered about you until you clear it from that assistant’s own settings — we have no way to reach it, and deleting our copy doesn’t ask it to.

How long we keep it

For as long as your account exists, unless the workspace owner sets a window (added 10 October 2026): under Data & privacy they can have conversations, and the contact details left in them, deleted once nothing new has happened for 30 days, 90 days, 6 months, 1 year or 2 years. It is off until they turn it on, runs nightly, and deletion is permanent. Conversations with an open handover, one a person is holding, and any follow-up still waiting are kept until they are closed. Delete an agent, a person, or your account, and the corresponding data goes with it — deletion is real deletion, not a hidden flag. Backups age out on the infrastructure providers’ standard schedules.

The credential log is the one thing that outlives what it describes. Removing a credential destroys the encrypted value and keeps its log, because a record of use that vanished along with the credential would not be a record. Each entry says what happened — stored, replaced, revoked, read by an action, or refused — which action asked, when, and, for something you did yourself, who did it. It never contains the credential value, and never anything from a conversation. Nothing expires it on a schedule of its own: it is kept for as long as the organisation it belongs to exists, and is deleted together with that organisation — for a personal account, that is the moment you delete your account.

What a webhook was sent is kept for 30 days and then deleted (added 11 October 2026), as described under “Endpoints an owner chooses”. It holds the same contact details and message text the webhook carried, so it follows the masking and deletion rules above: masked text is stored masked, and deleting a conversation, an agent or the account removes it.

A record of each lookup run is kept for 30 days and then deleted (added 11 October 2026). It says which lookup ran, when, whether it worked, how long it took, which conversation it belonged to, and the names of the values sent and the fields kept. It never holds a value a person typed, a key, or what the endpoint replied.

Fields an owner adds about a person (Contacts, for example a plan or a region) are stored with that person’s details and removed with them.

Your controls

  • Export everything from Settings → Data & privacy, any time. Where an answer was rearranged into steps, a comparison, or a list, the export includes that machine layout alongside the text — it is a view onto the same words, not extra data.
  • Delete any agent, which removes its content, its index, its conversations, and anyone who left contact details with it. Material it shared with your other agents stays in your Knowledge Hub.
  • Release an address you are holding from a deleted agent, or let it lapse on its own after 90 days.
  • Delete your account, which removes everything — the credential log described above included. It asks you to sign in again first when you haven’t signed in within the last ten minutes (added 10 October 2026), so a browser someone left open can’t erase it.
  • Delete an individual person from your People list or from Contacts. Their conversations and the fields you added about them go too, and Export everything includes them until you do.
  • Turn off conversation-driven features. Follow-up requests — where your agent offers to take a visitor’s email — are on for a new agent and off in one tap; an agent that was already live is never switched on without you choosing it.
  • Change what your agent emails you. A new agent starts with the new-people alert and the weekly summary switched on, and the unanswered-questions alert switched off. All three live in that agent’s settings, and each can be turned off entirely.
  • Account mail is not switchable, because it is about your account rather than your agent’s activity: a decision on your agent’s identity review, a warning before a trial ends, the receipt when you delete your account, and a password reset you asked for.

If you’re in the EU/UK you also have the formal versions of these rights (access, rectification, erasure, portability, objection) — email us and we’ll honour them without ceremony.

Children

hiy isn’t intended for people under 16.

Changes

If this policy changes in a way that matters, we’ll say so plainly on this page and update the effective date — not bury it.

4 September 2026. This update adds description, not new practice. Since August the product gained actions that collect a few values alongside contact details, a credential store with its log, and actions that call an endpoint a creator names — each is now described above. How long the credential log is kept is stated in “How long we keep it”. And “Your controls” now says which of your twin’s emails start switched on, and that account mail is sent whether or not those are.

19 September 2026. A creator can now give their twin a webhook, described above under “Endpoints an owner chooses” alongside the actions that already worked this way.

8 October 2026. A creator can now connect Zendesk or Intercom to open a ticket on handover, and connect Slack so their agent answers there, described under “Endpoints an owner chooses” and “If you talk to an agent in Slack”. Pressing Tell me when it’s live on a page marked Preview now records which feature it was, your workspace’s plan and how many agents it has, described under “What we store”.

10 October 2026. Owners can now add short notes (“fields”) about a person who already left their details, and merge two entries for the same person, in a Contacts page. The notes are stored with that person, removed with them, and included in the export. Nobody is added by hand, and the notes are never read by your agent.

11 October 2026. A creator can now set up an action that reads from their own system in the product, and each run of one is recorded for 30 days. A lookup that only reads can carry a short value the person typed, such as an order number, in its address, described under “Endpoints an owner chooses”. The record of each run holds names and outcomes only, never a value, and is described under “How long we keep it”.

11 October 2026. A creator can now see what each webhook was sent over the last 30 days, with whether it arrived, and send a delivery again, and can choose to receive every message or a note when a conversation is closed. What was sent is kept for 30 days and readable only by the agent’s owner, described under “Endpoints an owner chooses” and “How long we keep it”.

12 October 2026. Messages to a twin in Slack are now recorded in a short-lived queue before they are answered, and the reply in another until it is posted, so a message is neither answered twice nor lost when a server restarts. Each holds the text as Slack sent it, before masking; it is emptied once the message is answered or the reply is sent, and anything left is deleted after 30 days. Described under “If you talk to an agent in Slack”.

12 October 2026. In test, a creator can connect their WhatsApp Business number so an agent answers it. Meta delivers those messages, described under “Third parties we rely on”, and what is kept and how to have it deleted is under “If you message a business on WhatsApp”.

12 October 2026. In early access, a creator can forward their support address to an agent so it answers by email. Resend receives that mail for us and sends the replies, described under “Third parties we rely on”. The mail is kept as a conversation like any other, under the same masking, retention and deletion rules; attachments are not stored, only their names.

17 October 2026. ElevenLabs, which carries voice calls for workspaces in the voice pilot, may process a call outside the EU, and on a call the agent’s first words say it is an AI, described under “Third parties we rely on”.

19 October 2026. The wording now says “agent” for any agent and “owner” for the person who runs one, where it said “twin” and “creator”. Expert Twin keeps its name. Nothing it describes has changed: no practice, retention period, processor or right is different.

20 October 2026. hiy now keeps ElevenLabs’ short summary of each voice call with its conversation, masked like the rest of it and deleted with it, described under “Third parties we rely on”.

23 October 2026. A helpdesk ticket opened on a handover now carries the email the person gave (and, in Zendesk, their name) and a link to the conversation, so the owner’s team can reply from Zendesk or Intercom. Described under “Endpoints an owner chooses”.

25 October 2026. A business’s chat replies that a visitor hasn’t read within 3 minutes are now emailed to the address they left, through Resend. A conversation handed to a person keeps a read-only key on the visitor’s device for 7 days so its replies can be reopened from Messages. This page now also says that Messages keeps up to 20 conversations on the device. Described under “What we store” and “What we do with it”.

26 October 2026. In the voice pilot, a business can have its agent answer a phone number. Telnyx, the phone carrier, is now named under “Third parties we rely on”, and the ElevenLabs entry now covers phone calls: both receive the caller’s number, and hiy keeps only a one-way code made from it, unless the caller gives it to be called back. To register a number, hiy keeps the business details the owner gives (and, when they forward their own number, that number and its carrier) and passes them to Telnyx.