AI content disclosure

What OwnersPal generates

Short videos for ads and social posts, including AI-generated instrumental background music under the footage by default. Draft replies to reviews (4–5 star ones can auto-send within the limits of your plan; lower-rated ones always wait for your approval). Captions and copy for posts. Suggested ad audiences and budgets. None of this is the output of a real person on our team — it's machine-generated content shaped by the brand profile, preferences, and feedback you give us.

Words on your videos

Your videos can show a short line of text on each shot and a closing card with your business name, address, contact details and a call to action. An AI model writes the short lines from what your own photos show and from details you gave us or published yourself on your website or Google listing. Before a line appears, an automatic check removes any line that names a price, number, currency or claim like "best in town" that your own records don't support. The text is drawn over your real footage; your photos themselves are never changed. The AI also picks the lettering style for each video from a set of styles we designed. On one or two shots, the words can be attached to the scene so they move with the camera.

What we use as input

The information you give us when you sign up (business name, location, industry, style preferences). Photos you upload. Reviews that are already public on Google, Yelp, Facebook, etc. Leads who fill in your forms. Ad-platform tokens you've explicitly granted. We don't crawl your inbox, your personal social media, or anything you haven't authorized.

Which AI models we use

We don't train our own foundation models. OwnersPal coordinates a handful of third-party providers, each chosen for one job: • Kling, Seedance, Veo, Grok, Vidu, Wan, and MiniMax H3 — image-to-video motion for short ad clips, all reached through Fal's inference gateway. • OpenAI (GPT-class) or Anthropic (Claude models) — whichever of the two we run at the time handles captions, ad copy and other text drafts, translations, image understanding and video planning, photo quality scoring, extracting your menu and other details from your website, website-assistant answers, and, for Brand Photoshoot, reading each photo, planning each shot and checking each styled version. • OpenAI only — content-safety screening and the monthly AI-visibility check. • ElevenLabs (via Fal) — optional voice-hook synthesis using preset voices only (no voice cloning). • ElevenLabs (directly, not via Fal) — the optional phone assistant: it turns a caller's speech into text and speaks the assistant's replies in a preset voice (no voice cloning). Each reply is written by Google's Gemini language model, reached through ElevenLabs. • fal.ai's Stable Audio, CassetteAI, ace-step, ElevenLabs Music, and mmaudio (all via Fal) — AI-generated instrumental background music mixed under your video footage by default. • Bria (via Fal) — when a product photo needs a backdrop, generates an AI-created scene around your real product photo; the product itself is never altered or invented. • OpenAI GPT Image and Google Nano Banana 2 (both via Fal) — image editing for Brand Photoshoot's optional styled versions: they add generic set dressing and lighting around your real photo. Google's model embeds an invisible SynthID watermark in its images. For product photos, BiRefNet (via Fal) first cuts your real product out so its own pixels can be kept, and Black Forest Labs' FLUX Fill (via Fal) is a fallback that paints the studio set when GPT Image is unavailable. • Fal — the model-inference gateway that brokers our calls to the image-to-video models above (Kling from Kuaishou, Seedance from ByteDance, Veo from Google, Grok from xAI, plus Vidu, Wan, and MiniMax H3, an open-weight model from MiniMax), to ElevenLabs, and to the image-editing models above. The authoritative list and the data each provider receives is at /subprocessors. If we add, remove, or swap a provider, we update that page in the same release.

What we DON'T do with your data

We don't use your customers' data to train public AI models. We don't sell or rent your data. We don't share AI prompts with other OwnersPal customers. The third-party AI providers we use (listed at /subprocessors) operate under data-processing terms that prohibit them from training their public models on data passed via OwnersPal's API.

What we learn across all OwnersPal businesses

One setting is tuned across businesses: three video-pacing numbers — how long the first frame holds, how long the last frame holds, and how long transitions cross-fade. Each week we average the values used in videos that owners rated highly or accepted (a 1–5 rating, or approval, of their own video), pooled across all customers, and nudge the defaults by at most 10%. Only those three numbers leave a business's account. No prompt, caption, photo, review, customer record, business name or written feedback enters the pool, and nothing in it can be traced back to any business. Everything else OwnersPal learns is per business and is never shared.

How AI-generated content is labeled

Every video OwnersPal generates carries a visible “AI-generated” badge burned into the frame for its full duration — it travels with the file wherever it's posted or downloaded, and our pipeline treats that label as mandatory: if it can't be applied, the video isn't published. Where a platform lets a publishing app declare AI content, we now do: we set that flag on every video we post to TikTok, YouTube, and Instagram, so their apps can show their own AI label. Facebook doesn't yet offer that switch to publishing apps — it instead reads provenance metadata embedded in the file itself, and that machine-readable marking is still rolling out on our side. X offers no publisher-side AI-content flag either, so we don't set one there. For a video we post there, we append a visible “(AI-enhanced)” label to the post text instead. LinkedIn offers no publisher-side AI-content flag either, so we don't set one there. For a video we post there, we append a visible “(AI-enhanced)” label to the post text too.

How Policy Guard checks every post before it goes out

Before OwnersPal publishes a post, sends a review reply, or launches an ad, Policy Guard checks it against the actual policies of the platform it's going to — Google Business Profile, Meta, TikTok, YouTube, Google Ads, Nextdoor, LinkedIn, and X — plus OpenAI's content moderation on the text and any images involved. This is a real gate, not a suggestion: content that trips a blocking rule (an incentivized-review offer, a before/after health claim, a protected-attribute ad targeting line, and similar platform-prohibited patterns) is stopped before it reaches the platform, with the plain-language rule text and a fix shown to you. This uses OpenAI (listed at /subprocessors) for the moderation step — no new provider.

Content Credentials (C2PA)

Our video pipeline is built to carry provenance forward rather than wipe it. A shared FFmpeg chokepoint keeps container-level metadata tags — including any provenance a model writes into standard MP4/QuickTime tags — across a re-encode. It now covers the resizing, trimming, end-card, on-screen badge and background-music steps of our main video pipeline, plus the final storage remux, but not yet every step: the optional sound-effects mix is composed by a third-party audio provider outside this chokepoint, so on the rare video where sound effects are the last layer applied, we don't yet promise that embedded metadata reaches the finished file. Full C2PA Content Credentials use a richer manifest format (JUMBF) that a standard re-encode doesn't retain on its own; preserving that end-to-end — extracting the manifest and re-attaching it with a recorded "transcoded" action — is not yet available for video. What we already guarantee: we never alter or forge C2PA credentials, and every video carries the visible burned-in "AI-generated" badge.

Synthetic humans and "deepfake" content

OwnersPal will not generate content that impersonates a real, identifiable person (a named employee, a named customer, a public figure, a politician). If a video we produce features an AI-generated human likeness — an actor that doesn't exist in real life — it carries the same visible "AI-generated" badge burned into every OwnersPal video, with the platform-native AI-content tag added on top wherever that handoff is live, in line with EU AI Act Article 50(4) and New York General Business Law §396-b (effective 9 June 2026). We also will not generate sexually explicit imagery of anyone, real or synthetic, under any circumstance — this aligns with the federal TAKE IT DOWN Act (15 U.S.C. §1309) and is enforced at the prompt layer.

Your responsibilities

Don't upload photos of identifiable people without their consent. Don't ask OwnersPal to make claims about your product that you can't substantiate. Don't ask us to impersonate a real person (employee, customer, public figure) or to fabricate a testimonial. Don't run ads that violate platform policy or applicable law. If you do these things, the content is yours, not ours — but we may refuse to generate or publish it.

If someone asks, we tell them

If a viewer, a regulator, or a counterparty asks whether content in a OwnersPal-produced video, reply, or post was AI-generated, the honest answer is yes — and you are free (and encouraged) to confirm it. Some jurisdictions, including Utah, require generative-AI services to disclose AI involvement upon a clear and prompt request from a user (Utah Artificial Intelligence Policy Act, SB 149). OwnersPal's standing answer to such a request is: yes — and the production details for any specific piece of content are available to the business owner inside the OwnersPal dashboard. Contact privacy@ownerspal.app if you cannot reach the business owner directly and we will help.

Not professional advice

OwnersPal-generated content is marketing copy. It isn't legal advice, medical advice, financial advice, or any other kind of professional advice. If your business is regulated (medical, legal, financial, real estate), you remain responsible for ensuring every published piece complies with the disclosures and standards that apply to your profession.

Results aren't guaranteed

OwnersPal can make it dramatically easier to look alive online. It can't guarantee revenue, sales, calls, foot traffic, follower count, or platform reach. The figures or examples we may show in product or marketing reflect what some customers see in some situations — not what every customer will see in their situation. Your industry, market, budget, and how you use OwnersPal all matter more than the tool itself.

Your review and control

By default, every new video, every reply to a 1–3 star review, and every Google Business post waits for your approval over WhatsApp or email. Auto-approve for 4–5 star replies is off by default; you can turn it on in Reviews settings, and replies to 1–3 star reviews always wait for your approval. Either way, nothing is permanent: you can edit, delete, or replace anything OwnersPal produces, and you can revoke ad-account access at any time.

Turning AI features off

You can opt out of any AI-driven feature individually, without losing the rest of OwnersPal: • Auto-reply to reviews → Reviews → See details → turn "Auto-approve replies" off (every reply then waits for your approval). • Voice hooks in videos → Profile → Voice opening → toggle off (no audio is sent to ElevenLabs). • Review overlays in videos → Profile → Show reviews on ads → toggle off (Google reviews are not fetched for video rendering). • Phone assistant → Phone → turn it off (it no longer answers calls: Twilio rings your transfer number or plays a fixed voice message, and your saved call history is deleted). Turning off all AI features leaves you with a manual scheduler, a media library, and the analytics view. To delete your account entirely, follow the steps at /data-deletion.

Voice-hook synthesis (opt-in)

If you turn on voice hooks in profile settings, OwnersPal generates a short (5–8 second) text-to-speech audio clip that opens your video. We use ElevenLabs (via Fal) and only their pre-licensed preset voices — we never clone your voice, an employee's voice, or any other person's voice. The text sent to ElevenLabs is built from structured fields you've provided (your business name, hours, default CTA). We never send a customer review, a photo, or free-form copy to the voice provider. Voice hooks are a paid feature in higher OwnersPal tiers because they consume third-party API credits; if you don't opt in, no voice synthesis runs and ElevenLabs receives no data about your business.

Keeping the same person consistent across shots (opt-in)

If you turn this on in profile settings (available on higher OwnersPal plans) and the same person appears in more than one shot of a video, OwnersPal picks one of the photos you uploaded of that person and sends it to fal.ai's video model as a reference image for each of those shots, so the person's face and outfit look the same throughout the ad. Why: without a reference, each shot is generated separately and the same person can look slightly different from shot to shot. What we do not do: we never create, store, or compare a face scan, face template, or any other biometric identifier, and we never use a photo you did not upload. Retention: a private copy of the reference photo is saved in your OwnersPal storage when the video is made, so the video can still be finished if you delete the original photo; the link sent to fal.ai expires after one hour. Deletion: each stored copy is deleted 90 days after the last time it was used as a reference — using it again restarts those 90 days — and turning the setting off or deleting your account deletes every copy sooner. It is off by default; only turn it on if the people in your photos have agreed to appear in your ads. This setting is not currently offered to businesses located in Illinois or Texas. fal.ai is already listed at /subprocessors — no new vendor.

Review-overlay rendering (opt-in)

If you turn on review overlays in profile settings, OwnersPal can display your real Google Business Profile reviews inside the videos we generate for you. Three rules protect customer privacy: (1) only reviews with 4 or 5 stars are eligible — we never use lower-rated reviews; (2) we mask author names to first-name-plus-initial format ("Jane Doe Smith" → "Jane S.") at write time; (3) we keep the verbatim review text as the reviewer wrote it and never paraphrase. If a customer deletes their review on Google, our nightly sync removes it from the cache within 24 hours. The cached review data lives in your account only, is never shared with other OwnersPal customers, and is deleted with your account.

About rewrite in a chosen tone

You can ask OwnersPal to rewrite your public "About" blurb in a tone you pick (warm, professional, upscale, or playful). This is evidence-first: the rewrite only restates facts you already gave us — your business name, your current About text, and your key attributes — in a new voice. It does not invent awards, specialties, statistics, or any claim you didn't provide. The result is a draft suggestion that you review and approve before anything goes live; nothing is published or saved automatically. This feature uses OpenAI or Anthropic (Claude models), both listed at /subprocessors, to do the rewrite, and only the facts above are sent.

Profile text for your social pages

When you press "Write my profile text", OwnersPal drafts the short bio for your Facebook, Instagram and Threads pages. This is evidence-first: the draft is written only from facts your business gave us or confirmed — your business name, the kind of business, your town, your About text, the @name you chose and whether you take bookings. It is never drafted from details that only Google supplied and you have not confirmed, and your phone number, address, website and opening hours are not sent. Before you see it, the draft is checked: it may not make claims like "best in town", it may not contain a number, phone number, email or web address that is not in your own facts, and it is screened by OpenAI's moderation service. If the AI is unavailable or its text does not pass, you get a plain sentence made from your own facts instead. You review and edit the text before you use it, and you copy it onto your pages yourself: OwnersPal never posts it for you. This feature uses OpenAI or Anthropic (Claude models), both listed at /subprocessors, to draft the text, and only the facts above are sent.

Translated public content

When a visitor to your published website switches language, OwnersPal can show your real content — your About text, your menu item names and descriptions, your FAQ, your team bios, and the text, button labels and photo captions you add to your own sections — translated into that language. This is a faithful translation: it restates exactly what you wrote, in another language. It does not rewrite, summarize, shorten, or invent anything, and it keeps proper nouns, prices, and links unchanged. People's names are never sent for translation. The original-language content is the source of truth, and any field we have not translated yet is shown in your original wording. This feature uses OpenAI or Anthropic (Claude models), both listed at /subprocessors, to do the translation.

Creative Card Studio ad copy

In the Creative Card Studio, OwnersPal can write the headline (the "hook") for an image or animated ad card and choose the layout that fits it. This is evidence-first: the copy and design are built only from facts you already gave us — your business category, your menu or offers, your current promotions, and your real customer reviews. It does not invent awards, prices, or claims you didn't provide; a superlative like "best in town" appears only if one of your real reviews actually says it. Every card is a draft you review and edit before you post it — nothing is published automatically. The AI copywriting is a paid feature; without it, the Studio still creates cards from built-in templates and your own text. This feature uses OpenAI or Anthropic (Claude models), both listed at /subprocessors.

Google Search ad drafts

For a Google Search campaign, OwnersPal can draft the ad's headlines, descriptions and search keywords for you. This is evidence-first: the draft uses only facts you already gave us — your business name, your About text, your business category, your area, your listed services or menu items and the words you asked us never to use in ads — and it is told not to invent prices, offers, awards, ratings or claims like "best". The draft is only a suggestion: you edit it, and nothing runs until you save it and approve the campaign's spending. If you turn on "Let Google optimise" (AI Max), Google may show your ad for searches beyond your keywords, but it only uses the headlines, descriptions and landing page you approved; Google's own ad-text writing and landing-page expansion stay off. Drafting uses OpenAI or Anthropic (Claude models), both listed at /subprocessors.

Photo quality analysis

If you turn on photo quality analysis, a vision model from OpenAI or Anthropic (Claude models) reviews the business photos you upload and rates each one for marketing quality — composition, lighting, focus, and how clearly it shows your subject. You stay in control: scoring runs only when you ask for it, and low-scoring photos are moved to your trash for review only if you opt in. Trashed photos are recoverable for 7 days before they are permanently removed. OpenAI and Anthropic are both listed on our subprocessors page, and the feature is available on paid plans.

Website design suggestions

In the website editor, OwnersPal can suggest up to three complete design directions for your site — a layout template, a font pairing, a section order, a lead photo, and a short About draft, each with a one-line reason. This is evidence-first: suggestions are composed only from material you already gave us — your real photos, your menu or services, and your real customer reviews — and the lead photo is always one of your own uploaded or imported photos, never a generated image. These are suggestions only: nothing changes until you choose one, the result is a draft you can edit freely, and your website is never published without your explicit action. Reviewer names are never sent. This feature uses OpenAI or Anthropic (Claude models), both listed at /subprocessors.

On-site AI assistant

Businesses on the Presence plan and above get an AI assistant on their own website, which the owner can turn off at any time. It answers visitor questions using only that business's public website information — hours, menu or services, team, FAQ, address, phone, and whether it takes bookings or online orders. It never invents a price, availability, or any other fact that is not in that information, and it declines questions unrelated to the business. Every answer is labeled "AI assistant." Visitor questions are screened by OpenAI's moderation system and sent to OpenAI or Anthropic (Claude models), both listed subprocessors, to generate the answer. Questions and answers are not stored, unless the visitor chooses to send a question to the business owner — then it is saved like any other website inquiry.

AI-visibility report

Once a month, for businesses on the Presence plan and up, OwnersPal asks OpenAI a handful of everyday questions someone might ask an AI assistant — like "best cafe near Kirkland" or "is Sunrise Cafe good" — and checks whether your business, or named competitors, come up in the grounded answer. Only your own business's category, city, and name are sent; no customer data is involved. We show you what came back — whether you were mentioned and which other businesses were — so you can see how AI assistants currently answer people looking for a business like yours. Nothing from these answers is published anywhere for you; it is diagnostic information shown only to you. This feature uses OpenAI (already listed at /subprocessors) — no new provider.

Brand Photoshoot (AI-enhanced photos)

Brand Photoshoot starts from photos you already uploaded. The subject of every result is your real photo, and the shoot is built not to invent a dish, room, product, sign, or person.

First, a vision model from OpenAI or Anthropic (Claude models) looks at each photo the shoot considers and writes down what it sees: what the main subject is and where it sits, whether any face is in the picture, any words printed on signs, cups, or cards, and how the light and focus look. That note is saved with the photo in your account. Then a series of text-only calls to the same provider acts as an art director (the idea and mood for each photo), a stylist (the framing and where any set dressing goes), and a photographer (lens, light, and color settings). Those calls receive the text note and your business details (name, type, brand colors, words you asked us to avoid, menu item names) — not the photo itself.

Every photo gets a real-background version: cropped to the planned frame and color-adjusted (brightness, contrast, warmth, sharpness) with ordinary image processing only. No generative AI model touches it.

Some photos also get an optional styled version, which may include up to two looks planned from your business's own details. For most photos, an image-editing model — OpenAI's GPT Image, with Google's Nano Banana 2 as a fallback, both reached through fal.ai — receives your cropped photo and a text prompt that describes the scene, light, and set dressing and quotes any printed words so they stay unchanged. It may add generic set dressing around your real subject, chosen from a fixed list — either one general list (for example napkins, boards, empty plates, herbs, flowers, candles, small plants and string lights) or a list for your type of business (for example smooth pebbles, a rolled towel and a palm leaf for a salon) — and change the lighting and background around it. It is instructed never to add prepared food or dishes, branded products, signs, rooms, or people, and not to change your subject or any printed text, but an image model can still get this wrong. Images from Google's model carry an invisible SynthID watermark. For product photos (a dish, a drink, or a product), OwnersPal instead cuts your real product out with a background-removal model (BiRefNet, via fal.ai), keeps your real product's own pixels — resized and placed onto a plain studio backdrop, never redrawn by the AI — and, when there is more than one real item, may place them side by side on the same set. OpenAI's GPT Image (or Black Forest Labs' FLUX Fill as a fallback, also through fal.ai) then paints only the studio set around your real product. When your brand kit has colors, a few fabric props in the studio set (napkins, table runners, placemats, tea towels and draped fabric) may be shown in one of your brand colors or a neutral tone, never in a color close to one of your product's main colors; in photos edited in place, every prop keeps its natural color. When a list for your type of business is used, props that match something your business sells are left out, judged by its category and by English words in its details (for example flowers for a florist); this check can miss a match. No prop on any list carries lettering or a logo, but an image model can still add some.

A second AI model (a vision model from the same provider) then compares each styled version with your real photo, checking your subject, any printed words, and whether anything was added that should not be there. A version is offered to you only if it passes that check; otherwise it is discarded. The check is automated and can miss a difference, so look over each styled version before you keep it. Photos where the read detects a face, photos of people or logos, and photos the shoot could not read with confidence get the real-background version only. That read is automated too, so a small or background face can be missed.

Every styled version is labelled "AI-enhanced" in the image file's own metadata, and it also carries signed C2PA Content Credentials whenever our signing certificate is in place. The metadata label itself is not signed and can be lost if the file is re-saved or screenshotted. The picture itself also carries a visible "AI-generated scene" mark in one corner, so the label stays with the image when it is re-saved, screenshotted or reposted, unless the picture is cropped. A styled version you approve is saved to your media library, always shown with an "AI-enhanced" label, and the file you download is that same labelled file. It stays there only: it is never used for the photos your automatic videos are made from, your website, your ads, or your cards, and it is never sent to your Google Business Profile — only real-background versions go there.

You pick the results you want, photo by photo, and can mark any photo "Not for me" with a reason. Your picks and rejections tune later shoots for your business only: they are never combined with other businesses' choices, and no engagement numbers such as likes or clicks are used. OpenAI, Anthropic, and fal.ai (which runs the OpenAI, Google, and Black Forest Labs image models) are listed at /subprocessors.

Connected AI assistants (opt-in)

If you connect OwnersPal to an AI assistant you choose, such as Claude, ChatGPT or another assistant that supports the MCP standard, that assistant can read your business's information in OwnersPal within the permissions you approved. It cannot change your business information, content or settings. Any answer the assistant gives about your business, such as a summary, an explanation or a suggestion, is generated by that assistant from OwnersPal's numbers, under its provider's own terms. OwnersPal does not generate those answers and does not review them. OwnersPal adds no AI generation of its own for this feature: the assistant receives information already in your account, some of which, such as caption suggestions, was made earlier by the AI features described on this page. Check an assistant's answer against your OwnersPal dashboard before you rely on it. You can disconnect an assistant at any time in Settings › AI assistants. Removing OwnersPal inside the assistant ends the connection only if that assistant tells OwnersPal, and not every assistant does, so to be sure, disconnect it in Settings › AI assistants. Our Privacy Policy describes what a connection can and cannot share.

Phone assistant (opt-in)

If you turn on the phone assistant, an AI answers the calls made to your business's phone assistant number. It is not a person. Before anything else, every caller the assistant answers hears, in English, that they have reached an AI assistant and that the call may be recorded and shared with our service providers. Normally they hear: "Hi, you've reached the AI assistant for [your business name]. This call may be recorded and shared with our service providers." If the assistant picks up but our service cannot confirm the call, it says "Hi, you've reached an AI assistant. This call may be recorded and shared with our service providers.", tells the caller the line isn't available and ends the call. Google's Gemini, reached through ElevenLabs, may read anything the caller says before the call ends, and we keep no transcript of such a call. You can add your own words to the greeting; callers hear them right after this disclosure, and you cannot change or remove the disclosure itself. When the assistant answers a call only to take a message, because it was already on as many calls as you allow at once and no one at your transfer number took the call, the caller hears this disclosure and then "Sorry, everyone is busy right now. I can take a message and make sure the team gets it." Callers can speak English, Spanish, French, Hindi, Italian, Russian, Turkish, Chinese or Vietnamese: the assistant is instructed to detect the language and reply in it, and to begin its first reply in another of these languages by saying again, in that language, that it is an AI assistant for your business and that the call may be recorded and shared with our service providers.

What it does: it is instructed to answer questions only from the information you have published; if online booking is set up for your business, it can check open times and book a table, and it is instructed to book only after reading the details back and getting a spoken yes to your booking terms; it takes a message for you; and it can transfer the call to the transfer number you set, which it is instructed to do when the caller asks for a person. Twilio connects a transfer only to that number. Whoever answers there must press 1 to take the call (see below). If it doesn't know something, it is told to say so and offer to take a message. If a caller threatens your business, your staff or anyone, it is instructed to end the call and leave you a message marked URGENT; that message stays in your inbox, and we also email you that it is there, without the caller's words. If a caller who is booking mentions an allergy or a dietary need, it is instructed to put it in the booking notes, and we mark booking notes in which we detect one for your staff to confirm.

What it does not do: it cannot take orders or payments (if you have an online ordering link, it is instructed to read that out instead). It sends callers no text messages or emails, makes no outgoing calls of its own (it only connects a caller to your transfer number), and does not clone anyone's voice (it uses one preset synthetic voice). It is instructed never to ask for a card number, not to give medical, legal or financial advice, never to promise that a dish is free of an allergen, to tell anyone describing an emergency to hang up and call 911, and to tell a caller who mentions suicide or self-harm that they can call or text 988, the Suicide and Crisis Lifeline.

How it works: Twilio carries the call, and ElevenLabs processes the caller's voice live, turning it into text and speaking the replies. Google's Gemini language model, reached through ElevenLabs, writes each reply from the text of the call so far and your published information. Call audio is not recorded. ElevenLabs works with the conversation before anything is removed from it. We ask ElevenLabs and Twilio to delete their copies of each call once it ends: normally right after the call, and within two days if we do not receive the call's end notice. For a call made while our service was down, we ask within about an hour of our service coming back; the log of a call made more than three days before then stays with Twilio under its own retention. If that notice never reaches us and we also could not link the call to its ElevenLabs conversation when it started, we cannot tell ElevenLabs which conversation to delete, so only ElevenLabs' own 30-day limit removes it. In any case ElevenLabs keeps a conversation for no more than 30 days, and Twilio may keep records that the law requires it to keep. For each call the assistant answers, other than one our service could not confirm, we keep a written transcript and a short summary, with the caller's phone number and the call's time, length and outcome, so you can see what was asked. Before storing a transcript or summary, we automatically remove card, bank and government ID numbers and medical terms that we detect; detection can miss some, and medical terms are detected in English only. We delete this call history automatically after 90 days. A booking or message the assistant takes is delivered to you like any other booking or inbox message and is kept until you delete it or your account. It reaches you as the caller gave it, with only the card numbers we detect removed, so you can act on details such as an allergy. For each call that reaches the assistant, we also keep a usage record for billing, which holds no caller data and is deleted with your account. For a call the assistant answers, other than one our service could not confirm, that record also notes which version of the opening disclosure was used, the call's language and whether the disclosure was heard, which we check against the call's transcript before anything is removed from it; the record of any other call notes no version and does not mark the disclosure as heard. If a caller asks the assistant not to keep a record of the call, it is instructed to tell them the call won't be kept and to note the request; when such a call ends, we store no transcript or summary of it. We still keep the caller's phone number and the call's time, length and outcome in your call history; any booking or message taken on the call and the call's usage record are kept as usual; and we ask ElevenLabs and Twilio to delete their copies as for every call.

Your control: the phone assistant is off until you turn it on, and you can turn it off at any time. While it is off, or when it cannot answer a call (for example, when your included minutes and spending limit are used up, or when the assistant, or our service at the moment the call comes in, cannot be reached), calls are handled by Twilio alone: Twilio rings your transfer number if you have set one, and otherwise, or if no one answers, plays a fixed voice message saying the line is unavailable. No AI listens to these calls: neither ElevenLabs nor Google's Gemini receives them, and nothing on them is recorded. For them we keep at most the caller's phone number and the call's time, length and outcome, with no transcript. Twilio also keeps your transfer number in your phone assistant number's settings so that it can still connect calls when our service cannot be reached; we ask Twilio to remove it from there when you change or clear it or delete your account, and keep retrying until it is removed (until then, a call made while our service cannot be reached may still be connected to the old number). Whenever Twilio rings your transfer number for a call the assistant does not answer or for a call it transfers (other than a call made while our service cannot be reached), whoever picks up first hears a fixed voice message from Twilio, "Call for [your business name]. Press 1 to answer.", and the caller is connected only once they press 1 (they have about 8 seconds). If no one picks up or no one presses 1, the call is treated as not answered, so a voicemail on that number no longer picks these calls up, and the caller hears the fixed voice message saying the line is unavailable. The call to your transfer number is not recorded, and no AI hears it. One exception: if a call came in while the assistant was already on as many calls as you allow at once and no one at your transfer number takes it, the assistant answers it to take a message. On such a call it can answer questions from the information you have published and take a message, but it cannot book or transfer, and the call lasts at most about 3 minutes. It is billed like any call the assistant answers: it uses your included minutes and then counts toward your spending limit, and if not enough minutes are left, the caller hears the fixed voice message saying the line is unavailable instead. In every other way it is handled like any other call the assistant answers, as described above: the same processing by ElevenLabs and Google's Gemini, the same transcript and summary, the same usage record with its note of the opening disclosure, the same handling of a request not to keep the call, and the same requests to ElevenLabs and Twilio to delete their copies. These calls do not count toward your limit of calls at once. If the assistant cannot be reached for such a call, the caller hears the fixed voice message saying the line is unavailable, and your transfer number is not rung a second time. Turning the assistant off deletes your saved call history, you can delete any single call from your calls list, and deleting your account deletes all call data; each time, we also ask ElevenLabs and Twilio to delete their copies of the calls deleted from your call history. See our Privacy Policy and the subprocessors page for the details.