Security and trust

It can send your email. It cannot send it without you.

Q: Are you certified to a security standard?

SOC 2 compliance is part of the Enterprise plan , with the report available under NDA. For Google access we are completing CASA, the independent security assessment Google requires of any app that reads Gmail, and Microsoft publisher verification is under way. We do not claim ISO 27001. What every plan gets today is on this page: sign-in through Microsoft's, Google's and your CRM's own pages with no passwords stored, a confirmation step the AI cannot bypass, an audit log of every action, encryption in transit and at rest, deletion on a published schedule, a data processing agreement and a named list of sub-processors.

This page explains how the assistant is built so that a misheard word costs a re-ask and never a wrong send, who can get to your data and who cannot, what we keep and for how long, why it only calls you because you asked it to, and which independent checks we have and have not yet passed. If something here is unclear, send us a message and a person will answer.

Call with CallMyDesk Call length 00:58 On your phone

You said: “Send Tom the revised quote at eighteen-five and say we can start on the ninth.”
CallMyDesk said: “Here is the email to Tom Baxter at Lyndale Roofing. ‘Hi Tom, the revised quote is $18,500 and we can start on November 9. Best, Geoff.’ Shall I send it?”
You said: “No. Eighteen-two.”
CallMyDesk said: “Changed to $18,200. Everything else the same. Sending to Tom Baxter. Say yes, or press 1.”
You said: “Yes.”
CallMyDesk said: “Sent at 2:12 pm. It's in your audit log.”

In short

Nothing is sent, booked or logged without your spoken yes.
Every action is written to your audit log. Revoke access in one click.
Your data is never used to train models, ours or anyone's.
Call audio is deleted after 30 days and transcripts after 90, or sooner if you ask.
US-hosted. DPA and sub-processor list published.

The confirmation rule

It reads back. You say yes. Only then can it act.

Reading is free. Writing is confirmed. The assistant can read your inbox, calendar and CRM as soon as you have given your PIN. Anything that leaves your account or changes it (an email, a calendar invite, a CRM note, a deal update, a message to someone else) is read back to you first: who it is going to, what it says, and every name, number and amount in it. Then it asks. You say yes, or press 1 on the keypad. Anything else, including silence, means no.

Before the first step: PIN, and a one-time code for high-risk actions.
You ask.
It drafts.
It reads back the recipient, the gist and any numbers.
You say yes, or press 1.
Done: it acts and writes the audit-log entry.
Not right? It re-asks. Nothing is sent.

A mis-heard word costs a re-ask, never a wrong send.

That is the rule. Here is how it is built so that the rule cannot be bent, by the assistant or by anyone else.

1. Two steps, not one.

Preparing a draft and sending it are two separate operations inside the system. The first produces a draft and the words that will be read back to you. The second needs a confirmation that the first cannot produce.
2. The confirmation comes from the system, never from the AI.

The confirmation is issued by a fixed, non-AI part of the service, and only after two things have happened: the full read-back has actually been played to you (the system tracks what audio was heard, not just what was generated), and within 30 seconds of it finishing you said one of a short list of yes-words ("yes", "send it", "confirm") or pressed 1. Your yes or no is matched against that fixed list and the keypad, not interpreted by the AI. "No", "stop", "change", a new subject, silence or hanging up all cancel it.
3. The AI cannot forge one.

If the AI model asks the system to send without a valid confirmation, the system refuses and tells the model to read back and ask. Each confirmation is single-use, tied to that one draft and that one call, and expires after 60 seconds. There is no setting, prompt or phrasing that turns this off.
4. Some things have no button at all.

Payments, moving money, changing passwords and deleting mail are not things the assistant can do. There is nothing to confirm because the capability does not exist.

Names, numbers and recipients are always read back.

A misheard "eighteen-five" is corrected in the read-back, as above. A misheard "yes" is not possible in the way people worry about: a sound that is not on the yes-list is treated as no, and the worst case is that it asks you again.

Your PIN

Caller ID can be faked. Your PIN cannot be guessed.

Every call starts with your PIN: four to six digits, keyed in on the phone or spoken. We do this because the number a call appears to come from can be faked, and the checks that US phone networks now run on caller ID do not catch every case. Your phone number is a first check; your PIN is the one that proves it is you. A few wrong attempts lock the line and we email you.

For a small number of sensitive actions the assistant asks for a second proof before it goes ahead. Either it texts a one-time code to the cell phone you registered when you set up, and you read it back, or it hangs up and calls you back on that number. Someone pretending to be you from a faked number never receives that text or that call. (You agreed to those texts and call-backs at signup, when your number was confirmed by a one-time code; see "It says it's an AI" below.)

By default, the second proof is asked for when sending to someone it has not written to before on your behalf, when a figure is above a limit you set, and when changing any of your own settings, including the PIN and the number it calls you on. You can make this stricter on your account page. You cannot make it apply to nothing.

First check
Your number
Proves it is you
Your PIN
For sensitive actions
One-time code or call-back

Lockout after repeated wrong attempts; we email you when it happens.

A call is a session. Once your PIN is accepted, the assistant will read and act for the length of that call. Hang up and the next call starts again with the PIN.

What we do not do: we do not identify you by your voice. Voice matching is unreliable over a phone line and can be fooled by cloned voices, so we do not use it or claim it.

Connected accounts

We never see your password.

When you connect Microsoft 365 or Google Workspace, or HubSpot, Pipedrive or Salesforce, you sign in on Microsoft's, Google's or your CRM's own page and check what the assistant may do. That is the standard way one service is allowed to act in another, and it means your password is never typed into our site, never sent to us and never stored by us.

We ask for the least we can. A new account starts with permission to read mail and the calendar and to create drafts. Permission to send, to change a deal or to write a note is requested only when you switch that ability on, so you can see exactly what you are granting and why.

The keys the provider gives us are stored encrypted, under a key that is specific to your account and itself kept under a separate master key. The AI model never holds those keys. When the assistant needs to look at your inbox it asks the system to search; the system, not the model, holds the key and does the searching.

Revoking access is one click on your connected-accounts page. We delete our copy of the key and tell the provider to cancel it. You can also cancel from Microsoft's or Google's own security page at any time, and that works even if you never log in to ours.

Connected accounts

Microsoft 365 Connected September 28, 2026
- Read mail and calendar
- Send email after your yes
Disconnect
HubSpot Connected October 1, 2026
- Read and update deals
Disconnect
Google Workspace Connected October 5, 2026
- Read mail and calendar
Disconnect

The audit log

Everything it did, and the words that authorized it.

Every action the assistant takes is written to a log you can read: the time, what you asked for in your own words, what it read back, your yes, what it did and what came back. The log can only be added to. Nobody at CallMyDesk can edit or delete an entry, and neither can you, which is the point.

You can read it on the web as "What my assistant did today", and you can have it emailed to you at the end of each day. On the Team plan, an administrator can read the whole team's log from one place, which is what a sales manager or a compliance officer asks for first. The log can be exported as a file at any time.

This is also the record a future auditor would want, so it is kept in a form that stands up: each entry carries the exact transcript line that authorized the action, the version of the assistant that handled it, and a fingerprint of the result.

What my assistant did today

2:11:52 pm
You: "Send Tom the revised quote at eighteen-five and say we can start on the ninth." Read back: email to Tom Baxter, $18,500, November 9 You: "No. Eighteen-two." Not sent
2:12:20 pm
Read back: email to Tom Baxter, $18,200, November 9 You: "Yes" Sent Message ID recorded
2:12:31 pm
You: "Note for the file: Lyndale want scaffolding quoted separately." Read back You: "Yes" Note added to the Lyndale Roofing record in HubSpot

What we keep

Audio for 30 days. Transcripts for 90. Or sooner, if you ask.

Your data is never used to train models, ours or anyone's. Call audio is deleted after 30 days and transcripts after 90, or sooner if you ask.

In detail:

What	How long we keep it	Why
Call audio	30 days, then deleted	So that a mishearing can be checked against what was actually said. Ask and we delete it sooner.
Transcripts	90 days, then deleted	So you and the assistant can refer back to recent calls ("what did I say to Tom last week?"). Ask and we delete them sooner.
Audit log	For as long as your account is open, then 30 days	It is your record of what was done. Export it before you close the account if you want to keep it.
Drafts	Until sent or canceled, then 30 days	So "send the one I drafted earlier" works.
What it remembers about your work (who's who, open deals, your preferences)	For as long as your account is open	This is the continuity that makes it useful on the second call. Ask it what it remembers at any time, and tell it to forget anything.
Your mail, calendar and CRM	Not copied to our servers	It reads them over the live connection when you ask, in the amounts needed for that call. Your mailbox stays where it is.
Your consent to be called and texted	Four years from the date you gave it, or from when you withdraw it	The law that governs calls and texts to cell phones asks the caller to be able to show the consent it relied on. We keep the wording you saw, the time, and the result of the one-time code; never anything you said on a call.
Account details and billing	For as long as your account is open, then as the law requires for invoices	Standard.

Deletion is a scheduled job that runs daily. "Deleted" means removed from live storage and from backups within a further 30 days.

On the Team plan you set the audio and transcript periods yourself, shorter or longer than the standard ones. On the Enterprise plan you can add long-term archiving and legal hold to meet record-keeping rules for regulated firms.

Not used to train

Nothing you say, nothing in your inbox and nothing in your CRM is used to train or improve AI models. Not ours, and not those of any company we use to run the service. The same rule is written into our data processing agreement and applies to every sub-processor we use.

Encrypted in transit and at rest

Every connection between your phone network and us, between us and your providers, and between us and our sub-processors is encrypted. Everything stored (audio, transcripts, log, memory, keys) is encrypted at rest. Keys to your connected accounts are held under a separate, per-account key, as described above.

Where it runs

US-hosted. Here is exactly who handles what.

In plain terms: the application, the database and everything stored (audio, transcripts, log, memory) sit in the United States. Call audio travels through our telephony provider's US region. The speech companies that turn your words into text and the assistant's words into audio, and the AI model that understands what you say and drafts what it reads back, are specialist providers in the United States, each bound not to keep or train on what passes through.

We say "US-hosted" and not "your data never leaves the United States", because a company we use may route part of its work through another region, and we would rather tell you who does what than make a promise we cannot check. The table below says what each company does; the data processing agreement says under what terms.

The contents of your inbox, calendar and CRM are yours. We process them only on your instructions, as your service provider in the words of California's privacy law and as your processor in the words of UK GDPR. CallMyDesk is today below the size at which California's privacy law applies to a company directly; if your company is covered by it, we sign the service-provider terms it requires, and they are already written into our data processing agreement. That agreement is the document to send to your compliance officer, your IT lead or your own privacy counsel, and we sign it on request for Team and Enterprise customers. We carry out a data protection impact assessment before customer data is processed and update it as the product changes; the summary is available on request.

UK customers (secondary market). UK GDPR applies to you as the controller and to us as your processor. The UK GDPR annex to the data processing agreement covers international transfers (the service runs in the United States, so your data is transferred there under the UK's International Data Transfer Addendum to standard contractual clauses), the Article 28 terms, and your rights under UK law. Nothing else on this page changes for you.

Sub-processors

These are the companies we use to run the service, what each one sees and where it processes. The full list, with contract details, is on the data processing agreement page and we email you at least 30 days before adding or changing one.

Purpose	Provider	What it processes	Where
Telephony (the phone call itself, your PIN, one-time codes, call-backs and texts from the toll-free number)	Twilio	Call audio in transit, your phone number, the text messages it sends you	United States
Speech to text	Deepgram	Call audio, turned into text as you speak	United States
Text to speech	Cartesia	The words the assistant says, turned into audio	United States
AI model (understanding requests, drafting replies and read-backs)	Anthropic	The transcript of the call, the parts of your mail, calendar and CRM needed for the request	United States
Hosting (application, database, stored audio and transcripts)	Fly.io	Everything above, at rest	United States
Payments	Stripe	Your billing details; never call content	United States and Ireland

Full list and contract terms: Data processing agreement

Microsoft, Google, HubSpot, Pipedrive and Salesforce are not our sub-processors. They are your own systems, which you have connected to the assistant, and your agreements with them continue to apply.

Your rights

Export anything. Delete everything.

Export. From your account page, download your transcripts, your audit log and what the assistant remembers about your work, as plain files you can read without us.
Delete. Close your account and the assistant's access to all your connected accounts is revoked in one step. Deletion then runs on the schedule above: audio within 30 days, transcripts within 90, everything else within 30 days of closing. Ask and we do it sooner.
Ask. Requests about your personal data, including from people whose details appear in your correspondence, come to us through the contact page. We reply within 45 days, which is the longest the state privacy laws allow, and usually within a few days. Where your company is the one covered by a state privacy law and we are its service provider, we pass on and help with the requests it receives.
Stop the calls and texts. Say "stop" on any call it makes to you, reply STOP to any text, or switch scheduled calls off on your account page. It takes effect at once, not in ten business days. Everything else about your account carries on.
Complain. If you are not happy with how we handle your data, tell us first and we will put it right. You can also contact your state attorney general's office or, in California, the California Privacy Protection Agency.
UK customers. UK GDPR applies to you: we reply within one month, as it requires, and you can complain to the Information Commissioner's Office. We would rather hear from you first.

The details, including the state laws and the lawful bases we rely on and what we collect from this website, are in the privacy notice.

It says it's an AI

It never pretends to be a person.

It says it's an AI on every call it makes on your behalf.

When you call your own line it answers with "Hello, this is CallMyDesk, your AI assistant." When it calls you with the morning brief, it opens with "Hello, this is CallMyDesk, your AI assistant, calling with the brief you scheduled. You can call me back on this number at any time, and you can say 'stop' to end scheduled calls." When it places a call to someone else at your request (to confirm a meeting, say, or move a booking) it opens with "Hello, this is an AI assistant calling on behalf of [your name] at [your company]," and it says so again if asked. It never claims to be you or a colleague. That opening line is what the AI-disclosure laws in Colorado, Maine, Utah and California ask for, and it is our rule on every call, not only in those states.

The words beside that checkbox are these, and the box is never pre-checked:

We keep a record of your yes, the time and the code check, because federal law asks a company that calls a cell phone with an AI voice to be able to show it was invited. "Stop", "cancel", "quit" or "unsubscribe" all count, in any form, and we act on them at once.

Emails it sends after your yes are your emails, in your name, because you heard them and approved them. If your firm would like a line at the foot of each email saying it was drafted with AI assistance, you can switch that on, and advisors often do.

The call between you and your assistant is a one-to-one call. It is recorded and transcribed so the assistant can work and so a mishearing can be checked, and you agree to that at signup. You are the only person on that call, and your agreement is the consent that every state's recording law asks for, including the twelve states, California, Florida, Illinois and Washington among them, where everyone on a call must agree before it is recorded. It does not record anyone else: when it calls a third party for you, the other person's side is not recorded and not kept as a transcript; the audit log holds only what you asked for and the outcome. Our full statement is on the AI disclosure page, which is also where the assistant sends people who say "disclosure" on a call.

What it does not do

The short list of things it will not do, by design.

It is a one-to-one call between you and your assistant. It never joins or records your meetings, and it never calls your prospects.
It does not make marketing calls or send marketing texts, to anyone, ever. The only calls it places are service calls you asked for, and each opens by saying it is an AI.
It does not call or text you until you have said yes at signup and confirmed your number, and it stops the moment you say "stop".
It does not send, book or log anything without your spoken yes or a keypad press. There is no "just do it" mode.
It does not move money, make payments, change passwords or delete mail. Those capabilities do not exist in it.
It does not sell, share or rent your data, and it does not use it to train models.
It does not identify you by your voice.
It does not promise to hear every name and number perfectly over a phone line. That is why it reads them back.

Assessments

What is on the Enterprise plan, what is under way, and what we do not claim.

We only show a badge when we have earned it, and we would rather say plainly what each assessment covers than imply more.

On the Enterprise plan

SOC 2 compliance.

SOC 2 compliance is part of the Enterprise plan, and the report is available to Enterprise customers under NDA. Book a call to ask for it, along with a signed data processing agreement and answers to your security questionnaire.

Under way

CASA, for Google access (under way).

Any app that reads Gmail must pass CASA, the Cloud Application Security Assessment, as part of Google's verification process, carried out by a lab Google approves. We are going through that process now. The outcome appears here, in this list, when it is issued and not before.
Microsoft publisher verification (under way).

Microsoft checks that the company behind an app is who it says it is before IT administrators can approve it for a business. We are going through it; the result appears here when it is issued.

Not claimed

ISO 27001.

Not held and not claimed.

If a page on this site ever shows a badge, you will find the letter, report or listing behind it linked from this section.

For security researchers

Found something? Tell us.

If you believe you have found a security weakness in CallMyDesk, in this website or in the demo line, send us a message with enough detail for us to reproduce it. We read every report and reply within two business days. We will tell you what we found, fix anything that needs fixing, and, with your permission, credit you here.

What we ask of you: test only against your own account or the demo line; do not access, change or keep anyone else's data; do not run anything that would disrupt the service for other people; and give us a reasonable time to fix the problem before you write about it. We will not take legal action against researchers who follow these rules in good faith.

We do not run a paid bounty program at the moment. If that changes, this section will say so.

Questions

Questions people ask before they connect an inbox

Is it safe to give an AI access to my email?

It is as safe as we can make it, and here is what that means. Your password is never shared; you grant access on Microsoft's or Google's own page and can cancel it in one click. It can read your mail, but it can only send after reading the message back to you and hearing your yes, and that confirmation is issued by a fixed part of the system the AI cannot imitate. Every action is in a log you can read.

Can it send something I did not mean to send?

Not without your yes. Every email is read back in full, including the recipient and any figures, and sending needs a spoken yes or a press of 1 within 30 seconds. "No", "stop", "change", silence or hanging up all cancel it. A misheard word is corrected in the read-back, as in the example at the top of this page. A wrong send would need a yes you did not give, and the system has no way to produce one.

How does it know it's me calling?

By a PIN you choose, spoken or keyed on every call. It does not rely on the number you are calling from, because caller ID can be faked. For a few sensitive actions it goes one step further and texts a one-time code to the cell phone you confirmed at signup, or calls you back on it, before going ahead. Repeated wrong PINs lock the line. It does not try to recognize your voice: voices can be cloned, and a PIN cannot be lifted from a recording of you.

Is my data used to train models?

No. Your data is never used to train models, ours or anyone's. What it reads in order to answer you is processed for that call and that purpose only, under a published data processing agreement that names every company involved, and the same commitment is written into our contracts with them. The security page lists what is stored, for how long, and by whom.

How long do you keep call audio and transcripts?

Call audio is deleted after 30 days and transcripts after 90, or sooner if you ask. Your audit log, which records what it did, when, and the words that authorized it, is kept for the period set out in the data processing agreement, because it is your record. On the Team plan your firm sets the windows itself, down to deleting audio at the end of the call; firms that must keep records for years talk to us about Enterprise. When you cancel, the deletion clock starts right away.

Where is my data processed?

US-hosted. The service and your data run in the United States with the named companies on our published sub-processor list, and our data processing agreement is published too, so you can see exactly who handles what. UK customers get a UK GDPR annex. In plain terms: the application, the call audio and everything stored sit in US data centers, and the AI models that understand and draft are run by vetted providers named on that list. We do not claim that everything stays inside the United States, because a speech provider may process elsewhere, and the list says who and where.

Are you certified to a security standard?

SOC 2 compliance is part of the Enterprise plan, with the report available under NDA. For Google access we are completing CASA, the independent security assessment Google requires of any app that reads Gmail, and Microsoft publisher verification is under way. We do not claim ISO 27001. What every plan gets today is on this page: sign-in through Microsoft's, Google's and your CRM's own pages with no passwords stored, a confirmation step the AI cannot bypass, an audit log of every action, encryption in transit and at rest, deletion on a published schedule, a data processing agreement and a named list of sub-processors.

What do I send my compliance officer or IT lead?

Three things: this page, the data processing agreement (with the sub-processor list and the service-provider terms California's privacy law requires), and, on the Team plan, a sample of the audit log. They will also want the retention numbers (audio 30 days, transcripts 90, adjustable on Team), the fact that the assistant never joins client meetings, and the AI disclosure page. If they have a question we have not answered, send it to us through the contact page and we will answer it in writing.

Now see what it is like on a call.

Nothing is sent, booked or logged without your spoken yes. Every action is written to your audit log. Revoke access in one click.

See how it works Download the DPA

Security questions and data protection requests: send us a message. Everything else: contact us.

DT Global Ventures Ltd · Company registration (CRN) 16042336 · VAT GB480044905 · Registered office: 3rd Floor, 86-90 Paul Street, London, England, EC2A 4NE. Contact: send us a message