AI lets criminals clone your CEO's voice, fake your supplier's emails and write flawless invoices. VouchSafe makes sure no money moves and no bank detail changes until the real person confirms it on a number you already hold.
Business email compromise was the second-costliest crime type reported to the FBI in 2025, after investment fraud, and generative AI makes it cheaper, faster and more convincing. Small firms are exposed because approvals run on trust and familiarity.
VouchSafe texts or calls a one-time code to the number in your trusted directory, never one taken from the message. The real contact reads it back. An impersonator can't.
Larger payments need a different person to release them. Whoever opened the case or made the callback can never give the second approval.
Detecting AI-written text and deepfakes is an arms race defenders lose. VouchSafe scores what attackers can't hide — where a request really came from and what it asks for — then enforces the one control that beats every variant.
The Outlook and Gmail add-ins score each message in the reading pane: lookalike domains, display-name spoofing, failed authentication, payment-detail changes, secrecy, urgency, code requests.
Risky requests become a verification case with the exact steps your policy requires. Anyone on the team can stop a payment.
VouchSafe texts or calls a code to the number in your trusted directory — never one from the message. The real contact reads it back. An impersonator can't.
Above your threshold a different person approves. Every step lands in a tamper-evident audit log your bank and insurer can rely on.
Built for businesses with no security team: plain-language verdicts, policy you set once, and steps staff can follow under pressure.
Risk verdict and next steps inside the inbox, with one click to open a case. Works in Outlook desktop, web and Mac, and Gmail on Google Workspace.
One-time codes by text or voice to the number on file. Changed numbers are locked out of callbacks for 24 hours, so a hijacked account can't swap the number and approve itself.
We're testing a guard that watches supplier records in your accounting software and opens a case when they change. It isn't available yet, and what it can see depends on your ledger.
Requesters, verifiers, approvers and admins. The person who opened a case or made the callback can never give the second approval.
Every action, hash-chained with SHA-256. Verify integrity in one click and export it as evidence for your bank, insurer or auditor.
Realistic deepfake, voice-clone and BEC scenarios for team meetings, plus a minute-by-minute playbook for the first hours after a loss.
AI assistants now read invoices and prepare payments. They can be fooled by a fake invoice, or by instructions hidden in a document, just as people can. VouchSafe applies the same rule to them: the assistant asks, and your people confirm.
An admin gives your bookkeeping or bill-pay assistant an agent key. Before it pays, it tells VouchSafe who it wants to pay, how much, and to which account.
The request becomes a case for your team: a callback to the number on file, and a second approver above your threshold or when bank details are new. The assistant can't approve it, enter the code, or change a number on file.
It records who confirmed what, and when. Anyone you share it with can check it hasn't been altered, and it stops matching if the amount or the account is changed afterwards. Check a receipt.
What it doesn't do. It covers the payments an assistant asks about. It isn't connected to your bank, so keep your bank's own two-person approval switched on. Your team isn't alerted when an assistant asks: the case appears in the app. A refused request can't be asked again for 24 hours. This is new. We have run it end to end with Claude, connected by sign-in. It has not yet been tried with ChatGPT or other assistants. It connects through the Model Context Protocol or a plain web API.
We test the risk engine the way you'd want any fraud control tested: on messages it was never tuned on, with the uncertainty shown.
We don't publish accuracy figures yet. Our current test sets have 50 messages each and were written by the same people who wrote the rules, so any number from them would overstate how well VouchSafe works. We'll publish results once a test set has at least 500 messages and was written or collected independently.
Detection is the first line, not the last. A polished, error-free message is not evidence it's genuine. So a low score never removes a verification step: any request that moves money still needs a callback on the number on file and, above your threshold, a second approver.
A fraud control you over-trust is its own risk. Here's exactly where VouchSafe's job starts and stops.
One flat price per business, not per seat. Every plan keeps the rule that matters: no money moves until the real person confirms it.
For trying VouchSafe and training staff
For businesses that pay suppliers by email
For larger teams that want hands-on help
Prices in US dollars, excluding tax. Billed monthly by card through Stripe; cancel any time from your billing page. We run VouchSafe for you and send callbacks from our own phone account, so there's nothing to install and no Twilio account to set up. Callbacks over your plan's allowance are never blocked: they're billed at $0.15 each. Prefer your own Twilio account, or to run VouchSafe on your own server? Both are available. MSPs and insurers rolling VouchSafe out to many clients: ask us about volume pricing.
It doesn't try to judge whether a voice or video is synthetic — detectors are unreliable and attackers iterate faster. It treats any money or access request over voice or video as unverified and routes it to a callback on a number you already hold, which works however good the fake is.
We host VouchSafe for you on Render, a US cloud provider, in its Virginia (US East) region, on a disk that is encrypted at rest and backed up with encrypted daily snapshots. Each business's directory, cases and audit log are kept in a separate database file of their own, not mixed into a shared table with other customers. Accounting connection tokens, and your own Twilio credentials if you add them, are encrypted with a key unique to your workspace. If you'd rather keep everything in-house, you can run VouchSafe on your own server instead. The free offline checker stores nothing outside your browser.
Not yet. We're still testing that guard and won't promise it until it's proven. In our testing, a US QuickBooks sandbox company didn't share vendor bank details with outside apps, so the guard may not be able to see them. Your callbacks and second approvals don't depend on your accounting software.
Your plan includes a monthly allowance of text and voice callbacks (100 on Team, 300 on Business), sent from VouchSafe's own phone account through Twilio. You don't need an account of your own. Going over is never blocked, because a fraud check should never wait on a quota: extra callbacks are billed at $0.15 each. If you'd rather use your own Twilio account, connect it in Settings and Twilio bills you directly. Staff can always record a callback they place themselves at no cost.
Small teams are the main target precisely because one person can be talked into one payment. VouchSafe works with two people: one verifies, the other approves.
VouchSafe was built by Dr. John Négus Adu, a legal scholar who has spent his career on how criminals use technology to steal, and how law and process can stop them. His conclusion: you can't out-detect a good fake, but you can out-process it.
More about John and VouchSafe →
Set up your workspace, add your top suppliers and executives, invite a second approver. That's it.