Proof

Everything on this page is a number we can show you the source of, on a screen, while you watch. Nothing here is a projection, an average from someone else's study, or a customer logo we were not given permission to use.

As of 22 September 2026. Re-checked at every weekly business review; if a figure here is older than a month, ask us for the current one.

What is running

2 AI coworkers in daily production use, at two companies, each with its own mailboxes, its own credentials and its own model key. Source: the agent list in our cluster. Both were Running when this page was written.

38,000+ model requests those two handled in the last 30 days. Source: per-key usage in our own model gateway, the same records their invoices are built from. A busy day is several thousand and a quiet one a few hundred, so this page states the floor rather than an exact figure that would already be wrong by the time you read it. Ask and we will open the live number on a screen.

1 scheduled job, at one of the two companies, that sends email without a person reading that message first: cold outreach from the owner's own address, four times a weekday, capped at 30 a day. Every other message either coworker produces is a draft, and a person sends it. Source: that company's policy file, which names the one job and the one purpose allowed to send and is the only thing that grants it, and the job list on the agent itself. Both open on a screen.

What it takes to move a live business without breaking it

Four weeks ago these customers ran on a different architecture. Everything moved. These are the measured cutover windows, not estimates:

What moved	Downtime
An AI coworker with five connected mailboxes, a CRM, a mailing list, a job-signal feed and six scheduled jobs	about 17 minutes
A 38-user, 42 GB file server	about 2 minutes
A company website	about 5 seconds

Each was verified afterwards by using the thing, not by looking at a dashboard: every mailbox opened, every integration called, every scheduled job run.

How the platform is kept

Every public page and API is checked every few minutes, and an outage wakes an operator. A check passes only when the real page or API answers correctly — a login screen never counts as "up".
Backups run nightly and are verified. No volume is ever deleted without a completed backup of it first; that is enforced, not a habit.
A restore has been performed end to end, from off-box storage onto a different machine, and read back.
Every change goes through Git and is applied by an automated reconciler. A manual change to a live system is reverted automatically, so the repository is always what is running.
Certificates renew automatically and are watched. One that would have expired was found and fixed 25 days early, by a check that fails below 21 days.
Assistants cannot touch the infrastructure that runs them. Their tokens are scoped so the cluster's own API rejects them.

What we will not claim yet

A failure rate.

Our model gateway is configured not to record failed requests. Any failure figure we read out of it would be zero because nothing counts them, not because nothing failed, and a number like that is not proof of anything. So there is no failure rate on this page until the gateway records one.

Hours saved. Replies won. Revenue influenced.

We believe these are real, and we are not going to put a number on them until we have a baseline that was measured before we started, a count taken after, and the customer's written permission to publish it. Until then, the honest answer is: we have two customers doing real work every day, and we have not yet finished measuring the outcome.

If you are considering buying, ask us for a reference call instead. That is a real customer's answer, not ours.

How to check any of this

On a call we will share a screen and show you the live source of any figure above: the agent list, the usage records, the status page, the backup log, the commit history. If a number here cannot be shown that way, it should not be here — tell us and we will remove it.