That is the bit that keeps people out. Not the price. The fear that one night it emails a customer something daft in your name, or that your debtor book turns up on somebody else's screen. Here are the four questions people actually ask, answered the way you asked them. Where the answer is not good enough yet, it says so.
Not unless you tell it to. Out of the box every email waits for your tap: overnight it writes, in the morning you decide. Once a reminder stage (before due, polite or firm) has a record of drafts approved unedited, an owner or manager can let that stage send on its own. Escalation and final notices never can, and an "approve every send" switch turns it off. The AI itself has no send button in any of this.
Overnight it reads your ledger and writes the chases. In the morning you read them and decide: send, change it, or skip it. That is how it ships. Letting a stage run on its own is a switch an owner or manager can flip, only for the before-due, polite and firm reminders, and only once that stage has a record of drafts you approved unedited.
When you approve, the exact text you were shown is fingerprinted. If one word of it changes between your tap and the send, the send is refused rather than guessed at. You can never be shown one email and have another go out.
The AI side of Harness has read tools and drafting tools. Sending is not one of them. Ask it to send something and the most it can do is stage a draft and ask you. That is a wall in the code, not an instruction we hope it follows.
The one recipient trick worth knowing about: the address you approved is the address it uses. It never works out who to send to again at send time, so somebody replying into an email thread cannot quietly redirect your approved email to themselves.
No. Every question Harness asks its database carries your company on it, so a query that forgets whose data it is asking for does not return somebody else's, it returns nothing. Under that sits a second layer in the database itself, which holds the same rule at the row level for every per-customer query.
What your team teaches Harness is shared inside your company, so nobody teaches it twice. It never crosses to another company. We run no AI model of our own: drafting uses your own ChatGPT or Claude account, under your settings with them.
Every fact Harness knows is listed in your control room with who taught it and when, and a button to delete it. A viewer cannot change what it knows. An owner can.
The database-level layer binds when a flag is set on the server. It is set in production. A wrong deploy could still run without the flag, so welding it on permanently is on the list. We would rather you read it here than find it.
Harness runs on one server we rent from Hetzner in Falkenstein, Germany, inside the EU. Not on a laptop. Heoka Partners Limited is a UK company registered in England and Wales, so UK data protection law applies and the Information Commissioner's Office is who you complain to if we get it wrong. The privacy policy lists the few other services involved and what each one does.
Invoices, quotes, contacts, what was chased and when, the drafts, the approvals, and the facts your team has taught it. We do not hold your card details; the payment provider does that.
An owner can disconnect the accounting system, or the mailbox (Gmail or Microsoft 365), from the control room. It takes effect there and then: the stored keys are dropped, and with no connection there is no reading and no drafting. Cancel and the billing stops. You do not need us to agree to any of it.
An owner can download "Export my memory": what Harness has learned, your rules and lists, the names of your connections, and counts of your mail and invoices. It does not include the mail or the invoices themselves. You can delete any fact Harness knows from the control room, one tap each. A full export, and a one-click wipe of the account, are not built today. Ask by email and we do it by hand and confirm when it is done. When those buttons exist this page will say so.
Switching AI tools is different from exporting your Harness account. ChatGPT and Claude connect through MCP, while the facts, sources, rules and activity log stay in Harness. Move between those supported chats without rebuilding your business context. A full self-service export is not available yet.
The thing that matters most is the thing you already own: the accounts and the mailbox are yours, at your provider, and Harness is a guest in both. It never becomes the place your business lives.
This is the real thing, not a picture of it. Pick a provider below and the assistant on this site starts answering through it straight away. Then run the check: it asks the new provider a fixed question with the full business context attached, and shows you the fingerprint of the rules layer. Same fingerprint on every provider means the prices, the approve-first rule and the facts all survived the switch, because they live in the Heoka layer, not in the model.
The assistant on this site is now answering through Google Gemini. The chat box on any page uses your choice from this moment on.
Harness works the same way at company scale. Your facts, your sources, your rules, your approval trail and your activity log sit in your Harness, and the AI provider is a part you can swap. Move from ChatGPT to Claude and nothing about your business is rebuilt.
The dangerous failure is the quiet one: it stops reading your ledger in March and nobody notices until June. So the check is built the other way round. You see the work every morning, and we watch for the night it did not happen.
You read the drafts daily. If the queue is empty on a morning you know money is owed, that is the alarm, and it reaches you within a day rather than a quarter.
One page, every send, who approved it, and the count of emails that went out without a person behind them. That number stays zero unless an owner or manager has switched a reminder stage to send on its own, and it is printed at the top, not buried.
A connection that quietly expires or a night job that did not run alerts us rather than sitting in a log nobody reads, and the server itself is checked every five minutes.
Every night the database is encrypted and copied to a second provider, Oracle in London, so losing our main server does not lose your history. Rebuilding a whole replacement server is not timed yet.
It is built from the approval trail, not from a report we write about ourselves. Below is the real shape of it, filled in with made-up rows so you can see what you would get.
A plumbing firm in Doncaster · 2026-08-01 to 2026-09-01 · every send below was released by a person
| Date | To | Reference | Approved by | Release |
|---|---|---|---|---|
| 2026-08-04 | accounts@example-build.co.uk | INV-1187 | jane@example.co.uk | released by tap |
| 2026-08-11 | pay@example-estates.co.uk | INV-1163 | jane@example.co.uk | released by tap |
| 2026-08-19 | finance@example-fitout.co.uk | INV-1201 | mark@example.co.uk | released by tap |
| 2026-08-27 | accounts@example-build.co.uk | INV-1187 | jane@example.co.uk | released by tap |
| 19 further sends, same shape, trimmed for this example. | ||||
| approved | 23 |
| release_executed | 23 |
| skipped | 9 |
| snoozed | 4 |
| draft_updated | 7 |
Generated from the Harness approval trail. By default, nothing leaves Harness without a content-bound human approval, unless an owner or manager switched that stage to automatic. A send with nothing behind it would appear in this table in red, and in the first number at the top.
Read it yourself, on your own overdue invoices.
The fastest way to trust this is to watch it write a chase to somebody who genuinely owes you money, and then not send it. Twenty minutes, your ledger, your call.
Still not answered? Email connor@heoka.co.uk and ask the awkward version of the question.