How our AI works

What it does on its own, and what it always asks you about.

You are handing real work to a computer, and you should know how far it can go before it has to come back to you. This page explains that, without pretending you want to read a manual.

What it actually is

An assistant that can use your systems, not a chat window.

Plenty of software has added a chat box that answers questions about your data. Ours can do the job as well as talk about it.

When you ask it something, it works out which steps are needed, does them one at a time, and checks the result before moving on. Filing a receipt, looking up a customer, drafting an invoice: those are things it does, not things it describes.

It also starts work without being asked. A statement arrives, a customer writes in, a lead goes quiet, and it gets going. You see what it did when you next open it.

A morning, start to finish

How did last month go?

  1. Pull last month, compare it to the three before it, find what changed
  2. Looked at the profit and loss
  3. done
  4. Compared it to May, June, July
  5. done

You made 4.1 percent less on each job than usual. Sales were fine. Contractor costs went up by 6,240.00 against June, and almost all of that was the Henley job.

The brakes

It cannot talk itself into spending your money.

Every job it can do is filed under how much damage it could cause. That filing happens in our software, not in the AI, so no amount of clever wording gets round it.

  • Reading is free. Looking things up, adding things up and answering questions never needs your permission.
  • Writing is written down. Anything it changes is logged with the time, what changed, and what it was before.
  • Spending, sending and deleting stop. Money leaving, an email going to a customer, or a record being thrown away always comes to you first.
  • You set where the line is. Refunds under fifty dollars can go through on their own if you want. Refunds over it cannot. That is your call, job by job, and you can change your mind.
What always stops, whatever your settings say
  • Deleting anything with history attached to it
  • Reopening a month you have already closed off
  • Sending money, promising money or refunding money
  • Sending anything to somebody outside your business
  • Giving another program access to your account

This list is fixed. Loosening your other settings does not shorten it.

The AI itself

We built it, we trained it, and we run it.

Most AI products are a thin layer over somebody else's AI. When you type into them, your information goes to that other company to be read. Ours does not, because there is no other company involved.

We start from AI that is published openly for anyone to use, then train it ourselves on the specific job each product does. An AI taught bookkeeping properly beats a much bigger general one that has merely been asked politely.

It runs on computers we own. That is partly about privacy and partly about cost: it is the reason we can charge what a one-person business can actually pay.

Whose AI
Ours. Trained by us for each product.
Where it runs
On our own machines. Nothing is sent out to be read.
Your data
Never used to teach it. Not now, not quietly later.
If it is having a bad day
The product still works. Your records do not depend on the AI being awake.

If you build software

You can point your own tools at ours.

This part is for developers. If that is not you, nothing here changes what the products do, and you can safely skip it.

Every product has a proper interface for other software to talk to, covering the same jobs its own assistant can do. Whether you are wiring up an internal script or pointing a general-purpose AI assistant at your books, you get the same set of jobs we use ourselves. There is no better hidden version that only we can reach.

You choose what it is allowed to touch, and you can take that back at any time. The brakes above apply to your software exactly as they apply to ours. In fact they are tighter: anything arriving from outside that would change your data waits for a person, even when our own assistant would have been allowed to get on with it.

Ask us for access

your software asking ours to do something
→ your agent asks
{
  "tool": "issue_invoice",
  "customer": "Marchetti Design",
  "amount": "3500.00"
}

← we answer
{
  "status": "waiting for a person",
  "why": "Sending an invoice always needs approval.",
  "shown_to": "the account owner"
}

Still not sure you want a computer doing this?

That is a reasonable way to feel. Tell us what worries you and we will give you a straight answer, including when the answer is that we are not the right fit.