Automate the work your software was never able to touch.
The repetitive, judgement-heavy jobs that Zapier can’t script and your accounting package doesn’t do — reading the email, working out what it means, and getting it into your books correctly.
| What arrives | What it becomes | Before anything happens |
|---|---|---|
| An enquiry, written in prose | A priced quote, against your own rates and past jobs | You approve it |
| An invoice going unpaid | A chase, pitched for that customer and how late it is | You approve it |
| A supplier invoice, as a PDF | A bill, matched against what you actually ordered | You approve it |
| A booking or job request | A diary entry, a confirmation, and the paperwork | You approve it |
| A renewal or certificate notice | A dated reminder, with the document filed | Flagged, never actioned |
Different jobs, one shape: something arrives that a person has to read and judge, and it ends up correctly entered in a system you already use. Nothing leaves your business without you.
£0.04
Cost per run once it knows your business
0
Payments or emails sent without a human
Same day
Connected to running unattended
Scope
Not the jobs that were already easy.
If a rule could do it, you’d have automated it years ago. These stayed manual because they need someone to read, interpret and decide.
Inbox to ledger
A supplier invoice arrives as a PDF. A customer asks about a booking in plain prose. Both need reading, judging and entering correctly — and neither is a form.
Systems that don’t talk
The supplier portal with no API. The spreadsheet that feeds the invoice run. The monthly report someone assembles by hand from four places.
Judgement calls
Is this complaint urgent? Does this invoice match the order? Is this a refund or a reimbursement? Rules can’t answer these. Under supervision, software now can.
Controls
Nothing irreversible. Ever.
Invoices are drafted unsent. Bills are drafted unpaid. Anything the system isn’t confident about goes to a review queue rather than being guessed at.
The irreversible step stays with a person — enforced by the architecture, not by the system behaving well. These agents read text other people wrote, and text can mislead software as easily as it misleads a person. So the layer holding your credentials refuses to carry an instruction to pay, send or delete, whatever asks for it. A confused agent produces a wrong draft, which you reject. It cannot produce a payment.
Method
It gets cheaper the longer it runs.
Find it
Most owners can’t list what’s automatable — the work is tacit. We look at where the hours actually go and come back with candidates, ranked by what they’re worth.
Shadow it
Before it touches anything, it runs against your last two months and shows exactly what it would have done. You correct what’s wrong. Nothing goes live until you’ve seen it work.
Harden it
Every correction becomes a permanent rule. The system consults the model less each month, so it gets cheaper, faster and more predictable as it ages — the opposite of most AI.
Record
A UK youth charity, running daily.
A volunteer treasurer was losing evenings to turning email into bookkeeping. Hall-hire requests written in prose. Expense claims with receipts attached as PDFs. Supplier invoices. The occasional overpayment needing a refund.
It now runs once a day, unattended, and drafts each one into QuickBooks against the right customer, account and section. Anything it isn’t sure about is listed for review instead of guessed. The treasurer checks the work in QuickBooks, where they already reconcile — we built them no dashboard to learn.
Boundaries
We won’t replace your existing tools.
Zapier and Make are very good at moving structured data between systems. Your accounting package is getting better at reading receipts every year. Keep them.
We take the part none of them can reach: work that spans several systems at once, arrives as prose rather than fields, and needs a decision before anything can happen. If a cheaper tool can do your job, we’ll tell you on the first call.