Most AI automation is sold as intelligence. We think that’s backwards. The interesting engineering is in doing as little of the work with a model as possible — and making the rest deterministic, cheap and predictable.
01 — Find the work
Almost nobody can list what’s automatable in their own business. The work is tacit — it lives in habits, not processes, and the jobs that hurt most are the ones nobody has ever written down.
So we don’t ask you to specify anything. We look at where the time actually goes and come back with candidates ranked by what they’re worth — including the ones that turn out not to be worth doing.
02 — Shadow it before it’s trusted
Nothing goes live on a promise. Before the system touches a single record, it runs against your last two months of real work and shows you exactly what it would have done — every decision, on every item, that you can check against what actually happened.
You correct what’s wrong. Those corrections aren’t feedback for a training run somewhere — they become the system’s rules. This single step does three jobs at once: it proves the thing works, it builds your confidence in it, and it teaches it your business before day one.
03 — Harden it until it’s boring
A new workflow leans on the model for most decisions, because it doesn’t yet know your suppliers, your customers, or how you like things categorised. That’s the expensive, least predictable month.
Every correction you make afterwards is folded back in as a fixed rule. Six months on, the great majority of decisions are settled by lookups that cost nothing and behave identically every time, and the model only sees the genuinely novel cases. The system gets cheaper, faster and more predictable as it ages — which is the opposite of how most AI products behave.
In practice
In a system we run today, roughly one line of code in eight involves a model at all. The rest is ordinary, testable software: filters that stop obviously irrelevant work before it costs anything, lookups that resolve names and accounts, a ledger that guarantees nothing is ever processed twice, and a queue for anything uncertain.
How we keep it safe
The irreversible step is never automated
Invoices are drafted unsent. Bills are drafted unpaid. Money never moves, and nothing leaves your business, without a person approving it.
That limit is structural, not behavioural
This matters more than it sounds. These systems read text written by other people — supplier emails, customer messages — and text can try to mislead software just as it can mislead a person. So we don’t rely on the system to be sensible. The layer that holds your credentials refuses to carry an instruction to pay, send or delete, no matter what asks for it. A confused agent produces a wrong draft, which a human then rejects. It cannot produce a payment.
Uncertainty escalates, it doesn’t guess
If the system can’t confidently match a customer, a supplier or an amount, it stops and lists the item for a human — with the reason. It then retries automatically once you’ve fixed the cause. Guessing quietly is the failure mode that destroys trust, so we designed it out.
You review where you already work
Drafts land in the system you already use and already trust. We don’t build you a dashboard to check every morning, because you wouldn’t check it.