Skip to content

The AI proposes. You decide. Every time.

Nummio reads receipts and suggests where each transaction goes. That is the whole extent of its authority. Nothing posts to your books because a model said so; every suggestion waits in a queue with its confidence and its reasoning, and a person says yes, or changes it, or ignores it. The claim is demonstrable rather than assertable, so the real review queue is on this page, with the reasoning switched on.

How it works, in the order it happens

1Your rules run first.
A categorization rule you've written for a vendor is checked before any model is consulted. If it matches, the transaction is filed by your rule, the AI is never called, and no AI Op is consumed.
2Then the AI proposes, with a confidence score.
For everything your rules don't cover, the model suggests an account and says how sure it is. High-confidence items are grouped so you can accept them in a batch; the rest wait for your call one at a time.
3The reasoning is shown, and you can hide it.
Every suggestion carries its rationale inline. A toggle hides it for people who find it noise; it comes back with one click.
4You decide, and the product remembers.
Accept, override with one click, or leave it. Overrides persist per business and inform future suggestions for that vendor. Nothing posts without the review path.

Two more places this shows up, both outside the review queue:

The plan never lets a model compute a number.
The plan engine in packages/plan calls no model — every figure comes from the engine, and a test in packages/plan/src/copy.test.ts fails the build if a plan sentence ever recommends a business action. A model, when it's used at all, only phrases an explanation of a line from figures the engine already found.
Payment guidelines use no model at all.
No model is involved anywhere in Payment guidelines. The pool is computed in the database — the guideline a business sees comes straight from the pool.guideline_for function, which enforces the three-businesses, ten-invoices floor before it returns anything.

The metering you can see before it bills

One AI Op is one receipt extraction or one categorization. Each plan includes a monthly allowance, the meter is in your billing settings, it turns amber at 80%, and past the allowance each op is billed at your plan's per-op rate with the running total shown before the invoice exists. Hitting the cap never blocks reading, reporting, or exporting your books. The pricing page has the allowances and rates.

The contrast

QuickBooks Online has no single switch that turns off all of its AI features at once.

We should be as plain about our own switch, so here it is: one switch in your billing settings, controlled by the business Owner, turns every model call off. When it's off no model is consulted for your business, not for categories and not for vendor names. Receipts still arrive with their fields read, your rules still run, and suggestions made before you turned it off stay on their items until you decide them. The difference that holds either way is what the AI is allowed to do without you: nothing. It cannot post, it cannot categorize past the queue, and your rules outrank it.

Your own AI, your credentials

That switch is about Nummio's own AI. A different AI can also work in your books: your own assistant, Claude, ChatGPT, or Cursor, connected from Settings → Connected AI tools and acting with your own credentials, not Nummio's. It sees one business per grant, and only what your role allows: an Employee's assistant can't do what an Owner's can. Every action it takes is written to the audit log as you, via that tool, the same as if you'd clicked it yourself. Revoke it in one click and it's gone. The two switches don't touch each other: turning Nummio's AI off doesn't disconnect your assistant, and disconnecting your assistant doesn't touch Nummio's AI.

The real review queue, reasoning on

This is the exact component the product ships, on seeded data. The "show reasons" toggle is on. j and k move, a accepts, u undoes.

8 transactions need your call.

Help

5 Nummio is confident about

3 need your call

2026-08-16 · Emailed receipt

Riverside Catering

−$340.00

Could be a client meal or an office party — the amount and vendor name don't clearly point to one account.

62% confident
No receipt attached

2026-08-17 · Emailed receipt

Home Depot

−$178.22

Hardware store purchase — this could be a repair, a supply run, or equipment, and the description doesn't say which.

58% confident
No receipt attached

2026-08-18 · CSV import

Unknown vendor

−$95.00
Uncategorized
No receipt attached

Want the same queue on your own receipts?

Forward one to your trial's receipts address and watch it arrive with its reasoning.

Start the trial
AI you control — Nummio