How it learns — Likvid

A bank feed is for sale and the model is the same one everyone has. What isn't for sale is how your company's money behaves — and that is what your corrections teach it.

Questions per month, as it learns a company.
This is what the product is built and measured against, not an average of anyone's results.

the target curve

45MONTH 1

262

143

3MONTH 6

We read it

An invoice number in the memo, an amount that matches. Free — and free for everyone else too.

A person told us

Someone who understands the business answered one question. Permanent. This is the asset.

Inherited from a decision

A new item settles itself from an answer you gave weeks ago, to a different item. This is the compound interest.

The model guessed

Named on the item as the Likvid AI's candidate — separately from what's verified, with the evidence that would settle it. Worth exactly as much as the next human nod, and marked as such until it gets one.

The one structural rule

There is no such state as "unmatched." A movement is either explained — and you can see which source explains it — or it carries one question for a person, whose answer is kept and reused. No third state, and no confidence score standing in for one.

What an answer carries

A decision isn't a tick on one row.

When you answer, three things are recorded with the answer. That is what makes the next item settle itself instead of coming back to you.

Its scope

This partner, this kind of movement, from this date on. A scope that only closes a single row means the question was too narrow to be worth asking.

Its evidence

What was on the screen when you decided — the invoice, the memo, the amount, the history behind it. The decision stays checkable later, by you or your accountant.

Its authority

The next machine run does not overwrite it. A rejected suggestion stays rejected, and new information is appended, never written over what a person decided.

Memory

What it remembers — and where you can see it.

A decision on an item is kept with its scope — that has always been the asset. Three further things are kept, and none of them is lifted out of a conversation.

Lessons you approved

When you record a decision with your own reasoning, that reasoning becomes a candidate lesson. It goes live only when a person accepts it — accept, narrow its scope, or reject — and a live lesson can be revoked. The next conversation and the next investigation get it as context.

Precedent

When a new case looks like a closed one — same partner, same shape of amount — the closed case and what settled it are shown next to it, in the chat and on the case itself. A precedent is shown, not applied.

Preferences you ask it to keep

"Keep the replies short", "call that account the payroll account" — kept when you ask, listed on your account page, removable one by one.

The limits

What it cannot learn — and won't pretend to.

One-off movements stay one-off. Learning needs repetition: a partner that appears once, with an amount that never comes back, will be asked about once — and that is the correct outcome, not a failure.
Nothing is inferred from an empty source. If a system returned nothing, that is a gap in the data, not a fact about your company. An empty result never becomes an answer.
The model never grades its own certainty. There is no "0.73 confidence" here. An item is either explained with its source visible, or it carries one question.
What only exists in someone's head has to be said once. A verbal agreement, an owner's decision, a deal made on a call — the product asks, and then remembers. It cannot read minds, and it doesn't guess in their place.
Conversations are not memory. What you say in chat is not kept. A lesson exists only because you recorded a decision with your reason — and only once a person accepted it.

It asks once. Then it stops asking.