Trust — Likvid

The hard rules

Six rules that hold, every run.

Zero outgoing actions without approval

An invoice, a document or a letter leaves only after a person approves that specific item. The daily run writes the reminder, gathers the evidence, and stops. There is no send-without-asking path in the code. The same holds for anything an AI tool you connect proposes: it lands as a card and waits.

Bank access is read-only

Bank data arrives over a regulated PSD2 channel through a licensed account information service provider. It cannot move money and has no payment mandate. You grant access in your own banking app and revoke it there; consent expires roughly every 90 days.

Every machine claim carries its source

For any number on any screen you can go back to which system said it and which document it came from. A figure with no traceable origin is not displayed as fact.

Tax paragraphs come from a closed catalogue

Deductibility rules live in a hand-curated catalogue where each category is bound to a fixed deductible share and a fixed statutory reference. The model may only pick a catalogue entry — it cannot write a citation, so a fabricated paragraph cannot happen. Where a sub-paragraph isn't verified against the current statute, the catalogue cites only to section level.

If sources contradict, the chain stops

An unmatched credit from the same partner, a payment that arrived after the draft was written, a reversal nobody linked — any contradiction stops the chain and sends the item to a person. The most expensive mistake a finance system can make is chasing someone who already paid.

A human decision is sovereign

The next machine run does not overwrite your decision. Decisions carry a scope — this partner, this kind of movement, from this date — and a rejected suggestion stays rejected.

The split

What the agent owns, and what the engine owns.

Likvid is agent-native and it is still a finance product. Those two are compatible only if the boundary is explicit: the agent does the work, the engine states the fact, and you approve the decision.

The agent owns

The deterministic engine owns

Reading and interpreting: what this movement looks like, what's unusual about it, what is worth asking.

Which payment belongs to which invoice. Scored matching with tax-number direction, cent-level tolerance, IBAN and invoice-number anchors.

Composing: the reminder, the partner file before a negotiation, the summary of the day.

The chain to the end: reversal, part-payment, FX in the original currency, bank charge nested under the movement, card settlement split back to invoices.

Assembling an approved command: a completion certificate, a closing invoice, a payment run.

Which VAT rate comes from which statutory section — from the closed catalogue, never from the model.

Explaining a decision in plain language, and pointing at the evidence for it.

The confidence band, computed by a fixed weight table from what was observed. The model does not grade its own certainty.

Evidence

Evidence, not confidence.

When Likvid proposes something, it doesn't tell you how sure it feels. It tells you what it saw: the invoice is issued according to the tax authority · the invoicing tool still shows it open · there is no matching credit in the bank · the due date has passed. Those are facts with sources, and the band that comes out of them is computed by a fixed table, not by the model.

Why this matters

A number like "0.73 confidence" tells you nothing about why, and there is nothing to learn from it. So the product doesn't have one: an item is either explained, with the source visible, or it carries one question for a person — whose answer is then kept and reused.

Tax

Where the numbers stop being ours.

Likvid reads the tax authority's own records — issued and received invoices, and the tax current account, refreshed daily. That gives a live position: VAT payable and deductible, upcoming obligations, and any overpayment sitting at the authority. Where the company switches on the VAT-return access, it also reads the authority's own list of the month's documents with its tax codes, and lists the differences against the books.

Year-to-date is the reliable figure. The monthly allocation is there for decisions, not for filing. We say this out loud rather than letting a number imply precision it doesn't have.
Classification, filing and the official close stay with your accountant. Likvid produces the material they can check against their own analytics, with every line pointing at its origin.
Two layers, at two different stages. The deductibility engine has been checked line by line against a practising accountant's work. In the interface it only asks: a deductibility question on an item is answered by a person, and the machine's own proposal never becomes an active rule on its own. The tax-type catalogue that runs in the interface is closed by construction, and its line-by-line review with an accountant is still open.

Data

Your keys, your data, your records.

Read-only where it matters. To be precise: Likvid is not a purely read-only system, because it writes to your invoicing tool when it issues an invoice you approved. The bank side stays read-only, and what goes out, goes out item by item with a human approval on it.
Credentials are stored encrypted, in a form that cannot be read back out of the database.
EU data residency, enforced on every model call. Processing is pinned to the EU with zero data retention, and it fails closed — if the region can't be honoured the call errors rather than quietly running elsewhere. Infrastructure and storage in the EU; sub-processors published.
Your records stay yours. Deleting a company deletes its data — with one deliberate exception: the record of who did what and when survives, because a deletion is itself an event that has to stay auditable.
Auditable by design. Decisions, approvals and machine runs are logged with who, when, and on what basis.
A guard watches the sources. If a feed goes quiet or a consent is about to expire, you get told.
External AI tools get a door of their own. A tool you connect signs in with your Likvid account, is granted one company and named permissions on a consent page, is listed under External tools, and is revoked there. It reads; what it proposes waits as a card for your yes. Every call is logged — which tool, which company, which reader. What a connected tool can and cannot do →

Privacy notice · Terms · Sub-processors — currently in Hungarian.

Honesty

What it doesn't do — yet, or at all.

A trust page that only lists strengths isn't a trust page. This is the current state, and we keep it current.

It is not your accountant. Classification, the VAT return and the official close stay with a professional. Likvid gives you the financial picture between closes, and the material to check.
The books are not complete on their own. There is a double-entry ledger behind the product, with a trial balance and a journal — but the opening balances reach only as far as the bank, and the rest sit with your accountant. No revenue recognition, no statutory consolidation.
It does not remember your conversations. It keeps two things you can see: preferences you ask it to keep, listed on your account page, and lessons you approved from your own recorded decisions. What you said in chat last week is not one of them.
The deductibility engine never decides on its own. It asks; a machine proposal stays a proposal until a person accepts it.
What a connected AI tool does with what it reads is that tool's affair. Our EU-processing guarantee covers Likvid's own model calls, not the tool you connected.
Advance invoices have edge cases. Where a final invoice settles an advance, the invoicing provider can override what we send. We don't demo that part as finished.
Hungary only, today. The connectors are Hungarian rails.

In a finance product, this is the product.