Does Tresora AI calculate my figures?

No. The reconciliation, the positions and the reporting are computed by the engine and never run through a language model. The assistant reads what the engine produced, explains it, and points at the records behind it.

What can it do without anyone approving it?

It reads and it explains, inside the permissions of the person asking. Anything that changes a record or moves money goes through the approval path you already configured, with the same approvers and the same limits as if a person had started it from the screen.

Can it see data the person asking cannot?

No. It runs on that person’s role and entity scope, so an entity they cannot open is an entity it will not read. Widen somebody’s scope and their answers widen with it.

Where does the record of what it did live?

In Activity Log, under the name of the person who asked, beside every other action in the platform. It carries the routine, the records read, whatever was drafted and who approved it, and it exports with the rest of the trail.

Which model is behind it?

The model is a component we configure, and it is deliberately not something your figures depend on. Nothing on a statement, a position or a report changes when it is swapped, because none of them run through it.

On top of the engine

An AI treasury assistant that only does what you can do.

Ask it about any figure on the screen and it answers from records the engine already reconciled. Ask it to do something and it runs on your permissions, on your approval path, under the name of whoever asked — and it stops where that path says to stop.

Request a demo

Roles & Permissions · Tresora AI

Asking as

Country controller · Spain

Bank accounts in scope

Of 47 across the group. The assistant opens exactly these and no others.

In this session

Read the position for these accountsAllowed
Open the evidence behind a matchAllowed
Run a saved routine from Prompt LibraryAllowed
Draft a correcting entryNeeds approval
Release a payment batchNeeds approval

Everything it reads and everything it runs is written to Activity Log under this name.

Permissions are set once, per role, and the assistant inherits them — there is no second place where an AI is granted anything.

Three modes

What can it actually change?

Three modes, and what separates them is not how clever each one is. It is how far each one is allowed to reach.

01Ask: It reads the screen you are on and the records underneath it, then answers in words a person can check against those records. A question changes nothing in your books.
02Guide: It puts what is open in the order it costs you and takes you to the screen already filtered to it. What to do about any of it stays your call, and your threshold decides what counts as open in the first place.
03Execute: It runs a routine your team saved — pull the files, chase what did not arrive, draft the correction that always follows. Each step runs on your approval path, and the step that needs an approver waits for one.

The routines live in Prompt Library, so the question one person got right on a Tuesday becomes the one the whole team runs on the first of the month.

What it reads

What is it reading when it answers?

By the time the question arrives, the answer already exists. The matching, the positions and the reporting are computed by the engine, with no language model anywhere in them, and the assistant reads the result.

It reads

The group position, at whatever level you are answering for — group, country, entity, account.
Every statement line, and the ledger entry it was matched to.
The evidence behind a match: the signals that agreed, and the explanations that lost, with their scores.
What is open — exceptions, missing statements, batches sitting past their cut-off.
What falls due and when: maturities, obligations, the calendar the group actually runs on.

None of that is generated. It is read — out of the same tables the screens read, with the asker’s permissions applied on the way in.

The day behind an answer

Bank accounts in one position: 47; Every bank in the group, at whichever level the question was asked.
Statement lines read today: 1,284; Read on the day they arrived, rather than at close.
Waiting for a person: 33; What Guide puts in front of you first.

The record

How do I know what it did?

One morning, one saved routine, seen from the log rather than from the conversation: what the assistant read, what it left drafted, and the step it did not take.

Activity Log · Tresora AI

Routine

Duplicate payment review

Aug 14, 2026, 08:12 GMT+2

Ran as: Country controller · Spain
Records read: 3

Drafted

Reverse 1500042407EUR 2,760.00

Awaiting approval

Nothing was released. The reversal waits for the approver your policy names.

Open that entry a year from now and it holds the same records, the same draft and the same name beside it. The trail is what the assistant produces, not a setting somebody remembered to switch on.

What will not happen without you

It does not release a payment, post an entry or change a rule on its own. Each of those runs the approval path you configured — the same one a person standing at that screen would run — and the entry names who asked, who approved and what changed.

What comes back

What does a finance team actually get?

Four things, and each one lands somewhere in the platform you can point at.

An answer with its records attached: Every figure in it names the screen it came from and the entries behind it, and each of those opens from the answer.
A routine anybody on the team can run: Save the question that worked and the whole team runs it from Prompt Library, with each person’s own permissions applied when they do.
A line in the log for every run: Activity Log holds who asked, what was read, what was drafted and who approved it — beside every other action in the platform, rather than in a separate AI history.
A picture of where it is being used: Usage shows which teams lean on it and for what, which is how a routine worth saving and a screen worth fixing both get found.

Who uses it

What it answers about

The questions we get

What finance teams ask before they switch it on.

01Does Tresora AI calculate my figures?: No. The reconciliation, the positions and the reporting are computed by the engine and never run through a language model. The assistant reads what the engine produced, explains it, and points at the records behind it.
02What can it do without anyone approving it?: It reads and it explains, inside the permissions of the person asking. Anything that changes a record or moves money goes through the approval path you already configured, with the same approvers and the same limits as if a person had started it from the screen.
03Can it see data the person asking cannot?: No. It runs on that person’s role and entity scope, so an entity they cannot open is an entity it will not read. Widen somebody’s scope and their answers widen with it.
04Where does the record of what it did live?: In Activity Log, under the name of the person who asked, beside every other action in the platform. It carries the routine, the records read, whatever was drafted and who approved it, and it exports with the rest of the trail.
05Which model is behind it?: The model is a component we configure, and it is deliberately not something your figures depend on. Nothing on a statement, a position or a report changes when it is swapped, because none of them run through it.

Bring the question nobody could answer last month.

A month of statements and the ledger extract that should agree with them. We run the reconciliation on your own figures, and then you ask the assistant the thing that ended up going round the team on email.

Request a demo See the whole product

Nothing is posted to your books, and nothing is released, that you have not approved. The demo runs on the same permissions your own team would have.