Docs
/
EN DE

AI Chat

Some questions are faster asked than clicked. “Who owes me the most?” means picking a report, setting a filter, and reading a column. AI Chat answers it in a sentence — from your own books, with the invoice numbers linked so you can check the work.

It does three things: it answers questions about your data, it explains how to do things in neoo, and it drafts bulk corrections for you to confirm. It never changes anything on its own.

Open Cockpit → AI Chat (requires the AI chat permission).

What to ask it

Questions about your books. “Summarize this month’s cash position”, “Show my top customers by outstanding balance”, “List any overdue invoices over 30 days”. The numbers come straight from your ledger — nothing is estimated or rounded on the way — and the answer arrives as a table or a short list, with invoice numbers, customer names and account numbers turned into links that open the record or the filtered journal behind it. That’s the part worth using: the answer is a starting point you can click into, not a dead end.

How do I …? “How do I set up bank accounts?” Questions about using the software are answered from this documentation, with a link to the page it used and, where one exists, a link to the configuration page in the app. When a topic isn’t covered here, it says so and points you at Support rather than inventing steps.

Bulk corrections. “Move everything booked to 6500 last quarter over to 6510.” This is the one place the assistant reaches for your data rather than reading it — and it does so by proposing, never by doing. The next two sections cover it properly.

Where it genuinely beats clicking around is the exploratory question — the one you’d otherwise answer by opening three reports to find out it wasn’t worth asking. For a figure you’re going to file or send to someone, use the report that was built for it.

What it can see

It is confined to the company you have open. It reads that company’s books and nothing else — it has no route to another company’s data, whoever else you have access to.

Within that company it reads your accounting records: journal and ledger postings, customer and vendor transactions, invoices, your customers and vendors with their addresses and contacts, the chart of accounts, tax rates, currencies and exchange rates, projects and departments, your services, bank accounts, bank transactions and matching rules, payments, and your workstations with their approval rules and transfer history. It also knows your custom fields by name, so you can ask about them the way you labelled them.

It does not read the document files themselves — it works from the transactions those documents produced, not the PDFs. It has no view of user accounts, roles or permissions. And asked about anything that is neither your accounting data nor using neoo, it declines rather than answers.

Proposed changes — you are the confirm button

The assistant cannot change your data. What it can do is prepare a change and hand it to you to approve. Confirming is the only thing that writes anything, and confirming is only ever done by you.

What it may propose is deliberately narrow: a bulk update to transactions that are already posted, in one of three areas — vendor transactions, customer transactions, or the general ledger — changing only a small, fixed set of fields:

Field Where
Date All three areas
Due date Vendor and customer transactions
Vendor / Customer Move transactions from one to another
Department, Project All three areas — set or cleared
Record account, Expense account, Income account, Account Rebooking postings from one account to another

Everything else is outside its reach. It cannot change an amount, post, delete, reverse, pay, or create anything — ask it to and it will tell you plainly that it can’t, and where to do it yourself.

Two limits apply before a card ever appears. Each of the three areas is gated by its own batch-update permission — the assistant only offers what your permissions grant, so it can never be a way around one you don’t hold. For vendor transactions and the general ledger those are the same permissions as the manual Batch Update screens; for customer transactions the assistant is the only bulk-correction route. And transactions still waiting for approval are left alone; only posted ones are eligible.

What the card shows

When a request fits, the reply is followed by a Proposed change card. At this point nothing has happened.

  • A summary of the change, and the area it affects.
  • Changes — every field with the value it would be set to, resolved into something readable: the vendor’s actual name, the account with its number and description, None where an assignment is being cleared. If the target no longer exists, you get an explanation instead of a card.
  • The transactions themselves — a count (“12 transactions will be updated”) and the list, with invoice or reference, date, name or description, the account where one is affected, and amount. The first eight are shown; Show all opens the rest. This list is the change. Read it rather than the summary sentence.
  • Confirm update (12) and Cancel. The number on the button is exactly how many transactions will be touched.

Transactions in a closed period

Anything caught by the change that falls inside a locked period is listed separately, under its own warning, and skipped by default. Confirming leaves it untouched.

The toggle Also update transactions in the closed period includes them, and the count on the confirm button changes to match. It’s there for the case where you genuinely mean it — reopening a filed period is a decision, not a side effect, so it’s never made for you. See Lock Period for what a lock means.

What confirming does

Confirm update re-checks that the affected transactions are still exactly the ones the card listed. If anything changed in the meantime — someone edited one of them while you were reading — nothing runs, and you get a fresh card with the current picture to review again. What you confirmed is what gets executed, or nothing does.

The changes are then applied one by one through the same machinery as a manual batch update, and the card reports back: how many were updated, how many failed, how many were skipped (closed period) — and, if you included them, how many were updated in closed period. Failures are listed individually with the reason and what to do about it, and they don’t stop the rest. Every change is recorded in your transaction history under your name, marked via AI Assistant — it is your change, and the audit trail says so.

Cancel discards the proposal; the card says so and nothing is touched. A proposal is also good for about fifteen minutes — leave it sitting longer and it expires, and you ask again for a fresh one.

A single proposal covers at most 500 transactions. Past that you’re asked to narrow the request — with a date range, for example — rather than being handed a card too big to actually read. The cap exists because a change you can’t review isn’t one you can confirm.

Working with the answers

Any answer can be copied with Copy. When it contains a table, Export PDF and Export Excel take the tables out as they stand — useful for passing a quick figure to someone who doesn’t need the whole report.

The conversation lasts as long as the page. It remembers what you asked earlier, so follow-ups like “and the same for last year” work, but navigating away or Clear conversation starts fresh. Nothing is carried over to your next visit.

Limits and good habits

  • Check anything that matters. The assistant reads real numbers, but it also decides how to interpret a question — which dates, which accounts, whether “revenue” means what you meant by it. That interpretation is where it can be wrong while sounding perfectly confident. Click one of the linked rows and see whether the underlying transactions are the ones you had in mind.
  • An empty answer is usually the question. No results almost always means a date range or a filter that didn’t match — not that the data is missing. It will say so; ask again with a narrower or wider period.
  • Big questions can time out. Each data query behind a question gets about half a minute. A question spanning years of line-by-line detail may run out of time; ask for a period at a time.
  • Vagueness gets a question back. Asked something ambiguous, it asks you to narrow it down instead of guessing — particularly for changes. That’s the behaviour you want; answer it rather than rephrasing around it.
  • It is not the report. For what you file, send, or sign, the reports are the authority. The assistant is for finding out fast, and for the tedious corrections in between.