An ERP with AI at the centre.

Reads freely

Ask why receivables moved, which invoices aged past sixty days, what a project has spent against its budget. It reads the trial balance, the registers and the aging the way the screens do, for the company you are looking at.

Proposes, and stops

Anything that would write, a journal, a reversal, a purchase order, renders as an action card with the exact lines. Nothing happens until you press Approve, and the approval is recorded with your name on the entry it made.

Refuses what you could not do either

A manual journal to an account a module posts automatically is refused, line by line. The person who raised a document cannot approve it. Every action it proposes is authorised against the permission that action’s own screen demands, so someone who may raise a production order still cannot make it ship an invoice. The AI works under the same controls as everyone else, not around them.

Answers out loud, if you want

Press the microphone and speak, or turn on the wake word and say “erp, what is our cash position” with both hands full. The answer appears as it is written and can be read back to you. Approving stays a click: a spoken “yes” is not a signature.

Asks before it guesses

A report that needs an as-of date, or an order that needs a quantity and a due date, raises a small form with whatever it understood already filled in. A required field it could not work out stays empty rather than being guessed at, so an aging never quietly runs to the wrong date.

A briefing, every morning

At six, before anyone signs in, it reads each company’s books and writes what matters: what needs a decision today, what to watch, what is fine, and what has changed since yesterday’s. One briefing per company per day, stored, so everybody reads the same one. A module that did not answer is named as a gap rather than guessed at.

Flags on the row

A vendor invoice that looks like one already received. A journal line four times anything that account has seen this quarter. A manual entry posted on a Sunday. These are plain rules over your own rows, not a model’s opinion: each names the rule and the figures, so you can check it by hand. A flag is a question to look at, never a verdict.

A brief before you decide

On the approvals queue, Explain says what the document is, what to check on the module’s own screen, what stands out against the approval policy, and what to ask the person who raised it. It never recommends approve or reject — that decision is yours, on the screen that makes it.

An auditor that investigates, and cannot judge

When a control fails, a red light tells you that something is wrong, not what. The auditor agent picks up the finding and reads: the control’s history to find the run it began in, then the documents behind it, through the same read tools Ask AI uses. It works under its own identity — a principal that cannot sign in, holds read permissions and exactly one write, and has no key that could post, approve or decide. The code decides pass or fail; the agent never does.

Its one write is a note on the finding: what is wrong, the likely cause, how sure it is, and a fix a person could apply. Every claim cites what it rests on — document numbers, control ids, run ids, the calls it made. If it cannot work out the cause it says so and writes nothing. A person still decides whether to accept or resolve.

And a record of what the AI did

Every run the agent makes is one row: what triggered it, which model, every tool call and what it returned, the tokens, the cost, and a hash of what it wrote. The row is inserted once and then sealed — the database refuses a second update, a change to who or what or when, a step count that disagrees with the steps, and every delete. A daily spend cap per company is checked before each turn. The first question an auditor will ask about an AI auditor is “show me what it did”, and that is the table that answers it.