Xero and Ai — asking your books in English, and getting the audit trail for free

From my perspective, the most telling thing about most Ai-meets-accounting conversations is the order of the questions. Software vendors ask first whether the Ai can post invoices, reclassify journals and clear down suspense accounts. Accountants, when I actually speak to them, ask first whether the Ai can be stopped from doing any of those things. That tension is worth paying attention to, because it tells you something true about where Ai adds value in a regulated practice, and where it starts to become a liability.

I believe read-only is the right first step for Ai inside Xero, not because write-back is impossible, but because the teams who would benefit most from it already have too many ways to make a mistake. An accountant working across eight client files in a week does not need an Ai that might post to the wrong one. They need an Ai that answers their question quickly, gives them the number in context, and leaves the ledger untouched.

The questions accountants actually ask their books

If you sit with a practice accountant during a typical Wednesday, the questions they ask Xero out loud are not the ones Xero is designed to answer in a single click. Who is more than thirty days overdue, and how much do they owe us in total. What did our March P&L look like, top three revenue lines, top three expense categories. Is there anything unusual in last week's transactions that I should look at before signing off. Are there two customers with the same VAT number but different legal names, which usually means one of the client's bookkeepers has set up a duplicate contact.

Each of those questions is a short sentence in English and a long click-path inside Xero. Most practices solve the gap by building saved reports or exporting to Excel, and specially when the practice has many clients, the saved-report approach gets fragile quickly because every new client file needs its own set. Ai flips the problem. The question stays in English. The answer comes back with context. The practice manages one connection, not fifty saved-report configurations.

Why the audit trail is the actual product

When I talk to regulated accountancy practices, the part they care about most is not speed. It is evidence. The ICAEW, ACCA and their insurers all want to see who looked at what, and when, and under what basis. A rushed Wednesday answer is only useful if you can later prove that it was the right accountant asking about the right client at the right time, and not a junior with excess permissions or a system that cached a query across tenants.

Apart from this, the audit requirement is the reason we made Xero read-only by design. Every call is logged with the connected user, the Xero tenant, the specific question, and the redacted payload that the language model saw. Business customers retain that audit for ninety days by default. Enterprise customers configure retention to match whatever their professional body and insurer require, which specially for larger practices can be up to seven years.

This is where SCRS matters. If your practice asks about client balances through a public Ai tool, you are pushing named customers and tax identifiers into a third-party model. Our approach is to pseudonymise those fields before the model sees them, rehydrate them client-side so your accountant reads the real names, and log both sides to the audit. The accountant's experience is unchanged. The exposure surface looks nothing like it used to.

A scenario that prove to be fruitful

One of our Business customers is a boutique accountancy practice with around thirty clients across the UK. The partner running the practice used to spend Monday mornings working through aged-debtors reports in Xero, one client at a time, before deciding which clients needed a chase that week. After connecting Xero through Other Me, that Monday ritual is a single chat: show me every client with receivables over thirty days, grouped by overdue bracket, and flag the top five by value. The reply takes about eight seconds. The chase emails are drafted by the same assistant a minute later, with names and amounts already filled in.

The partner tells me the difference is not the time saved, although the time saved is real. The difference is that the partner now spends Monday mornings talking to clients rather than reading screens. That shift is small on paper and large in practice, and it is the kind of outcome I tend to focus on when I think about whether Ai is adding to a practice or just sitting on top of it.

Where this sits in the bigger picture

Ai in accounting is going to go through a phase where the loudest vendors promise automation of everything, including the bits that really should not be automated. My view is that the teams who come out ahead are the ones who are honest about what should stay manual, instrument the read-side properly so the Ai can answer questions in seconds, and demand an audit trail that would satisfy their professional body without having to be assembled by hand.

I believe read-only Xero is not a limitation. It is a deliberate choice about where Ai belongs, and it is the foundation on which write-back features can be added later with proper segregation-of-duties controls. If your practice needs a specific write action, tell us what it is and we will consider it with the controls the profession actually requires. Until then, the answer you get in English, with the audit trail attached, is the useful thing.