Ask your books: what the assistant will answer, and what it refuses
A chat surface over your own transactions that never does arithmetic and never gives advice. Here is why both limits are the point.
What it is
Ask your books is a text box on the dashboard. You type a question about your own spending. What did my Bali trip cost. How much went to software this year. Where did most of my money go in June. It answers in a sentence or two, with the figures.
Under the surface it is deliberately narrow. The language model is allowed to do exactly two things: choose which query to run, and narrate the result. It is not allowed to add anything up, and it is not allowed to tell you what to do.
Why it cannot do arithmetic
A language model asked to total forty transactions will produce a fluent, confident, occasionally wrong number. That is a property of how the models work, not a bug in any one of them. For a casual question it is annoying. For a figure you might set money aside against, it is unacceptable, because the wrong answer is indistinguishable from the right one.
So every number the assistant can say comes from a tested query in ordinary code. When you ask what a trip cost, it calls the same function the dashboard uses, which filters your transactions by country and date, sums outflows separately from inflows so a refund can never read as a purchase, excludes pending rows, and buckets by merchant and category. The model reads that result to you. If the model asks for a query that does not exist, it is refused by name rather than met with silence it might fill in.
Trips are inferred, not typed
The trip question is worth a closer look because it shows the approach. Nobody remembers the dates they travelled. So the assistant does not ask. It looks for clusters of transactions in one country, close together in time, and treats each cluster as a trip. A single foreign charge is not a trip, because that is almost always an online order from a foreign merchant. Two or more within a few days is.
The result names the country, the cities, the dates, the total, and how much of it is already marked deductible. That last figure is a filter over the classification the classifier already made. The assistant never reclassifies anything.
Why it refuses advice
Ask it whether you should elect S-corp status and it will decline and point you at the S-corp analysis on the dashboard and at a CPA. Ask whether a specific expense is deductible for you and it will do the same. Ask what you will owe next year and it will say it cannot know.
This is the section 7216 line the whole product holds. ExactSplit is an estimation and bookkeeping tool, not a tax preparer, and a chat surface is the easiest place to lose that distinction, because a fluent answer to a should I question does not feel like advice until it is. The refusals are written into the system instructions and there are automated checks that they stay there.
The disclaimer travels with the answer
Every answer carries a short disclaimer: the figures describe transactions from your connected accounts only, so cash, unlinked cards, and unpaid invoices are missing; classifications are best guesses you can override; this is not tax advice. It is attached to the response itself rather than shown once at the top of a chat window, so an answer you copy into an email keeps its caveats.
What it sends to the model
The assistant sends slices of your ledger to a third party model to narrate, which is more than the classifier sends. Because of that, it ships switched off, behind its own flag separate from the classifier's, and it stays off until a data processing agreement and zero data retention are in place with the model provider. The endpoint says so if you call it early. That is a compliance tripwire rather than a lock, and it is there so turning the feature on is a deliberate act.
Not tax advice. Estimates only. Confirm anything material with a CPA.
Run your number →