You are not asking your accountant for information. You are buying judgment from someone who is licensed, who bills by the hour, and who carries professional liability for the answer they give you. That changes what a good question looks like.
It particularly changes what a bad one costs. Ask we think this should be handled the way we handled it last year — can you confirm? and you will very likely get "yes, that's fine." Fast, cheap, and nearly worthless: you bought agreement with your framing rather than their judgment. They answered the question you asked. The one you needed answered — whether the frame was right at all — was never on the table, because nothing in the message invited them to look at it.
That is the trap worth understanding, because an AI session makes it much easier to fall into. An hour of reasoning produces a conclusion that feels solid, precisely because it is internally consistent and every alternative got discarded along the way. Confidence is what the process manufactures. It is not evidence the premise was sound.
The obvious correction — forward everything, let them see the full picture — just converts the problem into billable hours: you are paying a professional rate to have your own situation reconstructed for you, and the context that made sense in your session doesn't survive the trip anyway.
Send the facts and your reading of them, marked as different things
Write two things and keep them apart.
The facts. Dates, amounts, which entity, which jurisdiction, what documents you actually hold. Things you could point at if challenged. If you cannot source something — a number you remember rather than one you can produce — it still belongs here, said plainly as that. Half-remembered figures quietly promoted to established fact are how a clean answer turns out to be the answer to a different question.
Your reading. The interpretation you and your assistant landed on, and the steps that got you there.
The split matters because it lets a professional work in the order they would work anyway: establish the facts, then apply what they know to them. Braided into prose, they have to unpick the braid first, and the unpicking is the part you are paying for.
Ask them to check the reasoning, not the answer
"Is this right?" is a yes/no question, and yes/no questions get yes/no answers. It also puts the entire burden on their willingness to volunteer that you asked the wrong thing — which busy people answering quickly often do not.
Show the chain instead: because A, we assumed B; because B, we concluded C. Which link doesn't hold? Now there is something specific to disagree with, and the disagreement is what you came for. It also protects you from the most expensive kind of agreement, which is agreement with a premise nobody examined.
Where you can, ask the open version rather than the closed one — what treatment applies here, not confirm that this treatment applies. Naming the shape of answer you need costs one sentence and changes what comes back.
Say which parts your AI filled in
This is the step with no pre-AI equivalent, and it saves the most time.
Your assistant's inferences do not sound like inferences. A model states a step it invented in exactly the tone it uses for a fact you handed it — fluency is not evidence of a considered position. You know which is which. Your accountant cannot tell from the prose, and if they cannot tell, they either check everything or check nothing.
So mark it: the dates and amounts come from the invoices; the claim that these two payments are the same kind of thing is my assistant's read, not something I verified. One line, and a professional knows where to aim first. It is the cheapest thing you can hand them, and it is usually where the error is.
Ask what would change the answer
Close with: what else would you need to know for this to come out differently?
That question does three jobs. It converts a one-shot answer into a short list of things you can go and verify. It surfaces facts you did not know were load-bearing. And it tells you how much weight the answer can carry, because an answer that flips on a detail nobody checked is not the same as one that holds either way.
Where the label does the work
Everything above is a discipline, which means it survives on your ability to remember it at 11pm. A briefing makes it a format instead: facts and assumptions as fields, each assumption carrying stated by sender or inferred by sender's AI, open questions kept separate from decisions. That label is exactly the flag described above, produced by the system rather than by your good intentions — and the note you write yourself earns the verbatim label only when the wording is genuinely yours. Worked examples are here.
Your session stays with you; only the reviewed briefing goes out. The full send flow is here.
What this does not do
It does not make your AI's read correct. A labeled inference is still an inference; the label tells the reader where to aim their skepticism and nothing more. An assistant's reading of your situation is a starting point for the professional to check, not a substitute for their judgment, and not something to act on because it arrived well-organized. Structure is not validation.
You should also still expect to be billed. What you stop paying for is the reconstruction. What you stop risking is a confident yes to a frame nobody examined.
If your assistant already holds the session, packaging it is the easy part — connect it and let the labels come out of the same place the reasoning did.