The Question AI Should Answer Before Your Staff Do
Not every client query needs a person. Here's how to decide which questions AI can draft or resolve in your practice — and which ones it should never touch.
Every accounting practice runs on questions. "Has my BAS been lodged?" "What documents do you still need from me?" "Can you resend the invoice?" "When's my company return due?" Individually, none of these takes long to answer. Collectively, they eat the hours you meant to spend on actual client accounting work.
The instinct is to hire someone to field them, or to accept that answering client questions is the job. But a growing share of these questions have a single correct answer that already exists somewhere in your systems. That's exactly where AI belongs — not as a gimmick bolted onto your account practice management software, but as a layer that answers the routine so your people handle the judgement.
Sort your questions before you automate anything
The mistake firms make is asking "where can we use AI?" The better question is "which of our incoming questions have a knowable answer?" Sort a week's worth of client and internal queries into three buckets.
1. Facts your system already knows
These are lookups. The answer is a status, a date, a number, or a document that lives in your client accounting software right now.
- "Is my March BAS done?" — a work item status.
- "When's my tax return due?" — a tracked obligation.
- "What are you still waiting on from me?" — an open PBC list.
- "Can I get a copy of my last invoice?" — a record in the system.
None of these need a qualified accountant. They need a fast, accurate lookup — and if a client can self-serve through a portal, they don't need to become a question at all.
2. Answers a person must approve, but not compose
These have a knowable answer, but the wording, tone or context matters enough that someone should read it before it leaves the building.
- Explaining why a fee is what it is.
- Chasing a document for the third time without sounding annoyed.
- Summarising what happened in a job so the client understands the delay.
Here AI drafts and a human approves. The draft removes the blank-page tax — the staff member edits and sends instead of writing from scratch.
3. Judgement calls that are the actual work
These are the ones you're paid for. Should this expense be deductible? Is this the right structure? What's the tax position on this transaction? AI has no business answering these autonomously, and anyone selling you "tax return software" that promises to make the technical call for you is selling you risk. Keep these firmly with your people.
Once you've sorted your questions this way, automation stops being a vague ambition and becomes a specific decision per bucket.
Bucket one: answer it before it's asked
The cheapest question to answer is the one that never gets sent. Most bucket-one queries exist because the client can't see the thing themselves. When status, deadlines, outstanding requests and invoices live behind a portal the client actually uses, a large slice of "just checking in" traffic disappears.
What's left — the client who emails anyway — is where an AI triage layer earns its place. If a query maps cleanly to a fact in the system, it can be answered or routed instantly. In Finye, because client records, work items, obligations and documents already sit in one place, the AI has the context to answer "where's my return up to?" without a person opening five tabs to find out.
Bucket two: draft, don't delegate the judgement
This is where AI quietly saves the most time without introducing risk, because a human still hits send. The value isn't a robot replacing your team — it's removing the friction that makes routine messages pile up.
Think about the messages your staff put off: the fourth document chase, the awkward "your job is delayed because we're waiting on you" note, the recap after a messy quarter. They're not hard. They're just tedious enough to slide to tomorrow. A drafted starting point turns a ten-minute task into a one-minute one, and the message goes today instead of next week.
Finye's built-in AI drafts client communications from the actual state of the job — who owes what, which deadline is looming, what's already been sent — so the draft is grounded in reality rather than a generic template. Your team reviews, adjusts the tone, and sends. The judgement stays human; the typing doesn't.
The guardrail: AI reads your system, it doesn't invent
The single most important rule when adding AI to accounting client management software is that it should draw from your records, not fabricate around them. An AI that guesses a lodgement date is worse than no AI at all. One that reports the tracked status of a real work item is genuinely useful.
Practically, that means:
- Keep it grounded in your data. AI answers should reference actual client records, obligations and work item statuses — not general knowledge that may be wrong for this client.
- Approve anything that goes to a client. Drafts are a starting point. A person owns what's sent.
- Never let it make the technical call. Bucket three stays with qualified staff, full stop.
- Log what it does. If AI answered a status question or drafted a chase, you want a record of it.
What this actually changes in a practice
The point of sorting questions isn't to fire anyone. It's to move your people up the value chain. When bucket one is handled by the portal and triage, and bucket two arrives as a ready-to-edit draft, your team spends its day on bucket three — the work clients are actually paying for.
The compounding effect matters as you grow. A practice that answers routine questions manually hits a capacity ceiling and hires to break through it. A practice that lets its client accounting software field the knowable questions absorbs more clients on the same team, because volume growth doesn't translate one-to-one into more inbox time.
Start with one week of questions
You don't need an AI strategy. You need a list. Track every question your practice fields for a week — client and internal — and sort it into the three buckets. You'll almost certainly find that a big chunk is bucket one (should be self-serve or instant lookup), a meaningful chunk is bucket two (should be drafted), and only the genuinely valuable slice is bucket three.
That list is your automation plan. It tells you exactly where AI helps, where it drafts, and where it should stay well away — which is a far more useful place to start than "we should probably do something with AI."