The Draft AI Writes: Where Automation Ends and You Begin
AI can draft the email, summarise the file and flag the anomaly — but it shouldn't press send. Here's where to draw the line in client accounting.
There's a version of AI in practice that sounds great in a demo and terrible in real life: the one where the software talks to your client, makes the call, and tells you afterwards. That's not automation — that's abdication. Your name is still on the advice, the lodgment, and the relationship.
The useful version is quieter. AI drafts, summarises, sorts and suggests. You review, correct and commit. Get that division right and you free up hours of low-value work without ever handing over judgement. Get it wrong and you spend more time cleaning up confident nonsense than you saved.
This piece is about drawing that line deliberately — where automation ends and you begin — inside the day-to-day of client accounting.
The three jobs AI is actually good at
Most of the real wins in an accounting or bookkeeping practice fall into three buckets. None of them involve AI making a decision on your behalf.
1. Drafting the first version
A follow-up email chasing a missing bank statement. A plain-English summary of what changed in a client's BAS this quarter. The first pass of a query response. AI is genuinely fast at producing a competent draft from context you already hold — the client record, the work item, the previous thread. What it produces isn't finished. It's a starting point that saves you the blank page.
The rule: AI writes the draft, a human sends the message. Especially anything that touches a client. A draft sitting in front of a staff member costs nothing if it's wrong. The same words sent automatically to a client can cost you the relationship.
2. Summarising a pile you don't have time to read
A new client hands over two years of correspondence. A job's been sitting in your work management system with fourteen comments across three staff. Someone needs to pick it up and understand where it's at in ninety seconds, not ninety minutes. Summarisation is where AI quietly earns its keep — condensing a long thread, a document dump or a messy job history into something a person can act on.
You still read the summary. You still open the source when something matters. But the triage — the deciding what deserves your full attention — gets a lot faster.
3. Flagging the thing that looks wrong
An amount that doesn't match the prior period. A client who hasn't responded to three requests. A job that's been open for longer than jobs of its type usually take. AI is good at noticing patterns and outliers across a lot of records at once — the stuff a human would only catch if they happened to be looking at exactly the right screen at exactly the right moment.
The flag is the value. The response is yours.
Why the send button is the line
The temptation, once AI can draft a decent email, is to let it send. It feels like the logical next step. It isn't.
Client accounting runs on trust and accuracy. An AI-drafted email that gets a figure slightly wrong, misreads the tone of a stressed client, or chases something you already received doesn't just waste time — it erodes the thing you can't easily rebuild. And because the draft looked confident and correct, nobody caught it until the client did.
Keeping a human on the send button costs you seconds. It's the cheapest insurance in the practice. The right model is:
- AI produces — the draft, the summary, the flag.
- A human decides — send, edit, escalate, or ignore.
- The system records — so the decision is logged against the client and the job, not lost in someone's inbox.
This is why AI works best when it's built into the practice management software you already run the work through, rather than bolted on as a separate chatbot. The draft appears attached to the work item. The summary sits on the client record. The flag surfaces on the board where the job actually lives. Context in, action reviewed, decision recorded — all in one place.
What this looks like in a real week
Take a mid-sized bookkeeping firm running client work across boards. Here's where the line sits in practice:
- The chase email. Finye's AI drafts the follow-up to a client who hasn't returned their PBC items, pulling in exactly what's outstanding. The staff member glances at it, tweaks one line, and sends. Thirty seconds instead of five minutes — and it's still a human's judgement on tone.
- The handover summary. A job moves from bookkeeper to reviewer. Instead of reading the full comment history, the reviewer reads a generated summary of where the job stands and what's unresolved, then opens the two comments that actually matter.
- The anomaly. A recurring job is flagged as sitting well past its usual turnaround. The system doesn't reassign it or email anyone — it puts it in front of the person who owns that board to decide what's going on.
Nothing here replaces the accountant. It removes the friction around the accountant so their attention lands where it's worth the most.
The questions to ask before you automate anything
Before you let any AI feature run unattended, put it through three questions:
- What happens if it's wrong? If the answer is 'a client gets a bad message' or 'a lodgment obligation gets misread', keep a human in the loop. If it's 'someone reads a slightly clunky draft', let it run.
- Is there context it can't see? A client going through a bereavement, a fee dispute, a sensitive restructure. AI doesn't know what you know off the record. Anything relationship-sensitive stays with a person.
- Can I see what it did? If an AI action isn't logged against the client or job, you've lost visibility. Automation you can't audit isn't a time-saver — it's a liability waiting to surface.
Automation that respects the relationship
The firms that get the most out of AI in client accounting aren't the ones that automate the most. They're the ones that automate the right layer — the drafting, the summarising, the flagging — and keep human judgement exactly where the stakes are: the client, the advice, the send button.
Done that way, AI isn't a replacement for your team's expertise. It's the thing that clears the low-value work off their desk so there's room to use it. The draft is the machine's job. The decision is still yours.