Home / Guides / AI email drafts

AI email drafts for accountants: useful, and safe — if a human approves every send

Elio · practising accountant, MC&S · 6 min read · July 2026

Accountants are right to be wary of AI answering client email. A confidently wrong sentence about a deadline or a deduction, sent under your letterhead, is a professional problem — not a technology one. But "AI sends email" and "AI drafts email" are completely different systems, and the second one is quietly excellent. The difference is a human approval queue that nothing can bypass.

The design rule: nothing reaches a client without you

In our practice, flagged client emails get a drafted reply. The draft sits in an approval queue where an accountant reads it, edits it if needed, and either approves it or bins it. Approve is a deliberate click. Every action — who approved what, when, with what edits — is logged for audit. There is no mode where the system sends on its own, because we never built one.

This isn't AI with a safety feature bolted on. The approval queue is the product. It converts AI from a risk you manage into a drafting assistant whose worst possible failure is a bad draft you delete.

Why the drafts are actually good: grounding

A generic chatbot knows nothing about your clients, so it produces generic prose — and the occasional invented fact. The fix is grounding: the drafting system reads from the firm's own records before writing a word. Ours sits on a knowledge graph of the practice — clients, entities, directors and trustees, related parties, jobs, deadlines and past correspondence — refreshed nightly from practice management.

So when a client asks about a payment plan, the draft already reflects who they are, what entities they have, what was lodged, and what the firm last said to them. And when the accountant edits a draft before approving it, that correction becomes a lesson the system keeps — the voice converges on your voice.

Confidentiality: where does the client data go?

This is the question a professional firm should ask first. Pasting client details into a free public chatbot is the wrong answer — you have confidentiality obligations, and you don't control where that text ends up. The right architecture keeps the firm's knowledge base on the firm's own infrastructure: your software, your database, your audit trail, not a per-seat SaaS holding your client file. That's how we built ours, and it's the standard we'd hold any system to.

What it's worth

Email is death by a thousand cuts: fifteen minutes here, ten there, all day, mostly writing things the firm has written a hundred times before. Drafting is the slow part — reviewing a good draft takes a fraction of the time composing one does. Approval-gated drafting gives the time back while leaving every judgment call exactly where it belongs.

It's also the most requested item on our list of what accountants can actually automate — and the one where the human-approval line matters most. If you're weighing whether a system like this should be bought or built, the build-vs-buy guide covers the trade-off.