# The client email desk: first drafts of the replies that eat an afternoon

Module 11 of 19 in the Paidnice AI for bookkeeping hub. Part: The six AI workflows.

How to draft the technical client replies without letting an AI state a tax position it invented. The three-marker system that separates fact from what a human has to verify, and the two kinds of email this workflow refuses to touch.

Source: https://www.paidnice.com/ai-for-bookkeeping/client-email-desk

---

**The client email desk drafts the first version of the replies that eat a partner's afternoon, so the work becomes editing rather than composing. The whole thing rests on one rule: the AI never states a technical position it worked out on its own. Every tax or accounting point is marked for a human to verify before it goes anywhere. Build this second, after the monthly query chase. It saves the most time per afternoon and needs no script at all.**

**Key takeaways**

**Three markers, not one draft.** Plain text is a fact from the client's own data. [CONFIRM] is a technical position a qualified human must verify. [NEED] is missing information.

**A technical reply with no markers has overstepped.** That is the tell that the model answered rather than drafted.

**Two kinds of email it refuses.** Complaints and fee disputes, and scope questions. Both go to a person.

**Read the whole thread.** A position the firm committed to three emails ago is binding whether or not anyone remembers it.

## Why a confident wrong answer is the real risk

Tax and accounting answers depend on jurisdiction, year, entity type and facts that are often not in the email. A model will produce a fluent, confident answer from an incomplete brief, and in writing to a client that is a professional problem rather than a typo.

The fix is not to ban AI from the inbox. It is to make the draft show its own uncertainty, so the person editing it can see in two seconds which sentences carry risk.

| Marking | Meaning | Who resolves it |
| --- | --- | --- |
| Plain text | A fact from the client's data or from earlier in the thread | Nobody, it is checkable |
| [CONFIRM] | A technical position that must be verified before sending | A qualified human |
| [NEED] | Information that is missing and has to be asked for | Whoever replies |

## Classify before drafting

Not every email should get a draft, and knowing which is most of the value.

| Type of email | What the workflow does |
| --- | --- |
| Factual about their own data | Answers from the ledger, citing the figure and its source |
| Process, "what do I do next" | Answers plainly, low risk |
| Technical, tax or accounting treatment | Drafts the shape, marks every position [CONFIRM] |
| Scope, "is this included" | Does not answer. Flags for the engagement partner |
| Complaint or fee dispute | Does not draft at all |

The last two matter more than the first three. A cheerful AI-drafted reply to a complaint is how a small problem becomes a formal one.

Free and open source

Install it: skills/client-email-desk

Drafts client replies in your firm's tone, marks every technical position for human verification, and refuses complaints, fee disputes and scope questions. No script, because nothing here is arithmetic.

Get the six workflows on GitHub

Works in Claude, Claude Code, Codex and Cursor. Copy the folder into .claude/skills/ and ask in plain language.

## Matching the firm's tone

Ask for two or three previously sent replies and match them: greeting, sign-off, sentence length, whether the firm uses first names, how it delivers bad news. A firm already has a voice, and a drafted reply that does not sound like it is a reply the partner rewrites from scratch.

## Rules for the draft itself

- Answer the question asked in the first line. Clients scroll.
- One email, one topic. Two questions get two drafts.
- Never invent a deadline, a rate or a threshold. Mark it [NEED] and ask.
- No apology for a delay unless there was one.
- Do not promise an attachment nobody has confirmed exists.

## What it cannot do

**It cannot send.** Every reply is a draft a human reads and sends, which is deliberate rather than a limitation to work around.

**It does not give tax advice.** It drafts the shape and marks what needs verifying.

**It does not decide scope.** That is an engagement partner decision and the workflow routes it there.

The wider boundary is covered in [AI for accountants](/ai-for-bookkeeping/what-ai-does-in-a-practice), and the queries this desk often ends up chasing start life in [the monthly query chase](/ai-for-bookkeeping/monthly-query-chase).

## Common questions

### Is it safe to let AI near client email?

Drafting is safe. Sending is not, and nothing in this workflow sends. The risk is not access to the inbox, it is an unreviewed technical position going out under the firm's name, which the marker system exists to prevent.

### Does it need a connection to my ledger?

Only for the emails that ask about the client's own numbers. The rest works from the thread alone.

### What about a client who emails five questions at once?

It drafts five replies and says so, rather than producing one long answer that half-covers each. One email, one topic, is the rule that keeps a thread reviewable.

---

Part of [AI for bookkeeping](https://www.paidnice.com/ai-for-bookkeeping), a free hub for bookkeepers covering the receivables process, building a practice, and the AI workflows behind both.