Module S2 — Client-comm assistant in your voice

Capture your writing voice once, draft the seven recurring client communications from templates with placeholders, and keep client data out of the prompt. The voice-matched client-comm assistant built on a tool you already have.

AI Tax Practitioner Editorial

Published May 28, 2026 ·16 min read ·Last reviewed May 28, 2026

The 30-second version. The single most-wanted AI feature in tax practice is a client-comm assistant that drafts in the practitioner’s own voice. You don’t need a dedicated product to get it. A 20-minute one-time voice-capture, a prompt library for the seven recurring communication types, and a template-first pattern that keeps client tax-return information out of the prompt — together that produces drafts in your voice in 2-5 minutes. The discipline that matters: draft the template with placeholders, fill the client specifics manually. The AI writes the letter; you insert the facts. That pattern keeps the §7216 surface minimal and keeps the voice yours.

Why this module exists

During peak season, a solo CPA handles 80-150 emails per day. A meaningful share are recurring client communications: IRS notice responses, missing-document chases, engagement letters, fee-change notices, extension notifications, year-end planning outreach, refund-delay explanations. Each is a small writing task that, multiplied across a client book, consumes hours per week.

Practitioners already reach for AI to draft these. Two things go wrong:

The voice flattens. Generic AI output reads like generic AI output — over-formal, vendor-cheerful, hedge-heavy in the wrong places. A client who’s worked with you for ten years notices when a letter suddenly doesn’t sound like you. The trust cost is real and quiet.
The §7216 surface balloons. Pasting an IRS notice (which contains tax-return information) into a consumer LLM is a §7216 disclosure. Practitioners doing this without the Module S1 setup are in violation.

This module solves both: capture your voice once, draft from templates with placeholders, fill client specifics manually. The AI produces a letter that sounds like you; the client’s data stays out of the prompt.

The Monday-morning sequence

One-time setup (20 minutes)

Step 0 — Capture your voice

The voice-capture is what separates this workflow from generic AI drafting. You feed the AI 3-5 of your own past client communications (with client specifics redacted), and the AI extracts a reusable voice profile you’ll prepend to every future drafting prompt. Use Artifact 1 — Voice-capture prompt below.

The output is a “voice profile” — a paragraph describing your tone, sentence rhythm, salutation/sign-off conventions, hedging patterns, formality level, and characteristic phrases. Save it. It becomes the first thing you paste into every client-comm drafting session. You do this once; refresh annually.

Per-communication (2-5 minutes each)

Step 1 — Pick the communication type + its template prompt (30 seconds)

Seven recurring types, each with a tuned prompt. Use Artifact 2 — Client-comm prompt library below: IRS notice response, missing-document chase, engagement letter cover note, fee-increase notification, extension notification, year-end planning outreach, refund-delay explanation.

Step 2 — Draft the template with placeholders (1-2 minutes)

Paste your voice profile + the chosen template prompt. Do NOT paste the client’s actual notice or actual tax-return information. The prompt produces a letter with [PLACEHOLDER] markers where client specifics go. The AI drafts the structure and voice; you fill the facts manually. The AI never sees the client’s SSN, the specific dollar amounts, the actual notice number, or the client’s identity.

Step 3 — Fill the placeholders manually (1-2 minutes)

You drop in the client’s name, the notice number, the dollar amounts, the dates — by hand, in your own document. For communications where the AI genuinely needs some client context, use the sanitized-context variant in Artifact 3 — but the default is template-first, placeholders-manual.

Step 4 — Run the tone-calibration check (30 seconds)

Before sending, scan against Artifact 4 — Tone-calibration checklist below. Did the AI flatten your voice? Inject vendor-cheerful language? Over-formalize? Add hedges you wouldn’t use?

Step 5 — Log if tax-return information entered the prompt (30 seconds, per S1)

If you used the sanitized-context variant and any tax-return information entered the prompt, log per Module S1’s audit-trail template. If you used pure template-first (no client data in the prompt), no log entry needed — the §7216 surface was zero. This is the advantage of template-first: most client comms drafted this way generate no §7216 disclosure at all.

Artifact 1 — Voice-capture prompt

Run once. Save the output as your reusable voice profile.

I'm a US tax practitioner. I want you to analyze my writing voice from sample
client communications so you can draft future communications that sound like
me, not like generic AI.

Below are [3-5] past client communications I've written. I've redacted client
names, SSNs, dollar amounts, and identifying details — focus on the voice, not
the facts.

[PASTE 3-5 REDACTED PAST COMMUNICATIONS]

Analyze these and produce a "voice profile" I can reuse. The profile should capture:
1. Tone (warm / formal / direct / reassuring / matter-of-fact — be specific)
2. Sentence rhythm (short and clipped / longer and explanatory / mixed)
3. Salutation and sign-off conventions
4. Level of formality (first names, contractions, technical terms with or without explanation)
5. Characteristic phrases or constructions I reuse
6. How I deliver bad news (penalties, fee increases, delays) — directly or cushioned
7. How I handle technical tax concepts with clients (explain fully / reference briefly / assume knowledge)
8. What I do NOT do (no exclamation points / no jargon / no false urgency)

Output the voice profile as a single reusable paragraph I can prepend to future
drafting prompts, plus a short "do NOT" list. Keep the profile under 250 words.

Notes: redact the samples before pasting. Pick samples that represent the range (one IRS notice response, one client-relationship email, one bad-news communication). Save the output and bind it to a text-expansion snippet (::myvoice). Refresh annually or when your style shifts. The voice profile is the asset — once you have it, every draft prompt starts with “Here’s my voice profile: [paste]. Draft a [type] in this voice.”

Artifact 2 — Client-comm prompt library

Seven prompts. Each starts with “Using my voice profile [paste], draft a [type]…”

Prompt 1 — IRS notice response (template-first):

Using my voice profile [paste profile], draft a client-facing cover letter
explaining an IRS notice and the response plan. Use placeholders for all client
specifics. The structure:
- Open by acknowledging the client received the notice and reassuring them this is handle-able
- Explain in plain language what the notice [NOTICE_TYPE] generally means
- State the response plan and timeline
- Specify what you need from the client: [DOCS_NEEDED placeholder]
- Close in my voice
Do not invent dollar amounts, dates, or notice specifics — use [PLACEHOLDER]
markers for all of those.

Prompt 2 — Missing-document chase:

Using my voice profile [paste profile], draft a friendly missing-document
follow-up email. Use placeholders for the specific documents. The structure:
- Brief warm opener (not pushy)
- List of outstanding items: [MISSING_DOCS placeholder]
- The why (so we can finish your return / meet the deadline)
- An easy next step (upload to portal / reply)
- Close in my voice
No false urgency unless a real deadline is named via [DEADLINE placeholder].
Keep it short.

Prompt 3 — Engagement letter cover note:

Using my voice profile [paste profile], draft a cover note that accompanies an
engagement letter. The structure:
- Warm reconnect (returning) or welcome (new) — via [CLIENT_TYPE placeholder]
- One sentence on what the engagement letter covers
- The signature ask
- Any Section 7216 AI-consent note if applicable (reference the separate consent doc per Module S1 / the consent generator tool)
- Close in my voice
Do not draft the engagement letter itself — just the cover note.

Prompt 4 — Fee-increase notification:

Using my voice profile [paste profile], draft a fee-increase notification. This
is bad-news communication — match how my voice profile says I deliver bad news.
The structure:
- The opener (per my bad-news delivery style)
- The fee change: [OLD_FEE] to [NEW_FEE placeholder], effective [DATE placeholder]
- The why (brief, honest — via [REASON placeholder]) without over-apologizing
- The value reminder (brief, not salesy)
- Close in my voice
Do not over-justify or grovel.

Prompt 5 — Extension notification:

Using my voice profile [paste profile], draft an extension notification. The structure:
- Reassure: an extension is routine, not a problem
- State what's extended and the new deadline: [DEADLINE placeholder]
- Clarify the payment-vs-filing distinction if a balance is due, via [BALANCE_DUE placeholder]
- The next step
- Close in my voice
Match the matter-of-fact register from my voice profile.

Prompt 6 — Year-end planning outreach:

Using my voice profile [paste profile], draft a year-end planning outreach
email (sent November-December). The structure:
- Warm opener appropriate to the season
- The value proposition: a brief planning conversation before year-end can [GENERAL_BENEFIT — not specific tax advice]
- 2-3 general planning topics worth discussing via [TOPICS placeholder — e.g. retirement contributions, Roth conversion timing, the new $40K SALT cap, QSBS timing under the OBBBA tiered structure]
- The ask: schedule a call by [DATE placeholder]
- Close in my voice
Do not give specific tax advice — frame topics as "worth discussing," not "you should do X."

Prompt 7 — Refund-delay explanation:

Using my voice profile [paste profile], draft a refund-delay explanation. This
is a "manage expectations" communication. The structure:
- Acknowledge the wait without over-apologizing (the delay is usually the IRS, not you)
- Explain the general cause: [CAUSE placeholder]
- What you've done / will do
- What they can do (check Where's My Refund, or wait)
- Realistic timeline: [TIMELINE placeholder]
- Close in my voice
Do not promise a date you can't control. Keep it honest.

Every prompt produces a placeholder template, not a finished letter. Bind each to a text-expansion snippet. The library extends naturally — add prompts for the communications your practice sends repeatedly (audit-representation engagement, amended-return explanation, estimated-payment reminder). The pattern (voice profile + structure + placeholders + voice close) is reusable.

Artifact 3 — The §7216 sanitized-context variant

For communications where the AI genuinely needs some client context — a nuanced IRS notice response — use this variant. It requires the Module S1 §7216 consent + DPA on file.

Sanitization preface (prepend before pasting any client context):

The following client context has been sanitized: the taxpayer name is masked to
initials, the SSN is masked, and identifying details are removed. Use it only to
inform the tone and substance of the draft. Do not reproduce any specific
identifying detail in your output — use [PLACEHOLDER] markers for any name, SSN,
account number, or address.

[PASTE SANITIZED CONTEXT — mask name to initials, mask SSN, drop account numbers
+ addresses, KEEP the substantive facts the draft needs]

When to use template-first (default) vs sanitized-context (exception):

Communication	Default	Why
Missing-doc chase	Template-first	The missing docs are a list; no nuanced context needed
Engagement cover	Template-first	Standard terms; no tax-return info needed
Fee notification	Template-first	The numbers are placeholders you fill manually
Extension notice	Template-first	Routine; no nuance
Year-end outreach	Template-first	General topics, not client-specific advice
Refund-delay	Template-first	Cause + timeline are placeholders
IRS notice response	Sometimes sanitized-context	A complex CP2000 response may need the AI to understand the specific issue. Sanitize first; S1 consent + DPA required.

The principle: default to template-first (zero §7216 surface). Use sanitized-context only when the draft genuinely needs client facts to be good — and only with S1’s consent + DPA on file.

Artifact 4 — Tone-calibration checklist

Run before sending. 30 seconds. Catches the four ways AI flattens your voice.

                TONE-CALIBRATION CHECK (before sending)

[ ] VOICE MATCH — Does this sound like me? Read it aloud. If it sounds like a
    generic professional letter rather than how I write, the voice profile
    didn't take. Re-prompt: "match my voice profile more closely — too generic."

[ ] NO VENDOR-CHEERFUL — Did the AI inject upbeat-marketing language? Cut:
    "We're excited to...", "Great news!", "We're thrilled...", "Don't hesitate
    to reach out!", "We value your business!"

[ ] NO OVER-FORMALIZATION — Did the AI make it stiffer than I'd write? Cut
    unless I write this way: "Please be advised that...", "Pursuant to...",
    "Kindly note...", "It has come to our attention..."

[ ] NO WRONG-PLACE HEDGES — Did the AI add hedges where I'd be direct, or
    remove them where I'd be careful? Tax comms need hedges in the right places
    (uncertain outcomes) and directness in others (deadlines, fees).

[ ] PLACEHOLDERS ALL FILLED — No [PLACEHOLDER] markers in the sent version.

[ ] FACTS VERIFIED — The amounts, dates, notice numbers I filled in manually
    are correct.

[ ] SECTION 7216 CHECK — If any client tax-return info went into the prompt
    (sanitized-context variant), S1 consent + DPA on file, logged per S1.

The most-common failure: vendor-cheerful injection. Generic AI defaults to upbeat-marketing register (“We’re excited to help you with your IRS notice!”). Tax clients facing an IRS notice are anxious, not excited. The vendor-cheerful default is the single biggest voice-flattening pattern.

Artifact 5 — Before/after artifact

A real anonymized demonstration of the lift + the voice-preservation.

Scenario: missing-document chase, mid-March, returning client.

Before — raw ChatGPT (no voice profile):

Subject: Outstanding Documents Required for Your Tax Return

Dear Valued Client,

We hope this email finds you well! We're excited to continue working on your
2025 tax return. However, we've noticed that some documents are still
outstanding. In order to provide you with the best possible service and ensure
timely filing, we kindly request that you submit the following items at your
earliest convenience:
- Form 1099-INT
- Form 1098 (Mortgage Interest)
- Charitable contribution receipts
Please don't hesitate to reach out if you have any questions! We truly value
your business and look forward to serving you.

Warm regards, [Firm]

The problems: “Valued Client” (you know their name), “excited,” “hope this email finds you well,” “at your earliest convenience,” “don’t hesitate to reach out,” “truly value your business” — every vendor-cheerful tell. A ten-year client reads this and knows it wasn’t written by their preparer.

After — same task, voice profile + template-first (Prompt 2):

Subject: Three items left for your return — [Client first name]

Hi [Client first name],

Almost there on your 2025 return. Still waiting on three things:
- 1099-INT (looks like the [Bank placeholder] account — it was on last year's)
- 1098 mortgage interest
- Your charitable receipts, if you're itemizing again this year

Upload them to the portal whenever you get a chance. If you want to file by
April 15 rather than extend, I'd need them by [date placeholder].

[Practitioner sign-off per voice profile]

The difference: uses the client’s first name, references last year’s return (the practitioner-relationship signal), names a real deadline only where it’s real, sounds like a person who knows this client. Same task, 90 seconds, in the practitioner’s voice. The voice profile kills the vendor-cheerful default; the template-first pattern keeps it concrete; the relationship signal (“it was on last year’s”) is something the practitioner adds.

FAQ

Do I need a §7216 consent to use this module?

For pure template-first drafting (no client tax-return info in the prompt): no §7216 disclosure occurs, so no consent is technically needed for that use. But you should still have the Module S1 setup done, because the moment you use the sanitized-context variant or paste an IRS notice for the AI to interpret, the consent + DPA become required. Do S1 first, then use template-first as your default.

Won’t clients be able to tell their letters are AI-drafted?

Not if the voice profile took and you filled the placeholders with real relationship detail. The failure mode that gets noticed is generic AI output. The voice profile + template-first + your manual personalization produces a letter that sounds like you.

Is this different from a product like Karbon AI’s tone-matching?

Karbon AI does voice-matching natively — if you already pay for it and it’s §7216-defensible per your DPA, use it. This module is for practitioners who don’t have a dedicated product but do have a CLEAN-tier general-purpose tool (ChatGPT Team, Claude Team, M365 Copilot per the S1 decision tree). The voice-capture prompt replicates the voice-matching using a tool you already have.

How often do I refresh the voice profile?

Annually, or when your style shifts. The profile is stable — your voice doesn’t change month to month.

What about genuinely sensitive communications (audit representation, penalty abatement)?

Template-first works for the structure, but these carry higher stakes. Use the AI for the draft skeleton, then heavily edit. Never send a sensitive communication on the AI’s first draft without a careful read. The tone-calibration check is mandatory for sensitive comms.

Can I use this for non-English communications?

Yes — the CLEAN-tier tools handle Spanish, Mandarin, Korean, Vietnamese competently. Capture a separate voice profile per language. Verify the output with a native-fluency read before sending.

Module S1 — §7216-defensible AI setup — the precondition. Do S1 first; it makes the sanitized-context variant compliant.
Module S3 — Pre-signature anomaly review — the §10.22 review-trail discipline; same sanitization pattern
Module S5 — OBBBA rename trap — relevant for year-end planning outreach (Prompt 6), which touches OBBBA-affected provisions
Section 7216 AI consent: the rules and the template — the §7216 layer

If you’d rather have the client-comm assistant set up across your firm — voice profiles captured for each preparer, the prompt library customized to your recurring communications, text-expansion snippets configured, and the §7216 template-first discipline built into your workflow — Tunderman, the publisher of this site, does AI implementation for tax practices. Reach us at editorial@aitaxpractitioner.com.