The average partner spends 90 minutes a day replying to client emails. None of them think they do.
Most of those replies aren't judgment calls. They're status updates, document requests, scheduling questions, and answers to things the client already asked once. The hard ones get all the attention. The easy ones eat the morning.
Here's the workflow I run on my own inbox. It uses Claude, one saved Project, and a 20-minute setup. Per-email handling drops from 4–6 minutes to under 60 seconds for the replies that should never have been hand-typed in the first place.
Step 1: The triage prompt. Paste the client's email into Claude with this lead-in: "I'm a CPA. Classify this email into one of four categories — status request, document request, scheduling, or substantive tax/accounting question. Tell me whether a junior staffer can answer it with a template or whether it needs a partner. Don't draft a reply yet."
Step 2: The reply prompt. If Claude flags it template-able, follow with: "Draft a 3-sentence reply in a warm, direct, professional tone. Reference the specific detail in the client's email. Don't make up dates or dollar amounts. Don't give tax advice. Sign off as [your name]." Read it. Edit the one thing that's off. Send.
Step 3: The handoff prompt. If Claude flags it substantive, ask: "Summarize what this client is asking in one sentence, list the 2–3 pieces of information I need before I can answer, and draft a holding reply that buys me 48 hours without sounding evasive." That holding reply is the part that saves the day. Most partners stall because they don't want to commit; the holding reply commits to a timeline without committing to an answer.
The reason this works is the constraint, not the AI. Claude is a fine writer on its own. But "3 sentences, no fabricated numbers, no tax advice" is what keeps it from drifting into a beige paragraph that sounds like every other firm.
What this doesn't replace: written tax positions, anything quoting a regulation, anything to a client mid-audit. Those get a partner's full attention. The workflow exists to clear the easy 70% of the inbox so partners have time for the 30% that actually requires their license.
Anthropic's Projects feature lets you upload a firm tone document, redacted past replies, or your engagement letter template so every draft sounds like your firm. Setup takes 10 minutes. Available on Pro and Team.
Karbon's in-thread reply AI is now generally available, bundled into Team and Business plans. Worth a side-by-side with Claude before you pick a firm-standard tool. Karbon wins on workflow ties; Claude wins on rewriting.
Anthropic confirmed customer data on Team and Enterprise plans is not used to train models, and zero-day retention is available on request. That's the answer firm admins have been waiting for. Still redact SSNs and account numbers by habit.
Microsoft's Outlook 2026 feature is fine for one-liners but has zero context about your firm or your client. Don't confuse it with a real AI workflow. It's the email equivalent of smart-quote autocorrect.
The average daily time a partner at a US accounting firm spends on email, per a 2025 Karbon-commissioned time study. About two-thirds of those messages are template-able. The other third are where the firm's value actually lives.
What it is: A feature inside Claude that lets you create a persistent Project with custom instructions and reference files. Every conversation inside the Project pulls from your firm's tone, templates, and any source docs you upload.
What it does well: Keeps voice consistent across staff. If three associates and a partner all use the same Project, every draft sounds like your firm — not like three associates and a partner. Source files (style guide, approved phrases, redacted past replies) ground the output in your actual practice.
What it doesn't do well: Not a CRM. Won't pull client history, deadlines, or balances unless you paste them in. Won't flag conflicts of interest or stop you from sending a reply to the wrong thread. That's still on you.
Pricing: Claude Pro $20/user/mo or Claude Team $25/user/mo (5-user minimum). Team is the right tier for any firm that wants shared Projects across staff.
Three questions every firm admin needs answers to before turning Claude loose on client work. I get asked these every week.
Which tier is firm-safe? Claude Free and Pro can use prompts for model improvement by default. Claude Team and Enterprise do not — that's contractually committed in Anthropic's enterprise terms. If your associates are pasting client identifiers into Free or Pro accounts, you have a data-handling problem. Standardize on Team minimum for anyone touching client work, and turn off training opt-in inside the admin console.
What's safe to paste? On Team or Enterprise, customer data isn't used for training and isn't fed to the open web. The practical test: anything you'd run through your tax software is fine here. Names, SSNs, EINs, bank account numbers for 1099 work, balance data — same risk posture as Lacerte, Drake, or your client portal. Don't overcomplicate it. The one line that stays bright: authentication credentials (passwords, API tokens, MFA seeds) don't belong in any prompt log, on any platform. That's a different category from client tax data.
Who owns the prompts? Anthropic's terms grant you ownership of inputs and outputs. But if a staffer writes a great prompt on a personal Pro account and leaves, that prompt walks out with her. Set the policy now: prompt libraries live in firm-tier Claude Projects, not personal accounts. Write it into the tech-use policy next to the Outlook signature rule. The cleanup of a wild-west rollout costs more than the policy work.
Pick five emails from your sent folder this week that were under three sentences each. Paste each into Claude (Team account) with this prompt: "What category would you classify this as, and could the same reply have been generated from a template?" If three or more come back template-able, set up a Claude Project with your firm's voice document on Friday. 15 minutes to test, 20 minutes to set up if the test confirms it.
P.S. Reply with one prompt you wish Claude handled better. I read every reply.
