Greenleaf CPA

Every Monday in season your team sends 30+ follow-up emails asking clients for the same missing documents. It's not a discipline problem — it's that your inbox was never built to track who owes what.

A scheduled agent that reads your open requests, knows which client owes which document, and sends personalized SMS + email nudges on a cadence — escalating tone as the deadline nears. You see a single dashboard of who's outstanding; the chasing happens without you.

1. Schedule Trigger — weekday mornings, 8am.
2. Source of truth — a Supabase table (or a synced Google Sheet) of client → required docs → received?.
3. Claude node — drafts a short, on-brand nudge per outstanding client, escalating wording by days overdue.
4. Twilio + email node — sends SMS and email.
5. Write-back — log each touch so nobody gets chased twice in a day.

$0/mo infrastructure (n8n self-host or starter cloud). Claude API ~$10–15/mo at your volume. Twilio SMS ~$0.01 per message.

Only names, emails, and the document checklist live in your database — no client financial data. The nudge text never includes return figures. Everything stays inside infrastructure you control.

Recovers ~6 hrs/week across the team during season. SMS nudges convert roughly 3x better than email-only, so the docs land sooner and your close timeline tightens.

Your bookkeeper re-categorizes the same transactions by hand every month for a dozen of your 30 clients. The rules live in their head, so the work can't be delegated and doesn't get faster.

A Claude agent with per-client memory in your own database that pre-codes uncategorized transactions and flags only the exceptions for human review. The model gets sharper every month as each client's history grows.

1. Schedule Trigger — weekly, Monday 8am.
2. QuickBooks Online node — pull uncategorized transactions per client.
3. Claude node — classify using that client's prior coding history pulled from Supabase.
4. Supabase write — store predictions + a confidence score.
5. Review view — bookkeeper confirms low-confidence items only; confirmations feed back into memory.

$0/mo infrastructure. Claude API ~$5–15/mo at your transaction volume. No per-seat SaaS.

Transaction data stays in your Supabase. The only thing that leaves is the classification request envelope to the Claude API — no PII beyond transaction descriptions, and nothing is used for training.

Recovers ~5 hrs/week of manual categorization. Quality compounds week over week, and the codified rules mean a junior can finally own the review.

Your two associates interrupt you 6–10 times a day: "what's the IRS guidance on…", "how do we handle this client's S-corp election?" You're the bottleneck, and the answers already exist — they're just trapped in your head and old memos.

A private assistant trained on your firm's SOPs, past memos, and bookmarked IRS guidance. Associates ask it first; it answers with citations back to your source docs. You stop being a human reference book.

1. Ingest — drop your SOPs, position memos, and IRS bookmarks into a Supabase vector store (one-time load).
2. Chat surface — a simple internal page or Slack bot.
3. Claude node — retrieval-augmented answers that cite the source doc every time.
4. Feedback loop — partner thumbs-down corrects an answer and updates the knowledge base.

$0/mo infrastructure. Claude API ~$10–20/mo. Replaces the need for a paid per-seat "firm GPT" subscription.

Runs entirely on your own documents inside your workspace. No client data is required to answer process questions, and nothing is shared with a model provider for training.

Reclaims ~3 hrs/week of partner time — the highest-value hours in the firm — and your associates get unblocked instantly instead of waiting on you.