Finance is the role with the clearest dollar-denominated ROI from Claude and the loudest failure mode if you trust it wrong. The wins are real — month-end close in half the days, anomalies caught early, cash forecasts that update themselves. The discipline that makes them real: consistent COA rules, weekly variance review, and a CPA still owning anything tax-shaped. This page is the blueprint.
Block 60 minutes every Wednesday afternoon. Open your Finance Playbook project, point Claude at the week's transactions and the running P&L, and walk the variances. Not month-end close — that comes later. This is the forward-looking ritual: what's drifting, what's not normal, what's worth investigating before it shows up in the monthly report as a surprise.
Most small-business finance is reactive. Books get touched at month-end, problems get noticed at close, and the conversation with the owner is "we missed by X" — three weeks after the miss happened. The Variance Hour reverses that. Every Wednesday, you ask the small question — "what doesn't look normal this week?" — so you don't have to ask the big question — "why did the month go sideways?" — three weeks late.
Six rotating focus questions for the hour. Pick the one most relevant this week or rotate through.
"Walk this week's payables. Any invoice more than 20% above the 6-month average for that vendor? Any vendor we haven't paid in a year suddenly billing us? Any duplicate-looking charges?"
"What's our average days-to-pay this month vs. trailing 6 months? Which specific customers are slipping? Which are still inside terms but trending the wrong direction?"
"Run actuals against the 13-week cash forecast. Where are we diverging more than 10%? Is the divergence in collections, payments, or revenue timing — and which assumption needs updating?"
"Compare this month's gross margin to trailing 3 and 12 months. If it's down more than a point, give me the three transactions or trends most responsible — and tell me whether this is mix, pricing, or cost."
"What % of revenue is from our top customer this month vs. trailing average? Any customer that's grown more than 30% as a share without us deciding to grow them?"
"Which expense category has grown faster than revenue over the trailing 3 months? Be specific — name the accounts, the growth rate, and the transactions driving it. Flag anything that looks like silent vendor price increases."
Why this works. Small businesses don't fail because the numbers were bad — they fail because the numbers were bad and nobody looked at them in time. The Variance Hour is the cheapest insurance against that failure mode. Sixty minutes a week is a rounding error against the cost of catching a problem at quarter-end that started in week 2.
The two sections that make this project worth its weight: the categorization rules (so Claude codes consistently across the team) and the "what I'm avoiding" list (so Claude can flag your blind spots when you're working on adjacent things). Most finance projects skip both. Don't.
Download the template. Fill it in honestly — 60 to 90 minutes the first time. Upload your last 12 months of P&L, balance sheets, and current 13-week cash forecast as knowledge files. The categorization rules need to come from your actual transactions, not a template — review the trailing 3 months and write down what you've actually been doing.
Open Claude → Projects → New Project → "Finance Playbook." Paste the file, upload the docs. Now every variance walk, every close prompt, every cash brief operates with Claude knowing your COA, your normal patterns, and the open items you've been carrying.
The "what I'm avoiding in the books" section is the finance secret. Bookkeepers don't avoid bad things — they avoid boring things. The unreconciled item that's $80 off. The vendor invoice with no PO. The owner reimbursement account that hasn't been cleared in 4 months. None of these are crises individually; collectively they're how books slowly lose touch with reality. Write them down so Claude can say "you're asking about cash but you've got $14k in AR over 90 days — is that still real?" The moment it does that for the first time, the project pays for itself.
Used inside your Finance Playbook project so Claude already has your COA, normal patterns, and forecast assumptions in context. Fill in the [bracketed] specifics and copy-paste.
Pull current cash (all accounts) and produce the daily brief: 1. **Cash on hand:** total across all operating accounts, separated by account 2. **vs. yesterday:** delta in dollars and one-line explanation (collections came in / payroll ran / vendor batch paid) 3. **vs. same day last week:** delta and trend (better, worse, or flat) 4. **Today's expected movements:** payables scheduled, AR expected, payroll if applicable 5. **End-of-day projection:** rough cash by end-of-day Length: under 100 words. If anything is unusual — an account I don't typically see movement in, an unexpected large debit, a missing expected deposit — surface it explicitly. Otherwise stay quiet.
Pull this week's transactions and the running month-to-date P&L. Compare against trailing 3-month and 12-month patterns. Surface variances: 1. **AP anomalies:** any invoice > 20% above the 6-month average for that vendor; any new vendor; any duplicate-looking charges. Name vendor, amount, % deviation. 2. **AR creep:** average days-to-pay this month vs. trailing 6. Specific customers slipping (name them, days outstanding). 3. **Margin compression:** if gross margin is down more than 1 point month-to-date, the three transactions or trends most responsible. 4. **OpEx creep:** any expense category growing faster than revenue over trailing 3 months — named accounts and growth rates. 5. **The one thing I'd miss:** the variance that doesn't show up in the top categories but matters. (A small vendor's silent price increase. A category that's been stable for a year and just moved.) If anything on "What I'm avoiding in the books" is relevant to what you're seeing, name it.
Pull AR aging from QuickBooks. For every customer over 30 days past due: 1. **Customer, amount, days late, last contact date / channel** 2. **Cross-reference our AR policy** in the project — what action is due for this customer today (first reminder, firm follow-up, stop-work, write-off conversation)? 3. **Draft the appropriate email** in our voice — direct but not aggressive, names the specific invoice + amount + due date 4. **Flag relationship context** if relevant — long-standing customer, recent disputes, payment history that suggests this is unusual vs. typical Output as a table with the draft message in a column. I'll review and approve before sending. For any customer over 60 days, escalate the recommendation: are we still chasing, or is this a conversation with the owner about credit terms / write-off?
Today is day 1 of [Month] close. Walk our standard close checklist (in the project) against the current state of the books. For each item: 1. Status: ready / in progress / blocked / not started 2. If blocked: what's blocking it and who needs to act 3. If in progress: realistic completion date given the blocker (if any) 4. The 3 items most likely to slip the close timeline this month End with: - The single specific action I need to take today to keep close on track - The earliest realistic close date given current state - Any item on "What I'm avoiding in the books" that needs to be resolved before close, not after
[Month] close is done. Pull the closed P&L and compare to: budget, prior month, prior year same month, trailing 3 months. Draft the variance commentary for the owner: 1. **Revenue:** actual vs. expected. The single biggest driver of any variance, named specifically. 2. **Gross margin:** actual vs. trailing 3 months. If down > 1 point, the specific reason. 3. **OpEx by category:** any category > 15% over trailing 3-month average — what drove it. Don't list every line, only the ones that moved. 4. **Cash:** actual ending cash vs. forecast. If divergence > 10%, the cause. 5. **The one thing the owner needs to know that won't show up in the table** — a leading indicator, a customer-concentration shift, a category that's stable now but trending. Length: under 400 words. Plain English. No accounting jargon. End with the question the owner is most likely to ask and your answer to it.
Attached: this week's transactions flagged as "needs review" or uncategorized. For each transaction: 1. **Best-guess category** using our COA rules in the project. Cite the specific rule. 2. **Confidence (1-10).** If below 7, do not categorize — flag for me. 3. **What additional info would resolve it** — a receipt, a vendor description, a project allocation 4. **Pattern flag** — if this looks like a recurring transaction we should add a rule for Default to flagging when ambiguous. Wrong categorization compounds — every transaction miscoded gets multiplied by the time it sits before discovery. A flagged transaction costs me 30 seconds; a miscoded transaction costs me an hour at close.
Refresh the 13-week cash forecast using current AR, AP, payroll schedule, and recurring transactions. Output: 1. **Week-by-week cash projection** with starting balance, expected inflows, expected outflows, ending balance for each of the next 13 weeks 2. **The trough week** — when cash is at its lowest in the next 13, and what it is 3. **Assumptions that changed** — what got updated this week vs. last week, and why 4. **Risks** — any large lumpy items (annual renewals, tax payments, big customer collections) that, if they slip, materially change the forecast 5. **The "if X happens" scenarios** — what cash looks like if our biggest AR customer pays 30 days late, or if our biggest invoice this week doesn't go out Flag if our cash forecast assumptions in the project look stale based on recent actuals.
Pull trailing 12 months revenue by customer. Show me: 1. **Top 10 customers by revenue** — % of total revenue each 2. **Concentration ratio** — what % of revenue is from the top customer, top 3, top 5 3. **Trend** — how has the top-customer % moved over the trailing 6 vs. 12 months? 4. **Risk read** — anything that would constitute a concentration risk by our standards (top customer > 25%, top 3 > 50%, etc.) 5. **The customer that's growing without us noticing** — anyone whose % of revenue has grown materially without an explicit decision to expand that account End with the customer the owner should have a relationship-building conversation with this quarter — based on revenue trend, not on who's most likeable.
Pull trailing 12 months by customer or by service line (whichever segmentation makes sense for our business). Calculate gross margin per segment. Output: 1. **Margin by customer/segment**, ranked highest to lowest 2. **The 5 highest-margin** — what do they have in common? Pricing, service mix, tenure, channel? 3. **The 5 lowest-margin** — same analysis. Are they loss leaders, in-progress turnarounds, or quietly destroying value? 4. **The mix shift** — has our margin mix moved over the last 12 months? Which direction, and why? 5. **The pricing conversation that's overdue** — the segment where our margin says we're under-priced relative to value delivered End with the one segment to focus growth on next quarter (highest margin + signal we can sell more of it).
New customer or new large order: [NAME, AMOUNT, REQUESTED TERMS]. What I know about them: [CONTEXT — referral source, business type, anything we've found]. Help me decide on credit terms: 1. **Default terms per our AR policy** — what we'd typically offer 2. **The case for extending standard terms** — what supports it 3. **The case for tighter terms** (50% deposit, net 15, etc.) — what supports it 4. **The case for declining** — under what circumstances we'd walk 5. **What public info I should check** before deciding — D&B, court records, recent news, anything specific to their industry End with: what would you do, and what's the smallest concession I could ask for that protects us without souring the relationship?
I'm considering: [DECISION — equipment purchase, contractor vs. employee, S-corp election, distribution timing, capex vs. opex treatment, real estate, anything tax-shaped].
Frame the tax question for me to bring to the CPA:
1. **What's the actual tax-relevant question?** Not what I think it is — what is it really?
2. **The factors that would change the answer** (timing, classification, structure, our tax position this year)
3. **The 3 specific questions I should ask the CPA** — phrased so they can give me a real answer, not "it depends"
4. **What documents I should bring** — receipts, prior returns, projections
5. **The decision I can make without the CPA** — vs. the decision that's CPA-shaped
Critically: do NOT tell me the tax answer. Tax positions need a CPA. Help me prepare for the conversation; don't replace it. If this is a decision I should not make without a CPA in the loop, say so as the first line of your response. Pull trailing 12 months of OpEx by category. Audit for creep: 1. **Category growth ranked** — which expense categories have grown fastest over the trailing 12 months, both in absolute dollars and as a % of revenue? 2. **The silent SaaS audit** — list every recurring subscription. For each: monthly cost, last 12 months total, whether we're still using it (best guess based on category and transaction frequency). 3. **The vendor-price-creep flag** — vendors whose average invoice has gone up > 10% over the last 12 months without an explicit renegotiation. (Silent price increases are the most common form of OpEx creep.) 4. **The "we said this was temporary"** — any expense category that started as a one-time or short-term spend that's become recurring. 5. **The kill list** — three specific subscriptions, vendors, or recurring expenses I should cancel or renegotiate this quarter. Ranked by likely annual savings. Be direct. Don't pad. The goal of this prompt is to find money, not to be diplomatic.
Prompt 06 is the one that prevents the slow disaster. Miscategorized transactions are the most common quiet failure in small-business accounting — they compound silently for months and surface at tax time or audit time. Make Prompt 06 a non-negotiable step for any transaction the bookkeeper isn't 100% sure about. The 30 seconds you spend on a flag saves you the hour you'd spend untangling it at close.
The integration with the most measurable payback in finance: a wired QuickBooks connection plus the close-prep prompt suite that pre-stages reconciliations, flags ambiguous categorization, drafts accruals, and writes the variance commentary draft for the owner. Setup: ~60 minutes. First-month payback: 4–6 hours of close time saved.
In your Finance Playbook project → Connect apps → QuickBooks Online. Grant read access on all entities (transactions, accounts, customers, vendors, invoices, bills). If you're on Xero, the connection works identically; if you're on Wave or a regional system without a direct integration, CSV export at close-time works as a starter.
Read-only is doing the load-bearing work here. Claude doesn't post journal entries, doesn't reconcile bank accounts, doesn't clear receivables. It reads the books and produces drafts that you (or your bookkeeper) execute. That's the right division of labor for now.
Save this as "Month-end close — day 1" inside the project. Run it on the first business day after month-end.
Today is the first business day after month-end for [Month]. Walk our books and produce the close-prep package. **1. Reconciliation status** For every bank/credit account: reconciled through what date, what's unreconciled, what's the discrepancy. List items needing investigation, ranked by dollar amount. **2. Ambiguous categorization** Pull all transactions from the closing month flagged "needs review" or coded to a default catch-all account. Apply Prompt 06 logic — best-guess categorization with confidence score; default to flag if below 7. **3. AR and AP cleanup** - AR: any invoice over 60 days that needs an action call before close (write-off, partial recognition, customer conversation) - AP: any unrecorded bills you can find in the inbox via the email connector, or any standard recurring bills that didn't appear this month - Open items on "What I'm avoiding in the books" relevant to close **4. Accruals to record** Based on payroll schedule, recurring vendor patterns, and known commitments, draft the standard accruals (wages payable, prepaid amortization, deferred revenue if applicable). Show me each entry — I'll review before posting. **5. Variance commentary draft** Run Prompt 05 against the soft-closed numbers. Draft the owner brief. **6. The earliest realistic close date** Given the state of the above, when can we realistically close? Name the single biggest blocker and the action that unblocks it. Output as one document. I'll work through it section by section over the next 2-3 days.
Before you use this for a live close, run it against the most recent fully-closed month. Three checks:
1. Reconciliation status. Operating account: reconciled through Nov 28; $1,247 in transactions Nov 29-30 awaiting bank feed sync. Credit card: reconciled through Nov 30, clean. Savings: reconciled through Nov 27, $0 discrepancy. No reconciliation blockers for close.
2. Ambiguous categorization. 14 transactions flagged. Top 3 by amount:
3. AR and AP cleanup. AR over 60 days: Charlie Corp $14,800 (84 days) — this is on your "what I'm avoiding" list. Recommend a decision call before close: are we recognizing this revenue or writing it off? AP: 2 unrecorded bills found in inbox connector — DataCo $1,200 dated Nov 22, FreshTools $445 dated Nov 28. Drafts ready.
4. Accruals. Wages payable Nov 28-30 (~$3,400 based on schedule). Prepaid insurance Nov amortization: $850. Deferred revenue recognition for [client]: $4,200. Three draft entries attached.
5. Variance commentary draft. (Full draft in thread — 320 words. Lead with revenue beat of 4%, the gross margin compression from one specific customer mix shift, OpEx category creep on contractor spend that's worth flagging to the owner.)
6. Earliest realistic close. Day 4 of next month, assuming the Charlie Corp decision happens Day 2. The 84-day AR is the single biggest blocker.
First business day after month-end: run the master prompt, work through the package over the next 2–3 days. The reconciliations and AR/AP cleanup are the time-savings; the variance commentary draft is the quality upgrade — the owner gets a real read on the month instead of a number-dump.
After three or four monthly closes, you'll know which sections of the master prompt to tighten, which categorization rules to add, and where the project is still thin. Update the project each time. The close gets faster each month because the substrate gets better each month.
What this isn't. It isn't an autonomous close. Claude prepares; you (and your bookkeeper) post. The reason: every accrual, every reclassification, every write-off is a decision that has tax implications, audit implications, and owner-reporting implications. The right amount of automation here is "the package is ready, you make the calls." Anything beyond that is L4 — and Finance L4 has the loudest failure mode in this whole resource.
L4 in finance is where the savings get serious and the failure modes get expensive. Three sketches, ordered from lowest to highest risk. Build them in this order.
Every morning at 7am, Claude pulls cash balances, the day's expected movements, and runs the variance-walk logic from Prompt 02 against the trailing day. Posts one Slack message: cash position, today's expected net movement, and any anomaly worth a 30-second look. Lowest-risk L4 in this section — read-only, surfacing, you decide what to act on.
Cash: $84,210 across operating + savings. Down $2,840 from Monday (paid Globex invoice; flagged below).
Today's expected net: +$4,200 (Acme deposit scheduled to clear) minus $1,200 (DataCo AP scheduled). Net +$3,000 projected.
Three anomalies worth a look:
Every Monday morning, Claude refreshes the 13-week cash forecast: pulls current AR aging, AP schedule, recurring payroll, and updates the projection. Flags any assumption that's drifted from actuals (your collections-timing assumption was 31 days, last quarter's actual was 38). Posts the updated forecast to a shared doc; owner and you both have the same forward view, automatically.
Medium risk — the forecast informs cash decisions, so quality matters. The mitigation: the agent surfaces what's changed since last week, with reasoning, so a human always sees what moved before acting on it.
The most aggressive L4 in this section — and the one that crosses from observation into operational recommendation. A real working pipeline a serious finance lead can build on top of QuickBooks + bank feed integrations:
Today's AP recommendations:
Today's AR priorities: Beta Inc ($8,400, 38 days, history of paying after firm reminder — chase draft attached). Charlie Corp ($14,800, 84 days — escalate beyond chase, recommend owner conversation today).
Cash sweep: Operating cash at $79k, floor is $60k. Recommend sweeping $15k to savings, leaving $4k buffer above floor.
Anomaly flag: Adobe subscription invoice came in at $52 vs. $48 last 11 months. Silent price increase. Worth noting; not worth action.
Week to date: $312 captured via early-pay discounts. $11,400 collected via prioritized chases (Beta + 2 others). 3 categorization anomalies resolved before month-end. Net working capital impact: positive ~$5,400 vs. trailing 4-week average.
The outcomes are real and measurable. A small business with $50k–$200k monthly spend can capture $500–$2,000 per month in early-pay discounts and cash optimization that previously went unclaimed because nobody had the time to run the math daily. Multiply by 12 months — that's a real number.
The catch is significant. Finance is the role where the failure modes are most expensive. The agent recommends a delay on a critical vendor invoice and a relationship sours. It misses an anomaly and a fraudulent transaction goes through. It applies an early-pay discount logic that triggers a workflow you didn't intend. The dollars at risk per mistake are higher than in any other role in this resource.
Build this only after:
If those four conditions aren't met, this build will cost more than it saves. If they are, this build pays for itself in a single quarter.
If you're imagining any of these three right now — pause. The honest path: spend two full quarters at L2–L3 with the Variance Hour and the Month-End Close Assistant. By the end of Q2 you'll know exactly which of the three fits your business, and your CPA will know enough about your AI substrate to advise on the L4 build. When you're at that point, that's a good call to have — finance L4 is the one I'd most strongly recommend getting a second opinion on before building.
Finance has the highest stakes in this resource. Each failure mode below isn't just an annoyance — each one has cost a real small business real money. Catch them in yourself.
Claude categorizes 100 transactions, 96 are perfect, 4 are subtly wrong. The 4 wrong ones compound over a year into a P&L that misrepresents reality and a tax return that gets corrected (or audited).
Spot-check 10% of AI-categorized transactions for the first 90 days. Once the patterns stabilize and your COA rules are robust, drop to 5%. Never zero.
The brief is the briefing — it tells you what you're looking at. The decision is still yours. Owners who treat "Claude says we have $X" as the operating cash position without ever opening the bank account directly will eventually be wrong about it at the wrong moment.
Cross-check actual bank balance vs. AI-reported cash position weekly. The brief is for orientation; the bank is for truth.
"Can I deduct this?" "Is this a capex or opex?" "Should we do an S-corp election?" These look like categorization questions; they're tax questions. AI is dangerous on these because it'll give you a confident, plausible, occasionally wrong answer that costs you at filing time.
Prompt 11 enforces this rule: Claude frames the tax question for the CPA conversation, never answers it. Keep the discipline. CPAs exist for a reason; AI doesn't replace them, it makes their time more useful.
The 13-week forecast was built on assumptions that were true three quarters ago. The forecast keeps refreshing weekly, but the assumptions never update. The forecast is now confidently wrong.
Quarterly: re-validate every assumption in the project against trailing actuals. If collections-timing was 31 days when you wrote it and is 38 days now, update it. The forecast is only as good as the assumptions it sits on.
The bank feed pulls transactions automatically. Claude can read them. So you stop manually reconciling. Six months later, a duplicate transaction the bank feed didn't flag has been miscategorized for half a year, and your P&L has been off by $X every month since.
The reconciliation is the trust audit. Bank feeds are convenient; reconciliations are non-negotiable. Monthly, at minimum. Don't let AI's competence make you skip the control loop.
The L4-3 build (Cash Optimization) is the one where the dollars at risk per mistake are largest. Build it before the books are clean, before the categorization rules are real, before the forecast assumptions are validated — and the agent will confidently optimize a cash position that doesn't actually exist as represented.
The four pre-conditions in the L4 section are not suggestions. They're the difference between a build that captures real money and a build that creates a slow-motion finance incident. Walk up the ladder.
Download the template. Build the project — including the books you've been avoiding. Save the prompts. Wire QuickBooks. Run the Variance Hour. In a quarter, your monthly close will be half the days, your forecast will match reality, and you'll have caught at least one problem before it became a surprise.