Finance & FP&A

How to build a 13-week cash flow forecast with AI

Finance & FP&A3 AI tools7 steps6 friction points

Want an agent to handle this on your live data? See the Starch version →

A 13-week cash flow forecast maps every dollar coming in and going out across the next quarter at weekly resolution. It's the standard tool for managing liquidity — not just knowing your runway, but knowing which specific week a big vendor payment hits, whether collections from a slow customer create a shortfall in week 6, and whether you have enough cushion to make payroll in week 11 if a deal slips. Most operators build this once in a spreadsheet and then slowly stop updating it.

The reason people reach for AI here is that the mechanical work is substantial but not intellectually hard. Categorizing transactions, projecting forward from historical patterns, formatting weekly buckets, writing assumptions — these feel like exactly the kind of structured, repetitive reasoning that a language model should handle. You have the raw ingredients in your bank statements and accounting system. You just need something to do the assembly work faster than you can in Excel.

ChatGPT, Claude, and Gemini can genuinely help with this workflow. They're good at structuring a forecast template, writing formulas, categorizing transaction exports, and reasoning through assumptions if you describe your business clearly. The limitation isn't intelligence — it's data access. You have to bring the numbers to them, and you have to do it every time you want a fresh view.

Skip the manual work → See the Starch version

Finance & FP&A3 AI tools7 steps6 friction points

AI walkthrough

How to do it with AI today

A practical walkthrough using ChatGPT, Claude, and other off-the-shelf LLMs — what they're good at, what you'll have to do by hand.

Tools that work for this

ChatGPTClaudeGemini

Step-by-step

1 Export your last 90 days of bank transactions from your bank portal or QuickBooks, and download any outstanding invoices and scheduled bills. You'll need these as CSV or copied text — the LLM can't pull them directly.

2 Open ChatGPT (GPT-4o) or Claude and paste in the transaction export. Ask it to categorize every line into a standard chart of accounts — payroll, software, rent, cost of goods, etc. — and flag anything ambiguous for your review. This usually takes one or two back-and-forth messages to get the categories right.

3 Paste your cleaned, categorized transactions back in and ask the model to calculate average weekly spend by category over the last 12 weeks. This becomes your baseline burn rate, broken down by cost type.

4 Separately, describe your expected inflows — recurring revenue, outstanding invoices with expected payment dates, any known one-time receipts. Paste in your accounts receivable aging if you have it. Ask the model to build a weekly inflow schedule for the next 13 weeks, flagging assumptions it's making about collection timing.

5 Ask the model to combine the inflow and outflow schedules into a single 13-week table with opening balance, weekly net cash flow, and closing balance. Specify your current bank balance as the starting point. Ask it to output this in a format you can paste directly into Google Sheets.

6 Review the output in your spreadsheet, fix any obvious errors or miscategorized items, and then bring the corrected version back to the LLM to adjust assumptions — slower growth, a delayed payment, a new hire starting in week 5. Ask for a revised version reflecting each change.

7 Save the final version as your working forecast. Set a calendar reminder to repeat this entire process in two to three weeks when the numbers have moved.

Prompts you can copy

Here is a CSV of 90 days of bank transactions. Categorize each line into: payroll, software/SaaS, rent/facilities, cost of goods sold, marketing, professional services, taxes, and other. Flag anything you're unsure about.

Based on the categorized transactions above, calculate average weekly spend for each category over the last 12 weeks. Show me the table and note any weeks that were obvious outliers.

Build a 13-week cash outflow forecast using these weekly averages. Assume week 7 includes a $14,000 quarterly insurance payment and week 10 adds one new hire at $5,800/month prorated. Start with a $240,000 opening balance.

Here are my open invoices with expected payment dates and my recurring subscription revenue of $68,000/month collected on the 1st. Build a weekly inflow schedule for the next 13 weeks. Flag any weeks where collections look thin.

Combine these inflow and outflow schedules into a single 13-week cash flow forecast table: columns for week number, date, inflows, outflows, net cash flow, and ending balance. Format it so I can paste it into Google Sheets.

Reality check

Where this gets hard

The walkthrough above works — until your numbers change, the LLM hallucinates, or you have to re-paste everything next month.

No live connection to your bank or accounting data — every refresh means a new export, a new paste, and re-running the full prompt chain from scratch.

Transaction exports frequently exceed what's practical to paste in a single context window; large datasets either get truncated or require splitting into batches you then have to manually reconcile.

Categorization quality is inconsistent across runs — the same vendor gets tagged differently depending on prompt wording, which creates category drift when you compare week-over-week.

Nothing persists. The 13-week model you built this month lives only in that conversation thread; next month you start over with no memory of last month's assumptions or structure.

Scenario changes require re-prompting the entire model, not adjusting a single input. Changing one hiring assumption means re-running and reformatting the whole output again.

The LLM has no visibility into what actually happened versus what was projected — there's no mechanism to track actuals against the forecast and update automatically as weeks close.

Tired of the friction?

Starch runs the whole workflow on live data — no copy-paste, no hallucinated numbers, no re-prompting next month.

See the Starch version →

Starch alternative

The same workflow on Starch

Starch is an agentic operating system — agents build and run persistent software against your live business data. For a 13-week cash flow forecast, that means the forecast lives as a running app connected to your actual bank and revenue accounts, not a conversation you re-run manually each month.

The Runway Analysis starter app connects Plaid and Stripe directly — Starch syncs your real balances, transactions, and revenue on a schedule, so the forecast always reflects current numbers without any export or paste.

The Scenario Analysis starter app lets you test assumptions — a new hire in week 5, revenue growing 15% slower than plan, a delayed fundraise — and see the impact on week-by-week cash immediately, against your actual baseline, not a static spreadsheet.

Transaction Insights automatically categorizes every expense from your connected Plaid accounts, flags anomalies, and tracks recurring charges — giving you a clean, current picture of your cost structure that feeds directly into the forecast.

Describe the exact forecast structure you want in plain English — weekly buckets, specific inflow categories, payroll timing, accounts receivable aging — and Starch builds the custom app. You're not constrained to a template.

Set up an automation that refreshes the forecast every Monday morning, compares actuals from the prior week against what was projected, and Slacks you a summary of any variance above a threshold you set — all without re-prompting anything.

Get closed-beta access →

Toolkit