How to Build a 13-Week Cash Flow Forecast With AI

I had a client send me a cash flow spreadsheet once with a sticky note on it (physically, on her monitor) that said “DO NOT TOUCH THE FORMULAS.” The spreadsheet was four years old. It had four tabs that nobody could explain anymore. And every Monday, she spent 90 minutes manually copying bank balances into it, because that was the only way to update it without breaking something.

That spreadsheet was the cash forecast. That was how they tracked cash.

Here’s the thing: her situation is not unusual. Most finance teams run their cash position off a file somebody built years ago, maintained by whoever was unlucky enough to get assigned to it, and updated by hand every week because nobody ever went back to fix the root problem. The model usually works fine. The process around it is just expensive to maintain.

This article walks through how to build a 13-week cash flow forecast with AI using Claude Cowork on the desktop app, connected directly to your financial files, with a scheduling setup that runs the analysis every Monday automatically. No more manual rebuild. No more 90-minute Monday morning ritual.

https://www.youtube.com/watch?v=3wN5c0yDltg

The Problem With Cash Flow Forecasting

The forecast model is usually not the issue. Most finance teams have some version of a rolling cash view: opening balance, inflows, outflows, closing balance. Some have 4-week views, some have 13, some have 26. The structure is fine.

What breaks every week is the input collection. Someone has to pull the bank feeds, reformat them, and paste numbers into the right cells. Someone has to update the actuals, check the budget assumptions, and send the summary before the 9am leadership call. If that person is out, the process stops. If the bank changes its export format, the process breaks. If the file gets corrupted, everybody finds out at the worst possible time.

The insight: the fragility is in the input layer, not the model. Fix the inputs, and the rest mostly takes care of itself.

Claude Cowork attacks the input layer. You connect it directly to the folder where your financial files live, and it reads them without you copying a single number. Combined with its scheduling feature, you can have the entire analysis rebuilt and saved every Monday morning before you open your laptop.

What you need before you start

Files and folder structure

File	Contents	Why you need it
Bank data (Day 1)	Balances and transactions by account and location	Current cash position
Bank data (Day 2)	Same structure, next day	Day-over-day movement and flags
P&L file	Monthly actuals vs. budget by location and account	Revenue and cost patterns for the forecast

For the walkthrough in this article, I’m using the F9 Finance Coffee Shop dataset — a 3-location coffee shop business with actual transaction data and 6 months of P&L history. The principles apply to any business with similar data structure.

Put all files in a single folder. Consistent naming matters — if the bank exports files with different names each week, the scheduled task won’t know what to read. Use a naming convention you can stick to (e.g., bank-day1-YYYY-MM-DD.xlsx) and update the folder each week before Monday.

Setting up Claude Cowork

1. Open the Claude desktop app.
2. Click the Cowork option in the navigation. This is not the same as the browser-based Claude — Cowork has direct access to files on your computer.
3. Click “Select folder” and navigate to the folder containing your financial files. Select it.
4. Cowork now has live read access to everything in that folder. No uploads, no copy-paste, no file size limits to worry about.
5. Verify the connection by asking: “What files are in my connected folder?” It should list everything you just added.

One thing worth noting: Cowork reads the files, it doesn’t move or modify them. Your originals stay exactly where they are.

Step 1 — Read the bank feed and build your cash position

With the folder connected, this is your first prompt. It does three things at once: summarizes positions, categorizes transactions, and flags anything unusual.

Prompt:
I've connected a folder with financial files for a 3-location business.
There are two bank data files: one for Day 1 and one for Day 2. Each has
a Balances sheet and a Transactions sheet.

Please: (1) summarize the opening and closing cash position for each
location across both days, (2) categorize all transactions by type —
Sales, COGS, Opex, Cash Mgmt — and (3) flag any balance issues I should
be aware of.

What the output looks like

Cowork will return a structured summary with:

• Opening and closing balance by location and account for each day
• A transaction categorization table showing dollar volume by type
• Any flagged issues it identified without being prompted

In the Coffee Shop dataset, Cowork flags the Astoria Operating Checking account going negative on Day 2 — a $41,093 overdraft. That flag appears in the summary automatically. No additional prompt needed to surface it.

From there, you can drill in:

Follow-up prompt:
What drove the Astoria checking account into overdraft on Day 2?
Walk me through the specific transactions.

Cowork traces back through the Day 2 transaction data and identifies the exact line items that pushed the balance negative — rent payments, supply orders, and the timing of merchant settlement credits that hadn’t cleared yet.

Case study — how AI caught an overdraft a manual process would have missed

The situation: The Coffee Shop business has three locations running separate operating accounts, plus merchant settlement, payroll, and reserve accounts at each location. In a typical week, a finance team member would pull a consolidated balance report from the bank portal and paste the numbers into the weekly tracker.

What changed: Both days of bank data were connected to Cowork in one folder. A single three-part prompt ran across all six location/account combinations simultaneously.

The result: Astoria Operating Checking went from $185,400 on Day 1 to negative $41,093 on Day 2. The AI flagged it in the summary output — not because it was told to look for overdrafts specifically, but because a negative balance is anomalous data it surfaced automatically. The follow-up prompt identified the combination of a $394 rent debit, a $2,991 supply order, and delayed merchant settlement credits as the cause.

In a manual process, this gets caught when the bank calls or when someone happens to open the account in the bank portal. With Cowork reading the feed, it gets caught Monday morning in the first prompt.

For finance teams managing multiple entities, subsidiaries, or locations, the math on missed flags scales quickly. If each location runs a meaningful cash risk and you’re checking manually once a week, you’re exposed for most of the week.

Step 2 — Layer in P&L history to find your cash drivers

Once the bank position is established, bring in the P&L. This is where you stop looking at cash as a snapshot and start understanding what generates it.

Prompt:
There's also a P&L file in the connected folder with 6 months of actuals
and budget for Sales Revenue, COGS, Labor, Marketing, Rent, and Utilities
across all three locations.

Please: (1) identify which location has the strongest and weakest
operating cash generation based on the actuals, (2) calculate average
monthly net cash per location using actuals only, and (3) flag any months
where a location ran significantly over budget on a cash-impacting line.

Reading the cash driver output

Cowork builds the operating cash bridge from revenue down to the cash-impacting costs. In the Coffee Shop dataset, Astoria consistently leads on cash generation per dollar of revenue — strong sales, stable labor, and lower relative rent than Hell’s Kitchen. Lower Manhattan lags, partly because of labor overruns in several months (staff turnover and training costs hit in May and June).

The stress test makes this actionable:

Follow-up prompt:
Based on the 6-month actuals, what happens to consolidated cash if any
one location misses revenue by 10%? Which location creates the most
downside risk?

The output gives you a location-level sensitivity: a 10% revenue miss at Lower Manhattan has a smaller absolute dollar impact than the same miss at Astoria, because Astoria runs higher volume. But Lower Manhattan’s margin is already thinner, so it hits a cash floor faster.

This is the kind of analysis that usually takes an afternoon to build in Excel. Here it’s two prompts.

Step 3 — Build the 13-week forecast

With the cash position and the driver analysis in hand, you’re ready to run the forward view.

Prompt:
Using the bank data and P&L history we've reviewed, build a 13-week cash
flow forecast for the consolidated business — all 3 locations combined.

Assumptions:
- Revenue trends follow the Jan-Jun 2023 seasonal pattern, adjusted flat
  from June forward
- COGS holds at the June actuals percentage of revenue
- Labor, Rent, and Utilities hold at their June actuals
- No capital raises or debt payments

Show me a base case and a downside where revenue is 15% below base.
Format it as a weekly table: opening cash, inflows, outflows, closing cash.

What the forecast table looks like

The follow-up questions that make the forecast useful

A static forecast is interesting. A forecast with a decision trigger is useful.

Follow-up 1:
In the downside scenario, at what week does consolidated cash fall below
$100,000? What's the early indicator I should track in the actuals each
week to know I'm heading toward that scenario?

Follow-up 2:
If we hit the downside revenue trend in weeks 1-3, what levers do we
have on the cost side to delay hitting the $100K threshold?

The first question gives you a specific week and a specific leading indicator to watch (typically weekly revenue per location versus the June baseline). The second question surfaces the cost levers you actually control in the near term.

Now the forecast is a management tool, not just a report.

Step 4 — Schedule it to run every Monday

This is the part that changes the process permanently.

1. With the analysis complete, send this prompt:

Set this up as a scheduled task that runs every Monday morning. Each
Monday, read the latest files in this connected folder, rebuild the cash
position from any new bank data, update the 13-week forecast using the
same assumptions, and save a summary to the folder as
'weekly-cash-summary.docx'.

2. Cowork creates the scheduled task. You’ll see it appear in the task list with the trigger (Monday), the source folder, and the output file name.
3. Confirm the schedule and close.

What the output doc contains

Each Monday, weekly-cash-summary.docx will be in your connected folder with:

• Cash position by location (opening and closing, flagged anomalies)
• Updated 13-week forecast table (base and downside)
• Any flags from the current week’s bank data

You open the doc, review the flags, and move on. The 90 minutes of data gathering is gone.

Updating assumptions mid-week

The schedule runs on the assumptions you built into the original prompt. If something changes — a location closes temporarily, you take on new debt, the revenue trend shifts — you need to update the scheduled task.

1. Open the task list in Cowork.
2. Find the Monday cash forecast task.
3. Edit the prompt — update the specific assumption that changed.
4. Save. The next Monday run picks up the new version.

It takes about 3 minutes. You don’t need to rebuild anything.

Case study — what changes after setting this up

The situation: A finance team running a multi-location business spent the first 90 minutes of every Monday gathering cash data manually: downloading bank exports, reformatting columns, pasting balances into the tracker, updating the rolling forecast, and sending the leadership summary.

What changed: Cowork was connected to the financial files folder. The 13-week forecast workflow was built in one session and scheduled to run every Monday.

The result: The summary doc is in the folder before anyone logs on. The team’s Monday morning time goes to reviewing and acting on the output — discussing the flags, adjusting assumptions, having the conversation that the data surfaces. The data gathering itself is gone.

The less obvious change: because the analysis runs every week without anyone having to remember to do it, the forecast actually stays current. In the manual version, busy weeks meant skipped updates. The scheduled version doesn’t have busy weeks.

When this approach won’t work

• Non-standard GL structures: If your chart of accounts is highly customized or changes frequently, the AI’s categorization may need correction each week.
• Inconsistent file naming: The scheduled task reads files by name. If your bank export file names change format week to week, you’ll need to rename them before the Monday run.
• Real-time data requirements: This is a file-based workflow. If your leadership needs a live cash dashboard that updates throughout the day, you need a different tool.
• Complex multi-entity consolidations: Works well for 2-5 locations with similar structures. Gets harder when entities have meaningfully different chart of accounts or intercompany eliminations.

Six steps to get started

Here are the six steps to get this running before next Monday:

1. Create the folder — one folder, all your financial files, consistent naming.
2. Open Claude Cowork in the desktop app and connect that folder.
3. Run the bank feed prompt — the 3-part prompt from Step 1 above. Review what it surfaces.
4. Run the P&L prompt — bring in your actuals, get the cash driver analysis.
5. Build the forecast — use your actual assumptions, not the Coffee Shop ones. Update the revenue trend, cost structure, and scenario parameters to match your business.
6. Schedule it — use the scheduling prompt from Step 4. Set the trigger, confirm the output file, done.

The first time through takes about an hour. Every Monday after that, you’re reading a doc, not building one.