.avif){kind=icon}
What you'll take away
- A generic prompt fails because it never asks for anything. A finance prompt names the exact inputs it needs and instructs the model to stop and ask when something is missing, instead of filling the gap itself.
- Garbage in, garbage out is not a cliché here. Clean the data, redact it, and then paste it. Messy input produces a messy answer that looks just as polished as a good one.
- Redact anything sensitive on a personal account. If your ChatGPT or Claude account is provided by the company, you're on safer ground and can share properly.
- Treat the model like a junior accountant. It's useful, it's fast, and it will make mistakes that you're responsible for catching.
- The goal isn't to keep opening our page. Try the prompts, keep the ones that work for your close, tweak them, and build your own library in a spreadsheet.
What's in the library
Five categories: reporting, forecasting, reconciliation, policy and compliance, and vendor and contracts. It's free at alaan.com/tools, no catch, and it grows most weeks.
The prompts split into two kinds. Some just need you to fill fields, like the 13 week cash flow forecast, which spits out a template ready to populate. Others need you to paste or attach real data, like the variance and reconciliation ones. Every prompt page tells you what it's best for, exactly what you have to provide, and what the output will look like before you run anything.
The session demoed monthly variance summary, anomaly detection, the 13 week cash flow template, bank reconciliation checker, and the month-end close checklist generator. Vendor spend analysis and contract terms review didn't make it into the hour, but they're both in the library.
What a prompt looks like when you run it
The monthly variance summary asks for four things: reporting month, any short context like a one-off event or a pricing change, your materiality threshold, and the P and L table itself. Generate, then open it straight in ChatGPT with one click. Claude doesn't support that handoff yet, so it's copy and paste for now.
The first thing the answer did was state its own comparison basis: actual versus budget, with a note that no prior month or prior year data had been provided. Then an executive summary, a variance table, and a set of management actions for the following month. On the demo numbers that meant flagging marketing at 699,000 against a budget it had overshot by 19%, followed by a reforecast recommendation.
If you'd rather have it in Excel than in a chat window, just ask for it in Excel. That works for most of these.
The demo where nothing came back, which was the point
The anomaly detection prompt got a vendor transaction list pasted into it, and returned no flags at all. It said the dataset had transaction dates and amounts but no employee or cardholder identifiers, so it couldn't do what was being asked.
That's the behavior worth paying for. A vague prompt in the same situation invents a cardholder, flags them, and you find out at the wrong moment.
Where AI still gets finance wrong
- Hallucination. A missing row gets filled with invented data that looks fine, and the model will hold its ground if you question it. On a thousand-row reconciliation, spot check ten rows at random, and run the same job through two models. Matching answers are probably right. Different answers mean one of them slipped.
- Math. LLMs are not good at arithmetic and get worse as it gets complicated. Where you can do the math yourself and hand it over as an input, do that.
- UAE context. These models are trained mostly on US and European rules. Telling one that you're in the UAE doesn't give it regulations it never had, so anything compliance-shaped needs checking against the actual rules.
- Stale data. A chat won't remember last month's variance work. If you want month on month, point it explicitly at the earlier thread and ask it to compare.
Questions from the room
How do I use AI without leaking company data? Anonymize and redact by default. On a personal account, round the numbers rather than pasting exact ones, or switch the currency. On a company or enterprise account you can share properly.
Claude, ChatGPT, or Gemini? Claude is the most capable for heavy work, but it burns through usage fast enough that you'll want to save it for the big jobs. ChatGPT handles these prompts fine. Ali doesn't use Gemini for finance work.
Do I need a paid plan? Free tiers cover basic tasks. Once you're feeding in thousands of rows, expect to hit a wall.
Can I connect it to my ERP? Not natively. Export to CSV or Excel and work from there.
Will AI replace finance? His answer: AI will replace finance that doesn't use AI. It's a tool, and the useful move is learning to drive it.
The one thing to do this week
Pick three prompts, run them on fake or redacted data, and keep the ones that hold up in a spreadsheet of your own. That sheet is worth more to you in six months than any library someone else maintains.
Watch the session
The full recording is above. The library is free at alaan.com/tools, and if there's a prompt your close needs that isn't in there yet, Ali's email is in the session and requests do get built.
