AI use cases in finance, sorted by where they fit your month

Every list of AI use cases in finance is organized the same way, by tool. Here is what ChatGPT does, here is what Copilot does, here is a bullet about fraud detection that applies to a bank and not to you.

That is backwards. You do not have a ChatGPT problem. You have a close that takes nine days and a forecast that gets rebuilt from scratch every month.

So this is sorted by where the work falls in your month. Each one links to the place I have written it up, and each carries a verdict on whether it is worth your time yet, because a fair number of them are not.

Checked 29 August 2026.

The short version

If you read nothing else, this is the shape of it.

Where in the month	Worth doing now	Not yet
The close	Data cleanup, file consolidation, reconciliation prep	Anything that posts an entry without review
Forecasting	Building the model structure, scenario setup	The forecast judgment itself
Reporting	Commentary drafting, dashboard building, variance explanation	Numbers you have not checked
Analysis	Pattern finding, formula archaeology, market context	Causation, which it will confidently invent

The pattern is that AI is good at the assembly around the thinking and bad at the thinking. That holds across every one of these.

The close

This is where the hours are, so it is where to start. It is also the safest place, because close work is high volume with rules you can check.

Cleaning and consolidating the files

The single highest-return use case on this list, and the least interesting. Twelve exports with inconsistent headers, merged cells and three date formats, turned into one clean table. Power Query data cleaning covers the repeatable version, and automating data cleaning covers the wider case. Verdict: do this first.

Automating the month-end report itself

The full pipeline, Power Query into Power Pivot into Power BI, so the report refreshes rather than gets rebuilt. Automating a month-end report is the end-to-end build and automating the close covers the process around it. Verdict: the highest-value project on this page, and the longest.

Running the close with an agent

Newer, and further than I expected. Running a close in Claude Cowork is the full workflow, and Claude Cowork covers what the tool does generally. Verdict: real, but keep it on a copy of the folder.

Accounts payable and expenses

The most common entry point for agents in production across the industry, because the volume is high and the rules are clear. Automating accounts payable and expense management. Verdict: worth it if your volume justifies the setup.

Forecasting

Be careful here. AI is good at building the machinery of a forecast and bad at the judgment inside it, and the two get conflated constantly.

Building a cash flow forecast

A 13-week cash flow forecast with AI is the build, and the rolling version keeps it current. The structure, the scenario tabs and the flag logic are all fair game. Verdict: yes for the structure, no for the assumptions.

Scenario and stress testing

Setting up five scenarios with different multipliers is exactly the kind of mechanical work worth handing over. Deciding which five scenarios matter is not. There is a worked example in Claude for Excel on a real cash file. Verdict: yes, with the base case locked.

The forecast number itself

Verdict: no. A model will produce a confident number from bad assumptions and tell you nothing is wrong. The judgment is the job and it is still yours.

Reporting and commentary

The strongest category after the close, because most of it is a language problem rather than a numbers problem.

Writing the commentary

You have done the variance work and you know why the number moved. Turning that into three paragraphs a non-finance reader understands is a separate job. Automating Power BI commentary is the version that runs on a schedule. Verdict: the fastest win on this page.

Building the dashboard

Building a Power BI dashboard with Claude Code and the Power BI MCP are two routes to the same place. The CFO dashboard and budget dashboard builds are the finance-shaped versions. Verdict: yes, and it is more capable than most people expect.

Variance analysis

Automating Power BI variance analysis handles the calculation and the flagging. What it will not do is tell you why a variance happened, which is the part your CFO asks about. Verdict: yes for the what, no for the why.

Analysis and research

Reading a model somebody else built

Formula archaeology. You inherit a model, revenue looks right, and you have no idea whether to trust it. Asking what a five-line formula does and where its inputs come from is the use case I reach for most often. It is in Claude for Excel with a worked example. Verdict: yes, and it changes nothing in the file.

Finding the pattern you did not go looking for

Open the file, ask what is happening in this business, and let it orient before you direct it. That first pass surfaces things your own question would have skipped past. Copilot agents has the full sequence. Verdict: yes, and pressure-test whatever it flags.

External market context

Checking whether your internal trend is a you problem or an industry problem. Perplexity for finance is the tool built for this, because it cites sources. Paywalled research stays out of reach. Verdict: yes, with a specific number carried into the question.

Building the small tool you keep needing

A tracker, a calculator, a one-page summary app your team opens without being asked. Claude Artifacts covers the build and the sharing. Verdict: underrated, and the setup cost is minutes.

The ones I would leave alone for now

Being specific about what does not work is more useful than another bullet about what might.

• Anything that posts without review. The agent removes the typing, not the accountability, and no production example I can find skips the human step.
• Causation. It will tell you labor costs rose. Ask why and you will get a confident, plausible, unverified answer that reads exactly like the real thing.
• Audit-critical calculations. Both vendors say so themselves. Anthropic lists audit-critical calculations without verification as a case Claude for Excel is not recommended for, and Microsoft’s Copilot FAQ states plainly that it can make mistakes and misinterpret information. That is unusual candor from a vendor and worth taking at face value.
• Work on files you did not create. A downloaded template or a vendor workbook can carry instructions aimed at the model rather than at you. Anthropic documents this as a live risk and says its own testing found cases where it worked, so it is not theoretical.

Where to start if you only do one

Take the process that comes back every month, produces a deliverable somebody is waiting for, and currently involves you doing the same twenty minutes of cleanup before you can think. That is almost always the file consolidation at the front of the close.

Fix that one, measure what it saved, then pick the next. Finance automation is the hub for the process-shaped view of all this, and AI in finance is the wider map if you want to see how the pieces fit together.