My AI finance stack, and what I dropped
The question I get more than any other is some version of: fine, but what do you actually use?
Fair question, and most answers to it are useless because they list everything the writer has an affiliate link for. So this is the short version, organized by what each thing is for, with a link to the post where I put it through real work.
There is also a section on what I stopped reaching for, which is usually the more useful half.
Current as of 29 August 2026. This page gets updated rather than rewritten, so the date at the top is the one that matters.
The whole stack on one screen
| The job | What I use | Where I put it through real work |
|---|---|---|
| Thinking with a spreadsheet open | Claude | Claude for Excel |
| Multi-step work across a folder | Claude Cowork | a full month-end close |
| Building things that run | Claude Code | a Power BI dashboard, end to end |
| Reporting and the model layer | Power BI, Power Query, Power Pivot | the month-end report build |
| External research with sources | Perplexity | Perplexity for finance |
| Whatever the company already pays for | Microsoft Copilot | Copilot for Excel, revisited |
Six things. If that looks small, it is because the list stopped growing about a year ago and started consolidating instead.
Claude, which is most of it
I am not going to pretend to be neutral here. Most of what I do runs through Claude, and the reason is narrow: it explains its reasoning before it acts, which matters more in finance than raw capability does.
In the spreadsheet
Claude for Excel sits in a sidebar next to the open workbook. The job I use it for most is reading a model I did not build, which is a surprisingly large share of the week. It edits now too, which changed what it is good for.
Across a folder
Claude Cowork is the one that handles work spanning many files. I ran an entire month-end close through it and wrote up the whole workflow, including the parts that needed correcting.
Building things
Claude Code is how I build the things that then run without me. The clearest example is an entire Power BI dashboard, and Claude Skills is how I stop re-explaining context every session, which is the single highest-return habit on this page.
Deliverables
Claude Artifacts for the small interactive tools, and Claude for PowerPoint when a board deck needs building rather than polishing.
Microsoft, because that is where the work lives
Two different things get called Copilot and they are not the same product.
Copilot for Excel I reviewed in 2024 and was not impressed. I went back to it this year and the ceiling has moved a long way, mostly because of Agent Mode and the fact that it now edits. The side pane is still the weaker half.
Copilot agents, meaning Researcher and Analyst, are a different layer and worth knowing about if your company already pays for the license. There is a monthly query cap that shapes how you use them, which almost nothing written about them mentions.
The honest framing: if your organization has already bought Microsoft 365 Copilot, use it and use it properly. If you are choosing from scratch with your own money, the decision is more open than the marketing suggests, and I ran that comparison in Claude for Excel versus Copilot.
The reporting layer, which is not AI at all
This is the part people skip, and skipping it is why their AI experiments do not stick.
Power Query, Power Pivot and Power BI are still the backbone. They are deterministic, auditable, and they run on a schedule without asking anyone. An AI that cleans a file brilliantly on Tuesday and slightly differently on Wednesday cannot carry a monthly report; a query can.
The rule I use: if the mess arrives on a schedule, it gets a query. If it is a one-off, it gets a prompt. Power Query data cleaning has the long version of that argument, and automating a month-end report is the full pipeline.
The other half of the stack
The prompts and templates I run on all of it
The tools are half of it. The prompts and templates are the other half, and they are the part that transfers between tools. Free.
Research, which needed its own tool
Perplexity is in the stack for one reason: it cites its sources, and I can click them. For checking whether an internal trend matches what the market is doing, that is the difference between something I can put in a board pack and something I cannot.
It does not reach paywalled research. Bloomberg, PitchBook and the trade association reports stay where they are, and anyone telling you otherwise is selling something.
What I stopped reaching for
The more useful half, and the part most stack posts leave out.
Chasing every new model
I tested nine tools to find the best AI for Excel and the honest finding was that the gap between the top three mattered less than the gap between using one properly and using it casually. I have stopped re-testing every launch and started re-testing when something changes how a tool handles a multi-sheet workbook.
Anything that only works on a demo file
If it needs clean data to be impressive, it is not going to survive contact with a real export. That single filter removes most of what gets recommended.
Building where I should have bought, and buying where I should have built
The rule I landed on: if the process comes back every month, build it properly in the reporting layer. If it is a one-off, hand it to a model and move on. Getting that backwards is the most expensive mistake on this list, and I have made it in both directions.
How I would assemble this from scratch
If you are starting on Monday with nothing, in this order.
- Whatever your company already pays for. Free is the best price and the approval is already done. Usually that is Copilot.
- One general model you pay for personally. The step up from the free tier is where the useful features live, and it is the cheapest line in this stack by a distance.
- The reporting layer, properly. Power Query first. It is not AI and it is the thing that makes everything after it stick.
- A custom assistant with your context loaded. A custom GPT or a Claude Skill. This is the step almost everyone skips and it pays back fastest.
- Then, and only then, agents. The agent layer is worth it once you have a process worth automating, and useless before that.
Where to go next
AI tools for finance is the hub with the tool-by-tool comparisons, and and it sorts the field by tool. If you would rather see the same ground sorted by where the work falls in your month, that view is coming.
If you would rather start with a process than a product, finance automation is the better entry point, and the free tools list covers what to try before you spend anything.
