The AI Tools For Finance I Still Use (And 3 I Dropped)

I tested seven AI tools for finance in early 2025 and wrote up all seven. Three of them are no longer on my desk, and that turned out to be the more useful half of the exercise.

Nobody publishes the tools they stopped using. Every roundup you have read grows by one tool a quarter and never shrinks, which is how you end up with a list of fourteen things and no idea which one to open on Monday.

So this page is a short list rather than a long one. What I run today, what I dropped and why, the test a tool has to pass before it gets a seat, and links into every guide on this site for the ones that survived. Checked 28 August 2026, and I update it when something changes rather than on a schedule.

The three I run

Claude does the thinking work: models, analysis, anything that ends in a written explanation somebody senior will read. Copilot does the Microsoft work, because it is already inside the file and already covered by whatever my employer signed. ChatGPT does everything I want to try before committing, because it is the cheapest place to find out an idea does not work.

Three tools, three jobs, no overlap worth arguing about. Everything below is a longer version of those three sentences.

Claude for finance

The one I open first. Where it separates from the others is that it explains its own work in language a controller can review, which is what you need when the output lands in a board pack and somebody asks how the number was derived.

It has also spread furthest into the actual tools. There is an Excel add-in on the Pro, Max, Team and Enterprise plans, a Power BI route, a desktop agent, and a skills system you can extend yourself. The whole shelf sits under Claude for finance, and the ones worth starting with are Claude for Excel, Claude Skills for the work you repeat, and a full Cowork month-end close if you want to see how far it goes before it breaks.

The honest cost is usage. It burns through a plan faster than anything else here, so I keep it for the work that has to be right rather than the quick tidy-up.

Microsoft Copilot for finance

The one you are most likely to be handed rather than choose. That is its real advantage: it sits inside Excel, Outlook and Teams, under an agreement your legal team already signed, which removes the conversation that kills most AI pilots before they start.

It is weaker than Claude at analysis and stronger than everything at being where the work already is. The whole Microsoft shelf sits under Copilot for finance, which starts with the license question because that is where most people get stuck. The ones worth reading first are Copilot for Excel, Copilot agents and building a Copilot Studio agent.

Two things to check before you plan around it. Copilot in Power BI needs a paid Fabric capacity of F2 or higher, and Microsoft states plainly that a Power BI Pro or Premium Per User license alone is not sufficient, which rules out most finance teams as they are licensed today. And the =COPILOT worksheet function ends on 14 September 2026. Microsoft says that starting September 14, 2026, the COPILOT function will no longer be available, so anything built on the formula version has to move to the side pane.

ChatGPT for finance

The easiest one to start with and the one most finance people already have open in a tab. For drafting, for explaining something you half understand, and for finding out in four minutes whether an idea is worth an afternoon, it is hard to beat.

Start at ChatGPT for finance. From there, custom GPTs are how you stop retyping the same context every morning, ChatGPT for Excel covers the spreadsheet half, and OpenAI Agent Builder is where it gets ambitious.

The thing to watch is what you paste. A personal account is not covered by your company’s terms, and a customer-level P&L in a consumer chat window is how these programs get shut down from above.

Gemini, and why it is still a maybe

Google keeps shipping, and Gemini keeps being fine. It handles long documents well and it is genuinely good inside Google Sheets, so if your company runs on Workspace rather than Microsoft 365 it moves up the list considerably.

For a Microsoft shop it stays a second opinion rather than a seat. The coverage here is Google Gemini for finance, Gemini Notebook, which is the most interesting thing Google has shipped for this audience, and Google Sheets automation.

The three that dropped off

Llama, Nova and DeepSeek were in the original seven. All three still exist and all three are capable. None of them are on my desk, and the reasons are worth more than another feature table.

Llama is the one I most want to be wrong about. Running a model on infrastructure you control solves the data question outright, which matters enormously in this field. It also needs somebody to own that infrastructure, and in a corporate finance function that person does not exist and is not getting hired.

Nova never found a job that one of the other three was not already doing better. That is the whole review.

DeepSeek performed well on the tasks I gave it. It failed a different test, which is whether I could get it approved for company data, and for a page written for people inside corporate finance functions that is the test that decides. If you are a sole practitioner the calculation is yours to make.

Two of the three fell out for reasons that had nothing to do with output quality. That is usually how it goes.

What a tool has to do to earn a seat

I run every new one through the same four questions before it gets anywhere near a real close. It takes an afternoon and it has saved me several.

Can I put company data in it

First question, every time, and it eliminates more tools than performance ever does. If the answer needs a conversation with legal that nobody is going to start, the tool is already out no matter how well it demos.

Does it show its work

An answer I cannot check is worse than no answer, because I will present it. Give it a variance you have already worked out by hand and see whether the explanation matches the arithmetic or just sounds like it does.

Does it survive a real file

Every demo uses clean data. Hand it the export where the header row starts on row 7, the amounts are stored as text and three columns are called Unnamed. Most of the gap between tools shows up right there.

Will it still be doing this in six months

Not whether the company survives, whether the feature does. Microsoft ends a worksheet function on 14 September 2026 that people built classification workflows on. Build on the surface a vendor is investing in, not the one they announced.

The tools underneath the tools

The chat window is the part everyone talks about and the smaller part of the work. Underneath it, the Power BI MCP is what lets a model read your semantic model instead of guessing at column names, Office Scripts is what turns a good prompt into something that runs next month without you, and Python in Excel covers the statistical work none of them do well in a chat.

Two more worth your time if you are building rather than using. SQL for finance is still the language that pays back fastest for anyone who touches a data warehouse, and AI prompting for finance is the difference between a tool that works for you and one that you gave up on in week two.

If you want the free end of the shelf, the free finance automation tools I use are listed separately, and Perplexity earns a mention for research even though it never earned a seat for building.

Where to go next

Pick by the work rather than the vendor and you will get there faster. AI in Excel if the job lives in a spreadsheet, Power BI if it lives in a report, finance automation if it lives in your month-end, and AI agents for finance if you want it to happen without you. The career and policy side of all this sits under AI, careers and governance.