AI Agents For Finance (What I Run, And What Broke)

Most of what gets sold as an AI agent for finance is a scheduled script with better marketing. That is not a complaint, because scheduled scripts are excellent and I run several. It is a definition problem, and it costs people months.

The useful distinction is whether the thing decides its own next step. A flow that refreshes a report every Monday at 6am follows a path you drew. An agent is given the goal and works out the path, which is more powerful and considerably easier to get wrong.

I build these weekly, in Claude and in the Microsoft stack, and the honest summary is that the ones earning their keep sit lower on the autonomy ladder than the demos suggest. So this page goes rung by rung, with what I run at each level and what broke. Checked 29 August 2026.

What counts as an agent, and what is just a script

Three rungs, and you climb them in order whether you meant to or not.

It answers

You ask, it reads something it has access to, it replies. No changes to anything. This is where every finance team should start, because a wrong answer costs you a re-read and a wrong action costs you a journal entry you have to reverse.

Boring, and it is where a surprising share of the value sits. An assistant that can read your close checklist, your policy documents and last quarter’s board pack answers half the questions that currently land in your inbox.

It acts, and you approve

It drafts the journal, builds the file, writes the email, and stops. You look and press go. Almost everything I run in a real close lives here, and it is the rung most people skip on their way to something more impressive.

The reason to stay here longer than feels necessary is that the approval step is also your audit trail. Somebody looked. That is a sentence you will want to be able to say.

It runs unattended

No human in the loop. Reserve it for work where being wrong is cheap and obvious: pulling a file, reformatting an export, posting a summary into a channel where somebody will notice if it reads like nonsense. I do not have anything touching the ledger at this level and I am not in a hurry to.

The step from rung two to rung three is a question about what you are willing to explain to an auditor. The technology is the easy half.

Claude Skills, the one worth learning first

If you only learn one thing on this page, learn Skills. A Skill is a folder with a SKILL.md file in it, and the file starts with two lines of metadata: a name, and a description of what it does and when to use it. That is the whole entry requirement.

What makes them work is progressive disclosure. Claude loads only the name and description of each Skill at startup, roughly 100 tokens each, so you can install a shelf of them without paying for any of it. The instructions load only when a request matches, and bundled files and scripts load only when the instructions reach for them.

For finance that means you can encode the way your close really runs, the way your commentary is worded, the checks that have to pass before a number goes in a deck, and have it applied without pasting the same three paragraphs into every conversation. I go through building one in Claude Skills for finance work, and Claude Code is where they get properly useful because the Skill can run scripts against real files.

Where Skills stop being a team tool

Custom Skills on claude.ai are individual to each user. Anthropic’s own documentation says they are not shared organization-wide and cannot be centrally managed by admins. So the moment you build something your team should all be using, distribution is your problem: everyone uploads it themselves, and nothing stops six people running six slightly different versions by March.

They also do not sync across surfaces. A Skill uploaded to claude.ai is not available through the API, and Claude Code keeps its own on the filesystem. Three places, three uploads, three chances to update two of them.

And treat a Skill from anywhere else the way you would treat an executable from a stranger. Anthropic’s guidance is blunt about it: a malicious Skill can direct Claude to invoke tools or run code in ways that have nothing to do with what it claims to do, and Skills that fetch content from external URLs are the risky shape. Read every file before you install it. On Enterprise there is content scanning for Skills uploaded through claude.ai and Cowork, though it does not cover the API or the Console.

None of this makes Skills a bad idea. It makes them a personal productivity tool that your governance has not caught up with, which is worth knowing before you present it as a team rollout.

The Microsoft route

If your company runs on Microsoft, this is the route that survives a security review, because it is inside an agreement your organization already signed. That advantage is worth more than any feature comparison.

Copilot Studio is where you build an agent with a defined scope, a set of knowledge sources and a set of actions it is allowed to take. It maps onto the three rungs almost exactly, which is why I wrote building a Copilot Studio agent in that order. The lighter version, agents inside the apps you already use, is in Copilot agents, and Copilot Cowork is the desktop end of it.

Underneath all of it sits Power Automate, which is the part nobody finds exciting and the part that runs. A scheduled cloud flow takes about ten minutes to set up and it is the correct answer to more finance problems than any agent. Power Automate for finance covers it properly.

The no-code route

When the work crosses systems that were never meant to talk to each other, this is where you end up. n8n is the one I reach for because you can self-host it, which changes the data conversation entirely, and building AI agents in n8n walks through a finance build step by step.

Zapier is faster to start and easier to get approved, Make sits between the two on both counts, and no-code automation for finance is the wider view if you are choosing between them. The older shape of this work, before any of it could reason, is robotic process automation, and it is worth understanding because a lot of what your company already owns is that.

On the OpenAI side, OpenAI Agent Builder is the equivalent, and custom GPTs are the rung-one version that takes an afternoon.

Where I would start on Monday

Take the thing you explain most often and make it answer that. Something smaller than the close or the forecast. The question a colleague asks you every second week that has a documented answer you keep retyping.

Build it at rung one, use it for a month, and notice how often you correct it. That number tells you whether rung two is safe. If you never correct it, promote it and let it draft. If you correct it weekly, the problem is your source documents rather than the agent, and no amount of autonomy fixes that.

The full worked example of how far this goes is a Claude Cowork month-end close, where I ran the whole thing and wrote up the parts that held and the parts that did not.

What breaks

It fails silently

A broken flow shouts. An agent that has stopped finding the source file will produce a confident, complete, entirely stale answer without a word of warning, and it will do it every week until somebody notices the numbers have not moved. Build a check that fails loudly, even if it is one line comparing this month’s row count to last month’s.

The scope creeps

It works, so you give it one more job. Six jobs later it is the thing nobody understands and nobody wants to touch before quarter end. Keep agents narrow and let there be several, the same way you would not put eleven unrelated macros in one workbook.

It has no owner

You built it on a Saturday, so it is yours, and it stays yours through the reorg and the holiday and the day you change jobs. Write down what it does and what it touches before it becomes load-bearing, because the version of this that ends badly is an agent posting to the ledger that only one person has ever understood.

Where to go next

The process these agents plug into lives in finance automation. AI in Excel and AI in Power BI cover the two surfaces most of this work lands on, AI tools for finance has the tool-by-tool verdicts, and the policy question that arrives the moment an agent touches anything real sits under AI, careers and governance.