How to set up Claude for a finance team: skills, data rules and review
Giving a finance team Claude logins and hoping for the best produces ten people writing ten different prompts and getting ten different numbers. The setup that works is closer to writing a procedures manual: a few written playbooks, a rule for where numbers come from, and someone other than the author checking the output.
What a skill is
A skill is a written playbook that Claude loads when it does a specific task. It's a document, not code, and it usually contains:
- When to use it: the task and the trigger ("monthly flux commentary for the management accounts").
- Inputs: which reports, queries or files it needs, and where they live.
- Steps: the order of work, the thresholds, the judgement calls already made.
- Output format: the template, the length, the table layout.
- Checks: what must be true before the output is handed over.
The point is consistency. When the controller and a new analyst both run the flux skill, they get commentary in the same structure, from the same data, with the same thresholds. Improving the skill improves everyone's output.
Skills that earn their place on a finance team typically include: board prep, the revenue bridge, flux commentary, pricing proposals, QBR packs, meeting prep, Excel model standards, deck standards, VAT or sales tax return preparation, and a standard for building small internal apps.
The data sourcing standard
This is the most important page you'll write. It tells Claude, for every kind of number, where it must come from and what to do if it can't get it.
| Number | Must come from | Never from |
|---|---|---|
| Actuals (P&L, balance sheet) | Closed TB, via the metric definitions | Last month's deck, an email, memory |
| Budget and forecast | The current approved version in the planning model | A draft version or a pasted table |
| Customer and pipeline data | CRM, read at the time of the task | A sales update slide |
| Cash | Reconciled bank balances | Bank portal screenshots |
| Market or external data | A named, dated source, cited inline | The model's general knowledge |
Two rules sit on top of the table. Every number opens its source: each figure in an output links or footnotes to the rows behind it and when they were read. And if the source isn't available, say so: a gap marked "not available" is fine, a plausible invented figure is not. Point metric lookups at one set of definitions (see the number book guide) so revenue means the same thing in every skill.
House voice
A one-page style guide that every writing skill loads. Keep it concrete:
- Lead with the conclusion, then the number, then the reason.
- Numbers before adjectives: "up 14% ($210k)", not "significantly higher".
- A banned-words list: "robust", "headwinds", "leverage" as a verb, and whatever your CEO dislikes.
- Formats: currency and decimals, how to show variances, what a table header looks like.
The review pattern
The builder never grades its own work. The same session that wrote the output will find it convincing. A separate reviewer, briefed as the real reader, scores it.
For each skill, define who the output is for and have an independent AI reviewer play that person: an analyst checking the arithmetic, the COO asking what it means for operations, the CFO checking the message, a board member looking for what's missing. Each scores the output against the skill's checks and lists what they'd push back on. A person then reads the draft and the review together.
Track the scores. A skill whose output keeps scoring badly gets rewritten. A skill nobody has run in two months gets retired. A short list of skills people trust beats a long list nobody uses.
What not to let it do
- Post, pay or send. Draft the journal, the payment run or the customer email; a person with authority sends it.
- Hold write access to the ERP or bank. Read-only connections cover nearly every useful task.
- Make judgement calls on its own. Reserves, impairments, revenue recognition positions: it can lay out the analysis, a person decides.
- Work from pasted numbers for recurring tasks. Pasting is fine for a one-off. Anything that runs monthly should read from the source.
The first five skills to write
- Flux commentary. High volume, easy to check, and saves hours every close.
- Meeting prep. Before any meeting, pull the relevant numbers, open items and last notes into one page.
- Revenue bridge. A written classification rule plus the bridge table and two-line explanation.
- Excel and deck standards. So anything Claude builds looks like your team built it.
- A morning brief. Cash, collections, what's late in the close, and today's meetings, each figure linked to its source.
Run each on real work for a month before writing the next. Most of the value comes from tightening the first few, not from adding more.
Common mistakes
- Prompts in people's heads. If the method isn't written in a skill, it isn't repeatable and it can't be reviewed.
- One giant skill. A skill that does "all of month-end" is hard to test. One task per skill.
- No sourcing standard. Without it, the model fills gaps with plausible numbers. This is the failure that gets AI banned from a finance team.
- Self-review. Asking the same session "is this right?" gets a yes. Use a separate reviewer.
FAQ
What is a Claude skill?
A skill is a reusable written playbook that Claude loads when it does a particular task. It holds the steps, the data sources, the output format and the checks for that task, so the result is the same whoever on the team runs it. A finance team might have skills for flux commentary, the revenue bridge, board prep and meeting prep.
Is it safe to give Claude access to financial data?
It can be, with the right setup: a business or enterprise plan with appropriate data terms, read-only connections to the systems it needs, and no ability to post, pay or send on its own. Check the plan's data handling terms against your own policies and involve IT before connecting anything sensitive.
How long does it take to set up Claude for a finance team?
The first useful skill can be written and tested in an afternoon. A starter set of five skills, a data sourcing standard and a house voice usually takes a few weeks of part-time work, mostly spent running the skills on real tasks and tightening them where the output falls short.
Get new playbooks by email
Practical notes on AI in finance from a sitting CFO. Occasional, no selling.
Subscribe