The takeaways from my Claude for Finance workshop
Claude for your finance work
A practical starter guide for finance teams: how Claude actually works, the examples I run in the workshop, the exact prompts to try yourself, and the tools and links to get going.
Keep this page. It is written so you can come back to any part of it when you are ready to try something, without needing to remember it all at once.
Section 01
How to use this guide
In the workshop I show finance teams how Claude takes on real work: reading the accounting system, working through the inbox and the spreadsheets, and repeating the tasks that eat the month. This guide turns that into something you can act on.
- The prompts are yours to use. Every grey dashed box is a prompt you can copy and paste straight into Claude. Change the details in [brackets] to match your own work.
- Start small. Pick one task that eats your time each month and try it. You do not need to set everything up first.
- When in doubt, ask Claude. The single most useful habit: if you are not sure how to do something, ask Claude directly. It knows how it is set up and can walk you through it.
In the workshop everything runs on fictional demo data, never live books. When you connect Claude to your own systems, the connection to your accounting package is read-only: Claude reads the numbers, reconciles, and drafts the work, and a person still reviews and posts. That is by design, and it is the safe way to work.
Section 02
How Claude actually works
A few ideas make everything else make sense. If you only remember these, you are most of the way there.
It predicts the next word. It is not Google, and it is not magic.
Claude is a large language model. It was trained on an enormous amount of text, and at heart it does one thing very well: it takes what you have written and works out the next most probable word, then the next, and so on. That is why it can write and reason so well, and it is also why it sometimes gets things wrong or "makes things up". It is working on probability, not looking up a fixed answer. So treat its output as a smart first draft to check, not gospel.
The Memento idea: it wakes up with no memory
Think of the film Memento, where the main character cannot form new memories and leaves himself notes so that each morning he can work out who he is and what he is doing. Claude is the same. Every time you open a new chat it starts fresh. Before it can help you, it reads a set of plain-text notes about who you are, how you work, and the rules to follow.
Those notes are simple markdown files (plain text, no formatting). The practical lesson is the one worth tattooing on your arm: if you do not give Claude good context, you will get poor results, because it is like a person who wakes up each morning not knowing who they are. Good setup is most of the difference between people who find AI amazing and people who find it useless.
Treat it like a smart but forgetful junior employee
The best working mindset. Claude is capable, fast, and eager, but new to your world every time. So brief it clearly, let it show you its plan before it acts, correct it plainly when it is wrong ("this is wrong, here is the output I want, here is how I want it done"), and ask it to remember the correction for next time. You are not hunting for a hidden setting like you would in normal software. There is an intelligence behind it. Just tell it what you want.
Before letting Claude do anything big, ask for a plan first: "This is who I am, this is what I want, this is the information I have. Give me a plan. Do not take any action yet." Go back and forth until you are happy, then say "go ahead". This keeps you in control and catches misunderstandings early.
Section 03
Setting up safely
Two minutes of setup protects your data. Do this before you put any real work in.
Turn off two privacy settings
In Claude, open Settings → Privacy and switch off:
- Location metadata, so it is not recording where you are.
- Using your chats to train the models, so your information is not fed back into the AI.
Never paste sensitive personal data into the chat box
No ID numbers, no banking details, no confidential personal information typed directly into a chat. Anything you type becomes part of your chat record.
A connector is actually the safest way to work, safer than pasting a file or typing data into the chat. When Claude reads your data through a connector (your accounting system, your email, your files), it goes through that app's own security and permissions. Nothing sensitive gets copied into the chat. So the goal is to connect your systems, not to paste from them.
Use a business account for team work
A company (Team or Enterprise) account gives you stronger data protection and lets your team share projects properly. It works like Microsoft 365: an administrator manages the seats and the settings for everyone. A personal account is fine for exploring on your own, but the real team setup should be on a company plan.
Section 04
Models, sessions and credits
Models are like levels of seniority
Claude comes in several models, and the easiest way to think about them is as employees at different levels. Stronger models are more capable but use up your usage faster, so you match the model to the job.
| Model | Think of it as | Use it for |
|---|---|---|
| Opus (and the top "Fable" tier) | A senior executive with fifteen-plus years' experience | Your hardest, most important work |
| Sonnet | A solid person a few years in | Everyday work, most tasks |
| Haiku | A sharp graduate, fresh out of university | Quick, simple, high-volume jobs |
You would not ask a senior executive to quickly check your calendar. Same here: do not spend your strongest model on trivial tasks. A good habit is to pick two models, a strong one and a basic one, and switch between them. When something really matters, you can also raise the effort level from the default (High) up to Max to make the model think harder.
Sessions have a memory limit
Each chat can only hold so much before it fills up (roughly 200,000 "tokens" on a basic account, up to about 1,000,000 on the top plan). Picture briefing someone for three solid hours and then asking them about the very start: they cannot hold all of it. When a chat gets near full, Claude tries to compress it, and that is where it can quietly drop something that mattered and start making mistakes.
When a chat is getting long, ask Claude to save everything important to a file, then start a fresh chat and hand that file back to it. You carry the context across cleanly instead of letting it get squeezed out.
Watch your usage
Under Settings → Usage you can see how much of your allowance you have used, per model, and top up if you need to keep working. It resets on a rolling basis.
Desktop app vs the browser
The browser version at claude.ai can reach anything in the cloud (OneDrive, SharePoint, Google Drive) and everything in this guide. The desktop app can additionally reach files saved locally on your own computer. For finance work through your accounting system, email and spreadsheets, the browser is all you need.
Section 05
Skills, connectors and plugins
Three ways to make Claude more capable. You will find them under the customise menu.
Skills
A written note that teaches Claude how to do one task your way: how to do it, when to use it, what to watch out for. Just like writing down a process for a new colleague. You can use ready-made ones or build your own.
Connectors
Built by the software makers themselves to link their app to Claude. Xero, Microsoft 365, Google Drive, DocuSign and thousands more. Most are read-only for safety.
Plugins
A bundle of skills plus the connectors for a specific job, packaged together. For example, a marketing plugin brings marketing skills and the right connectors in one go.
You never need to hand-write Claude's instructions or settings. Work with it instead: ask it what it already knows about you and how you work, correct it, then say "make sure these settings are saved everywhere." It builds and keeps its own notes.
Section 06
Connecting your finance systems
The connections that matter most for finance work. Which ones apply depends on the systems you run.
Your accounting system
Most modern accounting packages now connect to Claude, though how depends on the package. Xero has a native connector that is ready to go and quick to set up: Claude can see your invoices, receivables, payables, cash position and top customers by revenue, and answer questions across them. It is mostly read-only, which is a safety measure. Sage Intacct connects through Sage's own AI Gateway (an official, read-only route); it is real but not a free switch, so plan for it (see the note below). If you are on a different package, use the trick at the end of this section to check.
For Sage Intacct specifically, connecting Claude is a small project, not an instant switch. On the Sage side you need an active Web Services Developer licence, the AI Gateway add-on subscription (a separate paid Sage add-on), and API credentials plus a one-time secure connection setup. These are arranged with your Sage account manager, and the connection stays read-only throughout.
Links Sage Intacct AI Gateway guide: developer.sage.com → AI Gateway guide. Connection steps: MCP setup and connection guide.
Microsoft 365 (or Google Workspace)
The Microsoft 365 connector links Claude to your Outlook email and your Excel, PowerPoint and SharePoint files. This is what lets Claude read an email and a spreadsheet in the same breath as your accounting data. You can either work from inside Claude, or use the small Claude panel that appears inside an Excel or PowerPoint document. If you run Google Workspace instead, the Gmail, Google Drive and Google Sheets connectors do the same job.
Sometimes a newly added connector shows an error. The fix nearly always works: fully close Claude (on Windows, open Task Manager and end every Claude task), reopen it, go back into the same chat and say "I just restarted, can you check the connection again please."
How to find a connector for anything
Not sure whether your system connects? Search the internet for the app's name plus the word "MCP" (for example, "[your payroll system] MCP"). If there is a connector, that finds it. If there is not one yet, there is almost always a workaround: export the data to a cloud folder Claude can read (OneDrive, SharePoint or Google Drive), then work over it from there, including on a schedule.
Section 07
Worked examples to try
The examples from the workshop, each one something you can try yourself. The prompts use placeholders in [brackets]; swap in your own details. Give Claude access to the right file or connector first, or simply start and it will tell you what it is missing.
Ask your books a question
Replace digging through reports with a plain-language question, answered straight from the ledger.
Reconcile a supplier statement
Match a supplier's statement against your ledger, find what does not agree, and get the exact query to send back. This is the email connector and the accounting connector working together in one place.
Find what should not be paid
Let Claude catch the problems before payment. This is a strong "we did not know what we did not know" moment.
A useful habit for anyone who handles supplier accounts: when a supplier emails asking to change their banking details, have Claude check it before anyone touches a payment.
Draft replies from the facts
Claude does the writing too, grounded in the actual invoice and purchase order rather than made up. A good habit: ask for the draft in the chat so you can read, edit and paste it yourself, rather than having it send anything.
Cost a new job in your own spreadsheet
Claude works inside your real Excel model, keeping your formulas, not a flattened copy. In the workshop it prices a job, adds a scenario sheet, re-checks the workbook, and drafts the reply.
Pull information out of documents, PDFs and scans
Claude can read invoices, contracts, statements and scanned documents and compile them into one place. Two things to know: scanned images and photos cost far more to read than text, and handwriting is the least reliable. So the smart approach is to define what you want first, then let it work in batches.
The reliable way to do it
- Define a schema first. Tell Claude exactly which fields you want from each document before it starts.
- Give it the documents and let it sort them by type and cost.
- Extract, then batch rather than doing everything in one go.
- Merge and reconcile into your chosen output, a spreadsheet or a document.
For messy or scanned documents, it also helps to convert them to clean text first with a free tool. See MarkItDown below.
Teach it a task, then have it run on a schedule
The shift from "chat" to "an employee that repeats your process". Scheduled tasks run in the cloud, so they happen even when your laptop is off. A good first one: the weekly payment run, pull the week's payables, fill the template, combine to a PDF for signature.
Where to set one up
- Open a Project, then Schedules in the left sidebar.
- Choose New task → Create with Claude (not "set up manually").
- Describe the task as if briefing a new intern. Claude fills in the details and asks you a few questions (where the data lives, how to deliver it, what time to run).
Section 08
Handy tools and links
The tools and resources from the workshop, with where to find them.
OpenRouter
openrouter.ai · leaderboards at openrouter.ai/rankings
A place to use and compare many AI models in one account, and see which models are best at which kind of task (writing, planning, data extraction). Useful when you want to match the right model to a job.
Wispr Flow (dictation)
wisprflow.ai · download at wisprflow.ai/downloads
Lets you talk instead of type, anywhere on your computer. Hold a key, speak, and your words appear as clean text. Modern AI understands natural, slightly rambling instructions, so talking to it is often faster and better than typing. (Claude also has a built-in dictate button if you would rather not install anything.)
MarkItDown (documents to clean text)
github.com/microsoft/markitdown · free and open-source, from Microsoft
Converts PDFs, Word documents, scans and more into plain markdown text. Because Claude works best and cheapest with clean text, running messy or scanned documents through this first is a good habit for the document-extraction work. It is free. Ask Claude to help you set it up.
Telegram, for notifications from Claude
telegram.org
If you want Claude to message you when a scheduled task is done or something needs your attention, Telegram is the easy way to do it (it is far simpler to connect than WhatsApp, which Meta keeps closed). Set up a Telegram account for your Claude assistant to message you.
See how AI reads text (the tokeniser)
A free page that shows how a model breaks your text into "tokens", each with its own ID. A quick way to see, rather than just hear, how the next-word idea works underneath.
Bringing your ChatGPT history across
If you have built up history in ChatGPT, you do not have to lose it. Ask ChatGPT to write a full record of everything it knows about you and how you work into a markdown file, save it, and hand that file to Claude.
Section 09
Spotting a good job for AI
The most useful thing to take away is a simple test for what to hand to Claude. A task is a strong candidate when it has all three of these.
It is document-based
The work lives in files, emails, spreadsheets or records, not only in someone's head or in a conversation down the corridor.
It is rule-based
There is a repeatable way you do it: check this against that, flag anything over a limit, follow the same steps each time.
You can verify it
You can look at the result and tell straight away whether it is right or wrong, so you stay in control and catch any mistake.
Most finance, admin and operations work passes this test, and the examples earlier in this guide all do. When you are wondering whether something is worth trying, run it through these three questions first, and start with the task that eats the most of your time.
Treat Claude like a smart but forgetful junior employee. Brief it clearly, ask for a plan before it acts, correct it plainly when it is wrong, and let it do the document-based, rule-based, checkable work while you keep the judgement and the final say. Two minutes to set a task up can save you thirty minutes every week from then on.