These are person directed tasks rather than production workflows. A person reviews and owns every output.
01
The tools
The task grid has been compiled with the Claude ecosystem in mind, because of its depth and ease of access for finance work today. This ecosystem has three main surfaces that are relevant to the tasks outlined:
- Claude chat for one off questions and file uploads in the browser or app.
- Claude Cowork for working on files already on your computer and developing workflows.
- Claude add ins that sit inside Excel, PowerPoint and Word, with Outlook in beta.
The grid is still useful for organisations that don't use Claude. ChatGPT with Codex and Microsoft Copilot do broadly the same, and indeed for a Microsoft organisation Copilot is a very sensible choice.
02
Cowork overview
Cowork is Anthropic's application that runs on your own computer, Mac or Windows and is the best surface for the majority of these tasks. Claude chat only sees a file when you upload it into the conversation whereas Cowork works directly with files already on your machine. Access is granted to specific folders or drives and Cowork reads and writes files inside those folders. Throughout the grid, the designated folder means an ordinary folder on your machine that has Cowork access.
03
Surfaces and automation
The tasks list in the grid outlines the main steps, rather than an exhaustive list, written to give enough guidance to easily implement with support from your AI tool. When seeking to implement it is useful to have the following points in mind:
- Match the surface to the task. A task that recurs against a stable reference, such as a policy, a template, a mapping or a set of contracts, runs in Cowork.
- A task that lives inside an open workbook or deck uses the Claude add in within that file.
- Use Claude chat for one off tasks, unless working inside a project. Claude like other LLMs struggle with context but switching on memory within the tool settings is advisable to assist in this respect.
Finally, the steps in the grid are written with manual file handling in mind for clarity: put the files in the designated folder, ask, review. However, once a task runs consistently ask Claude to write a Skill that describes it. Once the skill is saved within Claude and the relevant connections are enabled in settings (for example Outlook, OneDrive, SharePoint or Gmail) you can let Cowork automate future tasks. For example connecting Claude to a vendor or customer inbox it can then collect the documents itself, file them in the designated folder, and even run the saved review on a schedule (though this only works if the PC or Mac is on and logged into Claude) against the stored reference, such as the contract summary. The manual version is how you learn the task; the scheduled version is how you run it.
04
Plan, data and governance
Use a Team or Enterprise plan for finance work. This is a strong recommendation for two reasons. Anthropic's privacy documentation states that inputs on its commercial plans, which include Team and Enterprise, are not used to train its models by default. On the consumer plans they are, unless the setting is turned off. The commercial plans also give administrator control over who can access data and how long it is kept. Whatever the plan, keep personal data and client confidential material out of any task that does not need it, and ALWAYS out of the in app add ins in particular. Anthropic's documentation states the add ins currently sit outside Enterprise audit logging. Treat documents and emails arriving from outside the organisation as untrusted input. Anthropic's guidance is not to run the add ins on files from sources you do not trust, since a hostile document can carry instructions aimed at the model. Nothing in the grid sends an email, posts to the ledger, moves money or concludes an accounting judgement. Every output is a first draft that a person reviews and owns.
05
Writing the prompt
The risk with these tools is not the exception, which will always be reviewed, but the item that comes back matched or confirmed because of an AI hallucination. That happens when the model pairs things up from what looks right rather than working them out. Cut that risk by making the model calculate rather than recognise a pattern, and by deploying a prompt framework that makes its work checkable. Key elements of the prompt framework are as follows, a very effective approach is to feed these principles into Claude along with the task detail and ask it to provide a suitable prompt:
- Ask Cowork to do the matching and totals in code, meaning it writes and runs a short script that compares the values and adds up the figures, rather than deciding the pairings and sums by inference. The script gives the same result every time and cannot invent a match. When the model works it out itself, it can return a confident but wrong pairing.
- Require a control total that must cast. Ask Cowork for the totals it must reconcile to ensure completeness and to flag anything that does not tie rather than ignore it. A fabricated match breaks a total, and the broken total is what catches it.
- Forbid the model from typing in any number itself. Every figure in the output must be pulled or calculated from a file you supplied. The model sorts, matches, runs calculations and formats; it never keys a value from memory, because that is where a wrong number enters unseen.
- Make every line cite the exact source it came from: the invoice line, the ledger row, the contract clause. Where the source does not hold the answer, the model leaves the field blank and flags it as missing source data rather than writing something plausible to fill the gap.
- Confine the model to the files supplied. Ask Cowork to use only what is in the designated folder, not its general knowledge, so it cannot fill a gap with an invention.
Built this way, the confirmed output is far easier to trust, because a calculation produced it and the totals prove it. That is what lets a person check a sample of the high value items in the matched set rather than read every line.