Choosing an AI assistant for finance work, and what neither ChatGPT nor Claude should do with your books
Both assistants are good at drafting, summarising and analysing an export. Neither should be the source of your numbers. A fair way to choose, and the lines to draw.
Manoj Tutika6 min read
Working on this yourself? Talk to Mark, our co-founder. We’re here to solve it with you.
On this page
For a finance lead, OpenAI's ChatGPT and Anthropic's Claude are both capable general assistants: they draft, summarise, explain and analyse files you give them. The choice between them matters less than two other decisions, namely which plan and data settings you use, and which finance work you never hand to a chat window at all.
This post is for finance leads and controllers in growing companies who already use one of the two, or both, and want to use them on purpose. We will not crown a winner. Both products change every few months, and any comparison of features is out of date by the time you read it. What lasts is a way to choose and a clear line around your books.
What are both assistants genuinely good at in finance?
A lot of finance work is language work wrapped around numbers. That is where both assistants earn their keep.
- Drafting. The covering note for the monthly report, an email to a customer about a disputed invoice, a first version of a finance policy, the agenda for the budget meeting.
- Summarising long documents. A 40-page supplier contract, a lease, a loan agreement. Ask for the payment terms, indexation clauses, termination notice and anything that affects how you book it.
- Explaining. How deferred revenue works for an annual SaaS contract, what a covenant ratio means, how a reverse-charge invoice should look. Treat the answer as a starting point and check it against the source.
- Analysing an export. Both can read a spreadsheet or CSV you upload and run calculations on it. Upload an aged debtors list and ask for the ten largest overdue balances with days overdue, and you will usually get a correct answer in seconds.
- Translating. Dutch, English, German and French correspondence, including the formal tone a Dutch payment reminder needs.
For most of these tasks, a careful user gets a useful result from either tool. The quality depends far more on the prompt and the material you provide than on the brand.
Where do ChatGPT and Claude differ?
They do differ, in default writing style, in how long a document they handle comfortably, in which file types and integrations they support, in their admin controls and in their pricing tiers. Claude, for example, has an add-in for Excel, and both support connectors to other software through the Model Context Protocol. ChatGPT has a broad set of built-in features beyond text, such as image and voice.
But we would be careful with any claim that one is "better at finance". Public benchmarks rarely test the work you do, and the gap on a given task can flip with the next model release. So instead of trusting a comparison, run your own.
How should a finance lead test them?
Pick five real tasks from last month, with the identifying details removed, and give the same material and prompt to each assistant. A good set covers the range of your work.
- A variance explanation. Give this month's and last month's P&L by account and ask for the five largest movements with a one-line explanation each, marking what it is guessing.
- A contract question. Give a supplier contract and ask what the notice period is, whether prices are indexed and how the payment terms read.
- An export analysis. Give an aged debtors export and ask for the overdue total per customer and the share over 60 days.
- A reminder email. Ask for a polite second reminder in Dutch for an invoice that is 30 days overdue.
- An accounting question you already know the answer to. For example, how to treat a twelve-month software licence paid upfront under Dutch GAAP or IFRS.
Then score each answer on four things.
- Correct. Are the numbers right? Check one or two calculations by hand.
- Honest. Did it say what it did not know, or did it fill gaps with confident guesses?
- Usable. How much editing before you would send or file it?
- Traceable. Can you see where each number came from?
An hour of this tells you more than any review. Many teams end up using both, for different tasks, which is fine as long as the data rules are the same.
What should you check before putting finance data in either one?
The plan matters more than the product. Consumer and business plans have different terms, and those terms change.
- Training. Check whether your conversations can be used to improve the models, and switch that off where you can. Both vendors state that their business offerings do not train on customer content by default, but read the current terms of the plan you actually pay for.
- A data processing agreement. Under the GDPR you need one when personal data such as customer names, employee salaries or bank details goes into a tool. Business plans usually offer one. Personal accounts do not.
- Where data is stored. If EU or UK data residency matters to you or your customers, check what the plan offers.
- Retention and sharing. How long chats are kept, whether shared links are public, and whether you as an admin can see and delete your team's conversations.
- Who has an account. A controller on a personal plan with the payroll file uploaded is a bigger risk than any model choice.
If you are not sure, start with a rule the whole team understands. No personal data and no full ledger exports in any assistant until the business plan and the data processing agreement are in place.
What should neither assistant do with your books?
This is the part that matters most, and it is the same for both tools.
- Be the source of your numbers. A chat assistant does not know your revenue unless you give it a file, and it can misread that file. The ledger is the source. A number in a board pack must be traceable to the ledger, not to a conversation.
- Post entries unsupervised. Pasting an assistant's suggested journal into your ledger without review is how a wrong VAT code or a reversed sign ends up in your filing.
- Decide your tax treatment. It can explain how reverse charge or the OSS scheme works in general. Whether it applies to a specific transaction of yours is a question for your accountant.
- Reconcile from memory. Matching bank lines to invoices needs the complete, current data on both sides. A one-off export in a chat is a snapshot that is wrong by tomorrow.
- Sign off. Approving a payment run, closing a period or filing a VAT return stays with a person who is accountable for it.
The common thread is that a chat window works on a copy. Your books need work done on the original, with a record of who changed what.
What sits between a chat assistant and your books?
The gap between "the assistant explained it" and "it is correct in the ledger" is where most finance time still goes. That is the gap Thred is built for.
Thred works on top of the accounting software you already use. Through Thred Layer, agents do the recurring work such as reconciliation, invoice approvals and collections inside your live ledger data, within rules you set, and every entry is traceable to its source document. You can ask questions in Thred's own chat, or in Claude, ChatGPT, Slack or WhatsApp, and the answers come from the live books instead of an export. Thred reads by default and writes only with your permission, your data stays in the EU and is never used to train a model, which our security page explains in detail.
So the honest answer to "ChatGPT or Claude?" is that either is a good assistant for a finance lead. Just make sure it is talking to your books, not guessing at them. If you lead finance in a growing company, our businesses page shows how that works day to day.
If you want to see your own ledger answering in the chat you already use, book a demo.
Frequently asked questions
Is ChatGPT or Claude better for accounting?
Neither is consistently better across finance tasks, and the gap on a given task can change with each model release. Run the same real tasks through both and score correctness, honesty about uncertainty, usability and traceability.
Can I upload my trial balance to ChatGPT or Claude?
Technically yes, both can analyse spreadsheets. Do it only on a business plan with training switched off and a data processing agreement, and remove personal data you do not need for the question.
Do ChatGPT and Claude train on company data?
Both vendors state that their business offerings do not train on customer content by default, while consumer plans have their own settings. Read the current terms of the plan you use and check the settings for your workspace.
Can an AI assistant prepare my VAT return?
A chat assistant can explain how VAT rules work and help you spot lines to review, but it should not decide treatments or file anything. The return should be prepared from your ledger and checked by you or your accountant.
Do I need a data processing agreement to use an AI assistant at work?
Under the GDPR you need one when the tool processes personal data on your behalf, such as customer names, salaries or bank details. Business plans of the main assistants usually offer one, personal accounts generally do not.
Sources
Read next
- Connecting Claude or ChatGPT to your accounting system: what is safe
- Your finance data in Claude or ChatGPT through MCP: what it is and what to watch out for
- Coding agents for finance teams: what you can build and the guardrails
Where Thred comes in
Finance agents running on top of your existing ledger. You set the rules, approve the work and handle the exceptions. No migration.



