Invoice capture with ChatGPT or Claude: when DIY AI is enough
Yes, for a few documents a month that the owner checks personally. Without deterministic checks, avoidable errors reach the books, reviewers stop checking once the output has been right for a while, and the fixes become double work in the accounting system. When a colleague runs it from their own subscription, someone has to maintain it and document it for the day they leave.
Will better models solve this?
They already read new supplier layouts well and will keep improving. Better output makes careful checking harder to sustain, not easier. It also does not maintain the setup: someone still owns the instructions, the account and the route into the books.
In this article
Why it works when the owner does it
A sole trader with a few supplier invoices a month uploads them to ChatGPT, Claude or another assistant and asks for the supplier, date, net, VAT and total. The result goes into a spreadsheet or the accounting software. It takes minutes, on a subscription that is already paid for.
Current models are also good at the part that used to break capture tools. A new supplier layout, a crooked receipt photo or an unusual invoice design no longer needs a template. The assistant reads it anyway.
It works because the owner is the control. They know every supplier, remember roughly what each bill should be, and own the account, the instructions and the consequences. When something looks wrong, they notice it and fix it once.
Without deterministic checks, avoidable errors reach the books
A language model reads each document afresh and returns the most likely answer. It is not deterministic. In one published test, the same prompt run 1,000 times with randomness turned off produced 80 different completions. A chat window also has no view of the rest of your books.
A deterministic check is a rule: the same input gives the same result every time. In invoice processing, the useful ones are simple.
- Do the lines add up to the invoice total, and does the tax summary match?
- Has this invoice, or this exact file, been processed before?
- Does this supplier's recurring charge go to the same account as last month?
Without them, errors that a rule would have caught go straight into the books: a bill entered twice, a line missing from a long invoice, a recurring charge coded to a different account from last month. These errors are unnecessary. The information needed to catch them was already there.
Reviewer fatigue: at some point nobody checks
The usual answer is a human in the loop. The assistant extracts, a person checks. That works in the first weeks, and research on automation says it does not last. When software is usually right, the people supervising it check less carefully and miss more of its mistakes. Experienced professionals are affected too, and training alone does not fix it (Parasuraman and Manzey, Human Factors, 2010). The Financial Reporting Council's March 2026 guidance on generative and agentic AI tells reviewers to be aware of the same risk, automation bias.
The arithmetic makes it worse. Suppose an assistant reads each field correctly 97% of the time and an invoice has 15 fields you care about. More than one invoice in three will still have something to correct, and nothing tells you which one. So you either check every row of every invoice, or you stop checking. After a few clean months, it is easy to drift into the second.
This is the uncomfortable part of better models. The more often the output is right, the harder it is to keep reading it carefully. Without deterministic checks, the checking lasts only as long as someone's attention does.
Errors turn into double work in the system of record
The accounting system is where the record lives. Assistant output usually reaches it by copy, import or a second entry, so the work is already done twice before anything goes wrong. When something does go wrong, the fix happens in the books: find the bill, correct or void it, redo the reconciliation and check whether a VAT return already used it.
In August 2026, AccountingWEB reported an accountant who found that a large client had used Claude to convert PDF bank statements into an Excel cash book. The client did not check the output, and numerous transactions were missing. The time saved at capture was spent again at reconciliation, by the accountant.
Someone has to maintain it, and document it
A DIY setup is a small piece of software, even when it is only a prompt. It has instructions or a skill, a subscription, the computer or browser session it runs from, and a route into the accounting system. Each part needs looking after.
- The model changes under you. When OpenAI retired GPT-4o and other ChatGPT models on 13 February 2026, its help centre said custom GPTs using them would automatically move to the closest current equivalent. The instructions stayed the same. The model reading them did not.
- The instructions drift. A line added for one awkward supplier changes the output for others. Copies of the prompt on two laptops diverge.
- It runs where the subscription runs. A setup built on a personal subscription usually lives in one person's account, often on their own computer.
When the owner runs it, that maintenance is their own time, and that can be a fair trade. When a colleague runs it, it is a business process, so write it down: where the instructions live, which account and plan it uses, what the reviewer checks, how the output reaches the books and what to do when it fails. Otherwise the process leaves when the colleague does.
If the documents belong to clients, there is one more check. ICAEW's summary of the updated professional conduct guidance warns that uploading identifiable client data into uncontrolled environments can create serious data protection risks, and that responsibility stays with the accountant.
What a capture service maintains for you
A capture service still reads documents with software, and people still review new suppliers and exceptions. The difference is that someone else maintains the infrastructure, and the steps around the reading are deterministic. Datamolino maintains the capture service, the web and mobile apps and the connections to Xero, QuickBooks Online and FreeAgent. A change is made once, for every customer, not in a prompt on one laptop.
- Intake is shared. Suppliers email invoices to the folder's own address or the team uploads them, so documents do not sit in one person's chat history. Route supplier emails to a folder.
- Totals are checked. With Validate totals enabled, a line-item or tax-summary mismatch blocks export until someone resolves it. See invoice capture and checks.
- Duplicates are flagged. Identical files are identified before capture, and captured invoice data is compared within a folder. How duplicate checks work.
- Coding repeats saved decisions. Accounts and tax codes come from rules saved from your decisions, which you can inspect and test. A model update does not change them. Automate recurring coding.
- Approval leaves a record. When a folder uses approvals, the configured approvers sign off before export. Set up invoice approvals.
- The process stays when people leave. The documents, rules and settings live in the folder, available to everyone with access, not in one colleague's account.
The checks run on every document, so a lapse in attention does not switch them off.
When DIY is enough
| Situation | A DIY assistant is enough | A capture service earns its cost |
|---|---|---|
| Volume | A few documents a month | Dozens a month, or several clients |
| Who checks | The owner | Colleagues, or clients who approve |
| What catches errors | The owner's attention | Rules that run on every document |
| Who maintains it | The owner, in their own time | The vendor |
| If the person leaves | Not a concern | The process has to keep running |
If your answers sit in the middle column, keep the assistant. At that volume, a service is hard to justify on time saved alone. If they sit in the right column, the question is no longer whether an assistant can read an invoice. It is who maintains the setup and what checks every document when nobody is looking.
Datamolino's free trial covers 14 days and up to 100 documents, with no payment details to start.
What does a service like Datamolino add?
It maintains the capture, the checks and the accounting connections for you. Its checks and coding rules are deterministic: the same condition produces the same result every time, whoever is working that day.
