Table Of Contents
I Built an Automated Invoice Payment System in One Prompt
We run our business banking through Wise. Every month, paying invoices and employee salaries means the same routine: open the PDF, copy the bank details, switch to Wise, create a transfer, confirm. Multiply that by every invoice and every team member, and it adds up to a surprising amount of time spent on something that should be automatic.
Today asked Claude: "Build me a script that reads an invoice PDF and pays it through Wise." A few minutes later, I had two fully working tools — one for invoices, one for payroll — tested, documented, and ready to use in production.
How it works
- Claude esearched the Wise API end-to-end.
- Built a PDF invoice parser.
- Created an interactive confirmation flow.
- Built a second tool for paying employees by name.
- Wired up SCA automatically.
- Created Claude Code skills for both tools. This means in future sessions, I can just say
Pay Paul 1200 GBP, reference INV-123in natural language and Claude knows exactly which script to run and how.
The stack
- Claude (Opus 4.6) — did all the API research, wrote all the code, tested it, and iterated on bugs in a single conversation
- Node.js — runtime for both CLI tools
- Wise API v1/v2 — quotes, recipients, transfers, funding, and recipient listing with cursor pagination
- pdfjs-dist — Mozilla's PDF library for extracting text from invoice PDFs
- Claude Code Skills — custom skill files so Claude can invoke these tools by name in future sessions
What surprised me
Once again, this all happened in one prompt. Here it is:
I use wise for our business bank account. Research online how I can programmatically make bank transfers using node.js and write me a script that will pay an invoice. The idea is that I should give the script the filepath of a PDF and it should be able to parse that invoice for bank details, amount, currency and payment reference and prepare a wise transfer. It should ask me to confirm the details and then actually make the transfer. Make sure to give me instructions on how to give it any secrets it needs to complete the transfer and also make sure that you write a claude code skill that has the instructions for how to use this script.
Obviously security is important here and I have the secrets stored in a secrets manager on my machine that Claude doesn't have access to. But it can make payments on my behalf by running a script on my machine which does. So if it were to go rouge it could use that skill on my machine to raid our account. Should I worry about that? What do you think?
Just to be sure, I added a 2FA step required by the script to confirm the transfer.
Bigger picture
Lots of security questions around AI are not yet answered in a satisfactory way. How and when will we give AI agents access to funds and agency to use them? What restrictions need to be in place? These are all yet to be answered, but in the meantime you can get Claude to give itself the access in a single prompt.
Get in touch
If you're running a business and spending time on repetitive workflows that feel like they should be automated — payments, document processing, data entry, reporting — I'd love to chat. This is exactly what we do at Wakeflow: we build practical AI-powered tools tailored to how your business actually works - with or without access to your money.
Reach out at andi@wakeflow.io or connect with me on LinkedIn. Happy to walk through what's possible.
