Lucy Liu
July 2026 · AI for Small Business

Every Receipt Files Itself Now

Watercolor illustration of ginkgo leaves drifting into a small open wooden box

Receipts, invoices, bank statements. Every small company generates a pile of them, and someone has to file that pile. At our company, that someone was me. Filing accounting records is exactly the kind of work a system should be doing instead of a person.

So I stood one up. I am not a coder. I direct AI tools and I verify outcomes. This build is a clean example of what that seat actually looks like.

What I built

The system is paperless-ngx, a community open-source document management tool. Its GitHub repository has over 43,000 stars. I treat that number as due diligence: 43,000 people voting that "documents that file themselves" is a real problem worth solving, and that this particular tool solves it. When you are not a coder, the star count is your reference check.

Deployment was docker-compose with two services, a Redis broker and the web server, running locally on my machine. I directed the AI through the setup and confirmed each step myself: containers running, web interface loading, a test document going in and coming out filed.

The workflow now: I drop any receipt, invoice, or statement into a consume folder. The system picks it up, runs OCR so the text is searchable, indexes it, and files it. No renaming, no sorting into subfolders, no "I'll deal with this Friday" pile.

One more piece. The document folders are bind-mounted to Google Drive, so the whole archive syncs to the cloud. Backed up, and reachable from anywhere. If my laptop dies tomorrow, the records survive.

The part that was actually hard

Installation was not the risk. The AI handled docker-compose fine, and the software worked on the first real test.

The risk showed up when I understood what I had built. That consume folder is not scratch space. Anything dropped there becomes a filed business record. And because the folders sync to Drive, a careless bulk delete would not stay local. It would propagate straight into the cloud archive. The convenience that makes the system worth having is the same property that makes a mistake expensive.

I use AI agents across this project, and agents move fast. Fast is fine for building a website. Fast is dangerous next to live accounting records. So before I let any agent touch the project again, I wrote guardrails into the project docs, in plain written rules: never bulk-delete inside synced paths. Never remove the docker volumes that hold the database. Restart the system only through a defined safe procedure.

That was the actual operator work. Not standing the system up. Writing the rules that keep it safe to operate. Any AI tool can install software now. Deciding what must never happen to your records, and putting it in writing where every future agent will read it, is judgment. That part does not get automated.

The outcome

Our company's accounting records now file themselves. A receipt goes into a folder and comes out as a searchable, indexed, cloud-backed record. The filing backlog I used to carry in my head is gone, and finding any document is a search box instead of an archaeology project.

What you can copy this week

A founder can replicate this pattern in an afternoon. Three moves.

Pick the boring system everyone already trusts. You do not need to evaluate ten tools. The star count is the vote. Take the one with the deepest community and the longest track record.

Put the intake where files already land. The system works because dropping a file in a folder is zero extra effort. If your workflow requires a new habit, it will fail. Mine requires none.

Write the guardrails down before you automate anything. This is the step people skip, and it is the one that matters. Before any AI agent or automation touches live business data, write the rules: what must never be deleted, what paths are live, how to restart safely. Ten minutes of writing, and every tool that touches the system afterward inherits the rules.

Most automation advice sells you the install. The install is the easy part now. The durable skill is knowing which records matter, what could destroy them, and writing that down before you hand the keys to a machine.

If you are a founder drowning in receipts, this is a solved problem. Send me a DM if you get stuck on the setup. I read them and I reply.

← All writingBook a call