← Back

Digital agency · billing platform

The billing app I killed.

I spent sixty hours over three or four months building a time-and-materials billing app to replace a process that runs on a spreadsheet. It worked. I killed it anyway.

This is the case study for the build that didn't ship — why I started it, why I stopped, and the much smaller thing that ended up doing the job.

Status: dead, deliberately. Our client projects really are too nuanced for off-the-shelf billing software — I wasn't wrong about that. A custom app could handle them. It was just overkill, and I'd rather say so than ship something I'd be maintaining forever.

60 hours — and I'd do it again

What those hours actually bought: almost everything I know about how this agency bills, and what is genuinely possible to improve.

What billing looked like

Billing runs out of a shared spreadsheet. Time logs get merged against the estimate records in the ledger, and then every invoice is entered from that sheet by hand. It is time-consuming and almost entirely manual.

The obvious move — have the skills create the invoices — doesn't work here. Billing uses progress invoicing, which the accounting platform's API doesn't support, so our MCP connection can't create those invoices either. That constraint is what sent me toward building an app in the first place.

Why I killed it

Every nuance meant another rule. Each one was reasonable on its own, and each one was correct — that is what made it hard to notice. Before long I could see the shape of the next five years clearly: me, maintaining that rulebook, forever, as the only person who understood it.

The app wasn't failing. I was. I had built something that could only ever be as current as my attention, in a business where the rules change every time a new contract gets signed.

So I stopped, and asked a different question: what was the spreadsheet actually costing? Not the keying — the errors. Staring at a sheet for eight hours a day across two or three days is mind-numbing work, and things slip through. The expensive part of this process was never the typing. It was the tired eyes.

What replaced it

an agent, not an application

I built an agent named Billinda. She runs a rules engine over every invoice we create before it goes to a client, checking the math, the descriptions and the contact emails. She catches a lot of what tired eyes miss, and she can produce some of our non-estimate invoices herself.

The division of labour turned out to be the interesting part. Sometimes I enter and Billinda reviews. Sometimes it's the other way around. We share the work rather than one of us replacing the other, and either order is fine because the checking is the same either way.

The app tried to prevent every mistake by encoding every rule in advance. The agent reads what was actually produced and tells me what looks wrong. One of those needs a rulebook maintained forever. The other needs a checklist I can change in an afternoon.

The paid media half of the app went the same way, and became a skill rather than a screen. It reconciles the month's spend, recomputes every management fee from the pricing rules, drafts the invoices for approval and ties the fee income back to the sheet — paid-media-billing →. Between the two of them, everything the platform was going to do is being done. None of it is an application.

Billinda runs on a schedule → through billing week, republishing the review page every two hours so the findings are current whenever anyone looks.

What's next

Moving off the ledger's estimate records — and their progress-invoicing limits — to a client project database of our own. It is a much smaller build than the one I killed: one source for project details that invoices can pull from automatically. Project managers could request changes to their own project records without needing access to the accounting platform, which we'd rather keep with Finance anyway.

Those sixty hours taught me almost everything I know about how this agency bills and what is possible to improve about it. I didn't come out the other side with what I expected. I came out knowing which build was actually worth doing, which is not a thing I could have reasoned my way to from the outside.

Just because a build dies doesn't mean it's a failure. It's part of the process. And yes — I'm already building the next one.

Built with Claude Code. Reviewed, reconciled, and argued with by a finance person who now can't stop shipping software. One of several things built for the agency.