Hey!

Chris here. Welcome to Blueprint—the newsletter to help you build a winning engineering team.

My bookkeeper is a contractor. And if everything goes according to plan, they won't be working with us for much longer.

They’ve done a great job, don’t get me wrong. But the work they do—pulling transactions, reconciling accounts, making sure QuickBooks matches reality—is exactly what AI agents are built for.

Every business has a version of this. The tedious, repetitive work that has to happen no matter what, but that nobody with real leverage over their time actually wants to sit down and do.

It falls to the bottom of the priority list, gets done quarterly when it should happen weekly, or quietly doesn't happen at all until something breaks.

AI handles this kind of work exceptionally well. The catch is that the tasks nobody wants to do are often the same tasks where a mistake costs you the most. So the way you set this up matters.

Let me walk you through how to do it safely. 👇️ 

📒 DEEP DIVE

How to (Safely) Automate the Work Nobody Wants to Do

A practical look at the right way to use AI agents for tedious financial and operational tasks.

The Setup Almost Every Business Has

There's a financial mess sitting underneath almost every business.

2 or 3 bank accounts. A business credit card. And somewhere in the middle of all of that, a person manually reconciling transactions, categorizing expenses, and probably wondering whether the chart of accounts they set up a few years ago still makes sense.

Here's what makes bookkeeping a particularly good first target for automation: double-entry accounting has checks and balances built into the system. If something goes wrong, you know it. The numbers don’t lie.

That built-in feedback loop makes it a lower-risk place to start than a lot of people assume.

How the Mechanics Work

Almost every financial system you use is accessible electronically. Some have APIs—which are just ways to reach out across the internet and pull data from systems you're already a customer of. And it's not read-only access. You can write. You can move data, update records, create entries.

For systems that don't have solid API support, there are ways to give your agent browser access and have it interact that way instead.

The workflow looks like this:

1. Connect your agent to your accounting system

2. Tell it to pull the transactions from your bank accounts and credit card, and run the reconciliation.

3. Ask it whether your chart of accounts is structured intelligently.

I did exactly this—and my agent came back and said the way I structured my tax accounts wasn't great. So I had it reorganize them.

That was always worth doing and never got done. Until now.

The Mental Model You Can't Hand Off

I tell every CTO and CIO I work with the same thing: you cannot outsource the mental model of how everything fits together. You have to understand how the data flows from one place to another and why.

In finance terms, that means you need to understand what a bank reconciliation is actually doing before you automate it. You need to know what a trial balance is. What a P&L shows you. What a balance statement represents. What a chart of accounts is and why the structure matters.

You don't need to be a CPA. But you need enough vocabulary to evaluate what the agent does and recognize when something looks wrong.

If you don't have it yet, the good news is that AI is a more-than-capable teacher.

The cutting-edge models understand accounting concepts deeply. You can build a solid mental model through conversation before you ever touch a live system.

I’m telling you: do that first. It costs you nothing and protects you from a lot.

The Safety Practices

Not enough people talk about the fact that the most tedious tasks are often the same tasks where a mistake is the most consequential.

Reconcile your accounts wrong, corrupt your chart of accounts, or touch the wrong data, and you have a problem on your hands.

That’s why there are a few things I do without exception when building automations like these.

When I give an agent access to any financial system, I create API accounts that are read-only. The agent can pull data, analyze it, surface recommendations—but it cannot write anything, move money, or make changes. Write permissions are locked out at the account level.

Before I make any modifications in a system like QuickBooks—cleaning up a chart of accounts, adding entries, changing anything—I pull a full backup first. Every time, no exceptions. These are not things people naturally think about. They should be.

And for anything touching live financial data, I use the best available model at the highest thinking setting.

A cutting-edge model running high-level reasoning is really difficult to get to do something destructive. It will typically stop you before you do something dumb. That's not a guarantee, but it's a real layer of protection on top of everything else.

The goal is to have these guardrails built into the system from the start—baked in, not bolted on. You want the rules present in a way where screwing things up isn't really possible in the first place.

What This Looks Like in Practice

Once you've done the setup correctly, here's what the day-to-day actually looks like.

Your agent runs continuously. It checks the transactions. It flags anything that looks off. It reports back on whether your books are clean. If it sees something that needs attention, it sends you a message—Slack, text, wherever you want it. You come in for the things that actually require a decision and stay out of everything else.

The contrast with doing this manually isn't subtle. If you do a manual bank reconciliation, you're doing it quarterly at best. And if you do it quarterly, you're ahead of the overwhelming majority of businesses.

But quarterly means you're always chasing problems that are already 3 months old. Running agents continuously means you catch things when they happen.

That's the shift. And it's available to any business willing to set it up the right way.

BEFORE YOU GO…

Pick 1 thing in your business right now that someone does manually on a bad cadence.

Think about the task that should happen weekly but happens quarterly, or one that doesn't happen at all because nobody wants to sit down with it.

That's your first target. Set it up with the right safety practices and get it running.

Talk soon,

Chris.