Prepare S-Corp tax filings from bookkeeping data
Insights. Task: Insights By John Sambrook · February 15, 2026 · AI, Small Business, Practical Tips TL;DR I prepared my S-Corp tax filings using two AI coding agents — Claude Code writing Python scripts to compute the forms, Codex reviewing the code for errors. They caught real mistakes, including an incorrect Schedule L field mapping. The broader pattern: when building custom tools takes minutes, the skills that matter shift from operating software to understanding the domain. I’m doing my business taxes this year with AI coding agents. Not conceptually. Not as an experiment. I mean the actual tax forms, schedules, and filings for my S-Corp. Here’s what the setup looks like: financial data from QuickBooks and bank statements goes into a private Git repo. Claude Code writes Python scripts that parse that data and compute the tax forms. Codex reviews the scripts with a high-reasoning model. I direct the process, check the results, and make the decisions. Between the three of us, the taxes are getting done. Two Agents, One Checking the Other The workflow has a structure that turned out to be more important than I expected. Claude Code with Opus 4.6 writes most of the code. Codex then reviews it. This isn’t a formality. Codex finds real errors. Here’s a concrete example. Codex ran a review of my 1120S scripts and flagged three findings, severity-ordered. The most serious: an incorrect Schedule L field mapping where Line 18 (ending) was being computed from total liabilities instead of other current liabilities. That’s the kind of error that’s easy to make and hard to catch by staring at a spreadsheet. Codex caught it by tracing the data from source JSON through to the generated CSV. The review process is often multi-turn. Codex flags an issue, Claude Code fixes it, Codex reviews again. They go back and forth until both agree the code is correct. It’s not unlike a code review between two experienced developers, except it happens in minutes instead of days. The Photoshop Problem One of the unexpected benefits is having a QuickBooks expert available whenever I need one. I don’t spend enough time in QuickBooks to know it well. It’s the Photoshop problem: you know the software can do what you need, but you only use it a few times a year, so you never build fluency. Every session starts with fumbling through menus, trying to remember how reconciliations work, wondering what changed in the latest update. With Claude Code, that friction disappears. I don’t need to know what’s under every dropdown in the QuickBooks UI. I describe what I need in plain English. When I needed to reconcile my business credit card, I just said “Let’s reconcile -7301.” Claude Code knew that -7301 referred to my business credit card, told me what data I’d need, and gave me the full pathname to the proper credit card statement in the repo. No fumbling. No searching. This pattern generalizes far beyond QuickBooks. Any complex software you use infrequently imposes a recurring tax on your attention. You lose time re-learning the interface every time you sit down with it. AI agents eliminate that tax. You describe what you want to accomplish, and they handle the how. What I’m Actually Learning Since 1996, I’ve had my CPA prepare both my business and personal taxes. That’s thirty years of someone else understanding my tax situation better than I do. This year is different. By working through the process with AI agents, I’m learning how my own taxes actually work. Not because I set out to learn — the learning is a side effect of directing the process. When you have to describe to an AI agent what needs to happen with your Schedule K-1, you end up understanding your Schedule K-1. I found errors I made in 2025 in how I managed certain aspects of the business. They weren’t catastrophic, but they were real, and I wouldn’t have caught them if I’d just handed everything to a CPA as usual. I’ll be much less likely to make them again in 2026. What’s Next: Plain Text Accounting Once taxes are done, I’m planning to replace QuickBooks with Beancount, an open-source double-entry accounting system that uses plain text files for its records. That sounds crazy. I know. But consider what actually matters now. QuickBooks is designed for human eyeballs: menus, forms, reconciliation wizards. Beancount is designed for both humans and machines to read: structured plain text that lives in a Git repo, diffs cleanly, and is trivial for an AI agent to parse and modify.; tools: Claude Code, Codex. Human role: The original post does not say how a person stayed involved.; limitations: The source does not state specific limitations, costs, or integration hurdles. Reliability depends on the author's self-reported experience without independent verification. Human approval is critical for final decisions, as the agents only assist in computation and review. Proposed advice: validate all tax calculations with a qualified CPA before filing..
Tools
Published on Agent Work Atlas:
Overview
Insights
Workflow
Insights By John Sambrook · February 15, 2026 · AI, Small Business, Practical Tips TL;DR I prepared my S-Corp tax filings using two AI coding agents — Claude Code writing Python scripts to compute the forms, Codex reviewing the code for errors. They caught real mistakes, including an incorrect Schedule L field mapping. The broader pattern: when building custom tools takes minutes, the skills that matter shift from operating software to understanding the domain. I’m doing my business taxes this year with AI coding agents. Not conceptually. Not as an experiment. I mean the actual tax forms, schedules, and filings for my S-Corp. Here’s what the setup looks like: financial data from QuickBooks and bank statements goes into a private Git repo. Claude Code writes Python scripts that parse that data and compute the tax forms. Codex reviews the scripts with a high-reasoning model. I direct the process, check the results, and make the decisions. Between the three of us, the taxes are getting done. Two Agents, One Checking the Other The workflow has a structure that turned out to be more important than I expected. Claude Code with Opus 4.6 writes most of the code. Codex then reviews it. This isn’t a formality. Codex finds real errors. Here’s a concrete example. Codex ran a review of my 1120S scripts and flagged three findings, severity-ordered. The most serious: an incorrect Schedule L field mapping where Line 18 (ending) was being computed from total liabilities instead of other current liabilities. That’s the kind of error that’s easy to make and hard to catch by staring at a spreadsheet. Codex caught it by tracing the data from source JSON through to the generated CSV. The review process is often multi-turn. Codex flags an issue, Claude Code fixes it, Codex reviews again. They go back and forth until both agree the code is correct. It’s not unlike a code review between two experienced developers, except it happens in minutes instead of days. The Photoshop Problem One of the unexpected benefits is having a QuickBooks expert available whenever I need one. I don’t spend enough time in QuickBooks to know it well. It’s the Photoshop problem: you know the software can do what you need, but you only use it a few times a year, so you never build fluency. Every session starts with fumbling through menus, trying to remember how reconciliations work, wondering what changed in the latest update. With Claude Code, that friction disappears. I don’t need to know what’s under every dropdown in the QuickBooks UI. I describe what I need in plain English. When I needed to reconcile my business credit card, I just said “Let’s reconcile -7301.” Claude Code knew that -7301 referred to my business credit card, told me what data I’d need, and gave me the full pathname to the proper credit card statement in the repo. No fumbling. No searching. This pattern generalizes far beyond QuickBooks. Any complex software you use infrequently imposes a recurring tax on your attention. You lose time re-learning the interface every time you sit down with it. AI agents eliminate that tax. You describe what you want to accomplish, and they handle the how. What I’m Actually Learning Since 1996, I’ve had my CPA prepare both my business and personal taxes. That’s thirty years of someone else understanding my tax situation better than I do. This year is different. By working through the process with AI agents, I’m learning how my own taxes actually work. Not because I set out to learn — the learning is a side effect of directing the process. When you have to describe to an AI agent what needs to happen with your Schedule K-1, you end up understanding your Schedule K-1. I found errors I made in 2025 in how I managed certain aspects of the business. They weren’t catastrophic, but they were real, and I wouldn’t have caught them if I’d just handed everything to a CPA as usual. I’ll be much less likely to make them again in 2026. What’s Next: Plain Text Accounting Once taxes are done, I’m planning to replace QuickBooks with Beancount, an open-source double-entry accounting system that uses plain text files for its records. That sounds crazy. I know. But consider what actually matters now. QuickBooks is designed for human eyeballs: menus, forms, reconciliation wizards. Beancount is designed for both humans and machines to read: structured plain text that lives in a Git repo, diffs cleanly, and is trivial for an AI agent to parse and modify.
Reported result
The author reported that Codex caught real mistakes, including an incorrect Schedule L field mapping where Line 18 was computed from total liabilities instead of other current liabilities. No measured outcomes, such as time saved or cost reduction, were provided. The author noted learning tax mechanics as a side effect of directing the agents.
Human role
The original post does not say how a person stayed involved.
Limitations
The source does not state specific limitations, costs, or integration hurdles. Reliability depends on the author's self-reported experience without independent verification. Human approval is critical for final decisions, as the agents only assist in computation and review. Proposed advice: validate all tax calculations with a qualified CPA before filing.
Try it
I need to prepare S-Corp tax filings using AI coding agents. My financial data is in [replaceable: QuickBooks export path] and bank statements are in [replaceable: bank statement folder]. Please set up a private Git repo structure. Use Claude Code to write Python scripts that parse this data and compute the 1120S forms. Then, use Codex to review the scripts for errors, specifically checking field mappings like Schedule L Line 18. Iterate between the two agents until they agree the code is correct. I will direct the process and make final decisions. First, perform a dry run on a small sample of transactions and show me the output for confirmation before processing the full dataset. Do not send or file anything without my explicit approval. If inputs are missing, ask me for clarification.
Source reports are not independently verified. Publication date is the catalog publication date, not the original article date.