I Did My Taxes with Claude. Here's What Happened.
It’s tax day — so, shoutout to my tax friends. I hope you get some sleep tonight.
I got mine done. Twice, actually: once with Claude, once with TurboTax, then reconciled the two side-by-side to the dollar. I was curious how AI would handle a workflow we all know — and one that comes with the full force of the organization that brought down Al Capone.
Quick context: I’m a former individual tax guy from PwC. I can read a 1040 and write a reasonably informed prompt. This isn’t a “Claude does your taxes for you” story. It’s a “here is what AI in finance work actually looks like when it works” story — and why that matters far beyond April 15th.
The workflow
It started with me dropping files into a folder. Wage forms. 1099s. A brokerage summary. Blank IRS forms and their instructions. Claude would classify each document, extract the numbers, and write the figures to a plain-text working file. I could watch it work — no black box, no “AI magic” — just a terminal reading PDFs and writing to files I could open and inspect.
Then the data pool grew. Bank exports. Credit card statements. Order histories. Each export got dropped into a folder; Claude extracted the 2025 rows, categorized them, and asked follow-up questions when things were ambiguous. The output was a growing picture of my year in flat files I could audit.
We built a couple of small skills along the way — a workbook generator that produced a spreadsheet version of the return with live formulas so I could trace every number, and a signature-capture page for signing PDFs without a printer. This wasn’t Claude running wild. It was me telling Claude what to build and then using it.
What it got right and what it got wrong
Claude caught things I would have forgotten about. It noticed a passing comment about health insurance and told me to grab a specific form — which unlocked a credit I would have skimmed past. It flagged a deduction strategy I wouldn’t have considered, derived from reading the actual instructions. I pushed it on that strategy and learned a lot in the process.
It made mistakes, too. I tried a skill to fill the forms directly, and while it was pretty good, the biggest miss was that it checked “Married Filing Separately” instead of “Single.” The math was right; the box was wrong. Probably fixable with better prompting on my end — but good thing I caught it before anything got mailed.
The documentation might be more valuable than the return
While Claude was gathering data to fill the forms, it was also building a complete audit-defense file as a side effect. Every expense traced back to a credit card line. Every business deduction had a written rationale next to it. It even suggested categories of documentation I hadn’t thought to include.
TurboTax might be able to do something like this in a few years — which is eons in this age. But the real value of AI here isn’t the form-filling or some disconnected prompting. It’s that it can connect to everything. Your email. Your calendar. Your receipts. Your bank feed. It assembles the whole paper trail and writes the explanation as it goes. If I ever get audited, the file it built is the difference between a short conversation and a long one.
Two paths to the same number
I ran it through TurboTax anyway, for two reasons. First, TurboTax is a filing engine — it submits to the IRS. Claude isn’t there yet. It could theoretically file via the free-filing portals and browser automation, but I’m not willing to bet a return on “it clicked the right button at the right moment” until there’s an actual filing API — and I’m not holding my breath for one. Second, when you’re about to sign a significant refund under penalty of perjury, it’s nice to have two independent paths land on the same answer.
The reconciliation is where the real learning happened. TurboTax and Claude agreed on almost everything. Where they didn’t, the difference always taught me something — either Claude had missed a nuance or TurboTax had skipped a deduction Claude picked up from the prompting. By the time the two matched to the dollar, I had a return I was confident in and a mental model of every number on it.
If you’re reading this on April 15th
Here’s the narrow version of this workflow I’d trust Claude with today, end-to-end, without a TurboTax side-by-side: filing an extension.
Most people don’t realize this, so it’s worth saying loudly: an extension to file is not an extension to pay. You still have to estimate what you owe and send the check by April 15th. Miss that, and penalties and interest start the same day. The extension just buys you six months to finalize the paperwork.
Which means to file an extension properly, you have to do your taxes anyway — pull together income, estimate liability, pay a number. That conservative estimate is squarely in Claude’s wheelhouse right now. Drop your documents in a folder. Tell it to be conservative: round up on income, round down on deductions, assume the worst on ambiguous items. Ask it for a completed Form 4868 with a payment number, then file the extension by phone or drop it in the mail (certified, return receipt).
Now imagine doing this every 30 days
Next year I’ll probably trust Claude with the whole return — either I’ll have sharpened the prompt or I’ll borrow someone else’s.
But tax prep has a built-in ceiling: you only do it once a year. The prompt never gets to mature. You learn something, file, forget most of it, and by next April you’re starting over.
The workflows that compound into something a finance team can’t live without are the ones you do every month. A close. An AP run. A reconciliation. A reporting pack. Every time you run it, you sharpen the prompt, tighten the skill, add a check. Twelve reps a year, not one.
Strip a tax return down and that’s what it is: a miniature monthly close. Extract from sources. Categorize. Reconcile against a second opinion. Produce an auditable trail. Sign under personal liability. The monthly close is the same job, at 12x the frequency — and that’s where AI stops being a neat trick and becomes something you don’t operate without.
What made this work with Claude is what’s still missing when finance teams try to do AI at scale: primary-source anchoring instead of model memory, auditable flat files where every number traces to its input, reusable skills that encode how the work gets done, side-by-side reconciliation where the differences teach you something, and a human asking good questions that neither side knew to ask. Almost nothing on that list is the model. The model is the easy part now. The hard part is the scaffolding — the thing that turns an answer into something a pro will sign their name to.
So — is this better than a human?
Better than me, for sure. And certainly better than the 21-year-old version of me at PwC on two hours of sleep on April 14th.
But the thing that really shifted for me is this: my tax preparer this year also had the sum of human knowledge at its fingertips. It answered theoretical and strategic questions between breaths — “how aggressive are these deductions?”, “what’s the audit risk if I take this one?”, “what should I be doing next year to minimize this?” These are questions you would never dream of asking an overworked accountant in April, because they don’t have time and you don’t have the retainer.
AI is here. Just having specific knowledge isn’t a valuable skillset anymore, and being good at entering things into an interface is solved. This year I’m glad I did it twice. Next year I’m willing to bet I’m doing it all through AI — and filing it too, on paper if I have to, if I can stand the line at the post office.
