Dayjob: Tax – evaluating AI agents on 50 professional tax assignments
surgehq.ai170 points by landonxi 7 hours ago
170 points by landonxi 7 hours ago
I prepared my taxes this year using Astra + Codex. It did a thorough job on a complicated return, and I sent it to my accountant to fill out the tax form, with everything laid out in a spreadsheet. They sent it back for final review, and I had Codex check it; it was riddled with errors. The refund was nearly $2k lower than it should have been, and they put a bunch of stuff into the wrong fields. Had to go back and forth multiple times to get it 100% correct. Made me wonder how many previous years' returns were wrong.
I feel like we're going to get to a point where it's mandatory to use AI for certain tasks - humans just can't be trusted to get things right that require a level of precision.
I sent my business and personal returns through Claude this year before signing, and it turns out they were a horror show of unforgivably stupid mistakes. Pretty infuriating considering what my now former CPA was charging me.
He amortized the goodwill of businesses I acquired over 5 years instead of 15 (incredibly basic stuff, which Claude flagged with *CRITICAL ERROR*). When I asked about it, he said he had caught the error in his review but forgot to save after he fixed it. I don't know if there's a good response to making that kind of a mistake, but if there is, it's definitely not that one.
As a proof-of-concept, I've been using extra tokens to formalize the IRS tax forms and worksheets, mostly just to prove this can be done. The idea is that something like this could be adapted into a publicly maintained version of the IRS Direct File program (RIP). I'm up to an estimated 60% coverage of US federal tax situations so far.
That's neat, what aspect(s) are you formalizing? That a given tax statement is correct?
Not quite. The formalization covers the IRS's published procedure, i.e. the questions and calculations that go into the published forms, worksheets, and instructions. It's not a formalization of the tax code itself, and it's not a tax calculator, though a client of the underlying code could function as a tax calculator. (The project includes such a client, a CLI tax interviewer, just to exercise the code.)
In more detail:
- every line of every form gets a YAML representation which includes the source text, a data type, and an expression built from a small set of lean atoms, plus a citation pointing to the exact location of the rule in the original official document.
- underneath the yaml layer there's a lean formalization of the atoms showing that they behave as specified, that all comuptations use only the allowed atoms and operations, etc. There's also a typescript evaluator that's used to cross check the lean for agreement, so we have the arithmetic represented two ways.
The critical failure point for an LLM-produced formalization like this is the interpretation of the form text. Since I'm not a tax lawyer and don't intend to become one, this is an explicitly owned weakness and one I've structured the project around. Publicly available tax calculators (PolicyEngine, UsTaxes, OpenTaxSolver, etc.) are used to cross check derived numbers from this system. Mutation testing ensures the test suite is robust against bugs and arithmetic errors. Interpretations are cross checked by different model families (Claude, GPT, Deepseek). Etc.
The idea is to derive confidence from the formal properties of the system itself, and to allow theoretical future users of such a system to derive confidence in the interpretation of specific forms by making it easy to cross check line-level formalizations against the source documents.
There's a lot more detail in the (alas, Claude-authored) README. When I have time I'm going to write my own version of the README expressing my vision of the project, but I'm not there yet.
Nice. Is the lean used to formalize formulas?
Sort of. The formulas for each line live in the form YAML I described above. The formulas are built out of a set of primitive operations that are defined in lean, with their properties proved there (non-negativity, monotonicity, etc.). There's also a lean evaluator that's run on compiled form data and cross checked with the typescript evaluator. (There's a lot more detail in the README.)
That's very cool but don't you think the Big AI labs will soon have collected enough data from us that they can easily do this?
I'm not sure it matters. The IRS killed Direct File under lobbying pressure from the tax prep industry. In my opinion the public is owed a free version of this sort of tool. If the IRS doesn't produce it, then citizens ought to. If Google or Anthropic or whoever wants to offer a free version of this, more power to them, but I don't see why they'd bother. They could have done it decades ago.
And to be clear: my vision here isn't "have an agent do your taxes", it's to have agents build a hardened and thoroughly verified deterministic tool that can help you navigate the (somewhat bewildering) forms and instructions decision tree when filing your own taxes.
This is great, and I appreciate the approach. Yes, tax calculation should be a decision tree. Unfortunately it is truly pretty complicated (just this year I had to abandon FreeTaxUSA for TurboTax because the latter handled an important case I needed better). But I have a lot of faith that we can do this right. Good luck!
I have a probably stupid question: given that the IRS killed Direct File, even if your software can create ready-to-file (or even ones that need some work) tax returns, isn't that moot without a way to like... get the IRS to accept them?
If your software can fill out the forms then you can (at the very least) print them and mail them in. The IRS certainly still accepts that.
https://www.irs.gov/filing/where-to-file-paper-tax-returns-w...
Why wouldn't the IRS accept a tax return created by your own software? How would they even know it was custom software? Ultimately your tax return is just a stack of papers that gets sent to the IRS.
Software sending efiles to the IRS has to be tested with and approved by the IRS. The IRS is working toward eliminating paper filing as an option. At this point, paid preparers are legally required to efile all returns and if the return gets rejected and they must paper file, they have to include Form 8948, Explanation for Paper Filing. For the IRS to accept a tax return from software the software has to go through this process. https://www.irs.gov/e-file-providers/how-tax-preparation-sof...
You can use free file fillable forms to submit online. It's a bit of work to transcribe from other software. It also requires a particular interpretation when filling forms (e.g. you need to round before summing your W-2s).
I think that assuming work can be done more easily in the future should not stop people from doing work that needs to be done in the present.
(If a painter knew that the camera was coming out I would not want them to throw their canvas away)
> (If a painter knew that the camera was coming out I would not want them to throw their canvas away)
That's not a fair comparison because the painter produces different work, whereas tax computations will be the same.
I have a non-trivial tax situation, as a US citizen living abroad. I prepare my own tax returns, never having found a cross-border tax preparer I trust, and have for the past few years, fed my draft return along with all the supporting paperwork into Claude. It has:
1. found errors, e.g. where I used the incorrect exchange rate (and managed to work out how I had made that error, which I was impressed by)
2. found items which would appear to be errors except for context it did not know
3. flagged areas where it disagreed with me as to the correct tax treatment, but accepted that the law is ambiguous (and in some cases I preferred its interpretation after additional research and in other cases I did not)
4. flagged areas where it disagreed with IRS guidance, and only acknowledged said guidance after I provided links to it
5. provided highly inaccurate estimated tax calculations, and not been able to fix its estimates even when challenged
So my takeaway is that it is a very useful tool for tax, but perhaps unsurprisingly, it is not good enough to dump all your paperwork into and expect a shiny return you don't need to check. (Possibly it is if you just have a W-2 and no other substantial income, but I can't speak for that scenario.)
I have a company abroad in addition to living abroad. It’s kind of a nightmare to do my US taxes and costs me easily $2000 per year right now. I was hoping I could have an agent do them after I have a few years of data and point it at the source statements and the end result forms
Similarly, had Claude look at my past returns and found mistakes in my records.
I had it prepare an analysis and then found a professional to confirm everything and then handle doing the amendment.
Definitely useful as a second pair of eyes.
> I have a non-trivial tax situation, as a US citizen living abroad.
Same situation as you. I have a local accountant in the country I'm in plus my US accountant. Claude/ChatGPT have been a huge help for checking, explaining, and finding issues. They also found websites with pertinent information that Google did not, which surprised me.
I'm also in a non-trivial tax situation, GPT Sol did better than my specialist accountant this year. It needed a lot of steering from me and about 2B tokens, but that is better than spending $3-4k in preparation fees. Total time to prepare and handle steering was still less than I normally spend on tax prep with my accountant.
I think what helped is giving it 100+ pdf and csv documents of my full situation so it knew everything about me and could cross-reference with all my bank statements.
So, GPT Sol outputted tax return PDFs that you were able to submit to the IRS? (I'd imagine that you couldn't directly e-file like you could from TurboTax.)
I also had a great experience consulting LLM regarding registering a business as a foreigner. Except I'm still very sceptical since it's not my area of expertise and consequences for a mistake are very high.
https://claude.com/marketplace/connectors/aiwyn-tax
Try this, it uses a tax engine as the backbone and provides feedback for Claude if it goes off the rails. Albeit it doesn't support every tax scenario.
I appreciate the recommendation, but as it happens I have already solved the estimated issue (mostly by just preparing a full return with estimated next year's figures, which while time-heavy is going to be more accurate). And for what it's worth, I suspect it wouldn't work for me - cross-border is extraordinarily complicated because of the tax treaties (indeed there is a dedicated IRS form to disclose the effect and rationale of any treaty based position), and given there is no obvious note on the documentation for it that there is cross-border support I doubt there is.
(And as I said to another commenter, I think it would be advisable in this thread for people to expressly acknowledge or disclaim any interest in or prior experience with any particular product)
check out https://www.multiplierholdings.com/
it's the most ai-native tax experience for US citizens abroad
Are you saying I should use this (or it looks like one of their portfolio companies) as my return preparer, or as a replacement for Claude my return-checker? If the former, I would want to know why I should trust it; if the latter, I would want to know why I should pay for their services given I am already playing Anthropic for a subscription plan anyway.
(It would also help to know if you had a financial interest in or prior experience with this company)
I find TurboTax is easy enough to use, except that data collection and entering is a huge hassle (to me, merely logging onto each website and getting the corresponding tax doc is painful chore). What would be a good way to collect the data easily? If I solve that problem, tax filing will become an easy task.
My tax accountant mis-filed my taxes a couple years ago, which cost me about $20K (which I learned after I switched to another accountant later). Don't know if Claude / GPT would make the same mistake or not given SOTA today, but feels like there's an opportunity for models to be saving me thousands or tens of thousands of dollars a year.
>mis-filed my taxes a couple years ago, which cost me about $20K
What do you mean by "cost you $20K"? That your tax liability was calculated by the accountant to be higher than it actually should have been? Or that the accountant calculated your liability too low, which was caught and billed by the IRS?
In the first case, you can still file for a refund within 3 years (I'm omitting some technical details), but in the latter case, as mentioned by others, you still owe the tax whether it was correctly reported on your original return or not.
Years ago I had a similar issue. Accountant did something incorrect, increasing my liability by around $50k. Tax date came so I paid it, but moved accountants. New person fixed the issues, refiled, but it took a long time to get the refund back. IIRC, the IRS did include some nominal interest for the time they had the money.
Sounds like an opportunity which could end up costing or saving $10k and which may be difficult to evaluate.
Someone still needs to take liability for mistakes.
The real issue is that accountant most likely has a track record of those kinds of mistakes - but you have no info on other clients. So you weren’t as informed at the time you chose them vs the new one.
People misunderstand the concept of liability. Yes your tax professional is liabile for fraud or gross misconduct, but if they make a mistake which results in unpaid taxes or fees you're still the one on the hook.
Typically you are on the hook for paying the actual taxes and penalties, but if you acted in good faith on the negligent advice of a tax professional many jurisdictions allow you to recover the costs of penalties and interest accrued as a result of the negligent advice.
Yeah, I looked him up just now actually, and he has a 3-star rating on Yelp... Which I think is pretty low for accountants?
I think the red flag should have been when he printed out all my tax returns and made me pick them up in person, and didn't even send a digital copy.
In this case accountant DID NOT take liability for mistakes.
It's not the accountant that paid $20k for his mistake but the accountant's client.
i'm a bit unclear who has the liability even when you use a CPA. accountants are humans that sometimes make mistakes.
The IRS made it crystal clear to me long ago. The tax filer is responsible for everything on the tax return, regardless of the preparer. Of course, accountants can be liable in other ways, but the filer has to pay the fines if the accountant makes a mistake.
From the IRS' point of view, the tax-payer is responsible for paying all of it. You can then turn around and sue to recover any of the fees/fines or penalties incurred as a result of an error made by the accountant, not the actual taxes though.
> Someone still needs to take liability for mistakes.
I am not aware of setups available to me that would transfer liability onto accountant.
So in either case I am taking liability.
(also, I am pretty sure that if corporation loses 10 millions over accountant making bad decisions, they are not going to recover that funds either)
yeah i think liability ultimately lies on the individual. of course, you could sue your accountant, but that feels hard, maybe almost as hard as suing the model providers.
At least in Poland it is not viable. Friend's business lost noticeable amount of money because accountant has not filed taxes and lied that they filed and lied that they checked it.
And there was no real path to recovering funds, except small part, after long court case.
At least they were able to fire them without severance pay.
This seems to be usual. Accountants are liable for crimes like embezzlement, but tax authorities hold individual tax payers responsible. If tax payers want to recover losses they have to go to court, which is often risky because accountants will say the client didn't do A, B, and C.
My obligatory comment about my experience with tax accountants:
https://blog.nawaz.org/posts/2022/Nov/do-you-trust-your-tax-...
I'm also 0 for 4 on tax preparers / tax accountants, so I concur with you its best to just do it yourself, which is what I now do. Only time I ever got audited (and failed) was when I had employed an accountant, since then I've never had a problem including passing audits. The tax code isn't /that/ complicated for most people's situation, and it's well worth understanding it.
I'm 1 for 6. And my current firm is overworked. I'm moving to do it myself (three companies, fund admin and personal). Recently Found a Tax-Accountant, not a CPA, who does advice only, no paperwork, who I'm using to cross-check my AI tools.
I am a CPA with 20+ years of experience and I feel like I need to defend our industry here.
There is no such thing as "failing" an IRS audit. You either under reported income, overstated deductible expenses, or your "accountant" changed the numbers you provided and committed fraud, which would have resulted in immediate legal proceedings by the IRS against the preparer. So did you "fail" because your preparer committed federal tax fraud or did you underreport your income?
The statement that the one time you used an accountant you got audited implies that you are confusing causation with correlation. There are about 6 ways that the IRS selects returns for audit but only 3 are actually relevant here.
- Information Matching (ie. if your W2 does not match the W2 submitted by your employer). This mostly used for your basic W2, 1099DIV, etc. kind of taxpayer and is usually just a letter from the IRS telling you your numbers are wrong and you need to amend your return.
- If they are examining a related return (ie. you own a business and another partner in that business is being audited)
- Computer scoring which breaks down into the Discriminant Function System (DIF), basically, how does your return compare to similar returns, and Unreported Income DIF, what is the potential that your return underreported income. This is what is commonly used for businesses and self employed individuals' Schedule C. Basically, when you prepare a business return (1040 Sch. C, 1041, 1065, 1120, 1120S, or 990) you have to report your business industry's NAICS code, and provide a single line explanation what your business does. Using this and the return data, the IRS is able to score a return on the likelihood that profit is understated by comparing a businesses income and expenses to the norm for a similarly sized businesses in that industry. High scoring returns are then reviewed by IRS personnel.
One of those three things happened to you and none of them are the fault of a preparer that is using accurate information provided by the taxpayer.
Finally, you keep saying things like "accountant". You disparaging an entire industry without saying what type of tax preparer you used. Was it a CPA, licensed by a state with experience in tax? An Enrolled Agent, licensed by the IRS directly? Or a tax preparer, an unlicensed person preparing your return either on their own or through something like HR Block or Turbotax? Because if you used an unlicensed tax preparer and are blaming licensed individuals that is like hiring an unlicensed electricians and then when your house burns down from an electrical fire, turning around and saying all electricians are bad
Additionally, "the tax code isn't /that/ complicated"... tax law is so complex that it is the only form of US law in which ignorance of the law is an acceptable legal defense.
In his defense, he did say "... for most people".
And that's true if all you do is have W-2 income, and RSUs/ESPPs. ESPPs were the most complex thing I had to deal with until I stupidly decided to do things and get K-1s.
I wasn't able to quickly see what tools/system-prompts were provided in the tasks. [0]
These types of detail-oriented professional tasks seem like they would perform better with task-specific harnesses. The example they give, of Astra missing the property excluded gains requirement, seems easily fixed by scaffolding ensuring that requirements are compiled first, then checked.
If the scaffolding is absent, is this benchmark testing that a model can self-scaffold, or track requirements in-context?
0 - https://huggingface.co/datasets/surgeai/DAYJOB-tax/tree/main...
I'm not sure LLMs will be good at doing taxes any time soon primarily because they're very bad at even simple math unless it's explained in words like in proofs. A good illustration of this is asking an LLM something like this: "If I put a container on a food scale, tare it, add 100g of peanuts, tare it, weigh 100g of almonds and then remove the container from the scale, what does the scale read?" LLMs completely puke all over themselves on this one. They do not account for not taring after the almonds and they do not see the weight of the container needs to be subtracted. The correct answer is minus (100g + container weight). Although this particular question also exploits another LLM weakness which is physical world concepts. LLM are just guessing. For an LLM to do taxes well, it would need to be trained on many scenarios and it would need a math service that it can use as a support service.
You just need to give an instruction something along the lines of "any initial calculation or derived value must be computed through the use of a $LANG program."
Pretty sure the LLMs could implement tax calculations better than the average accountant, were they given the rules.
In my experience agents will reach for numpy for any non trivial math without being instructed
It's the judgement of "trivial" where I find issues.
Most calculations are trivial.
It is also trivial to specify deterministic tool calling for all math
As a veteran tax CPA, I think LLMs can be very good at preparing taxes but I would not trust them to complete a tax return start to finish. An LLM that can take data from tax documents and input that data into the correct boxes of tax preparation software has no need for math, it just has to identify data in a pdf, spreadsheet, or handwritten note and transcribe that data to the correct box in professional tax software that then does the math. In your typical local CPA firm, an LLM could replace the unlicensed preparers/interns/early career stage CPAs in filling out the return for the experienced professional to review and adjust according to their specialized knowledge. As with many industries, I think the LLMs are suited at replacing entry level automated turks but lack the skills necessary to fully complete the project.
“The scale reads −(100 g + the weight of the empty container).”
I literally gave it your question. What models did you test?
There is currently a bill in Congress to finally allow regulation of paid tax preparers. Not sure if addresses AI, but probably not.
https://www.accountingtoday.com/news/new-bill-calls-for-irs-...
There is also an IRS form (14157) to report a return preparer complaint, which of course does not include an option to report a faulty AI.
I'm surprised Sonnet 5.5 did noticeably better than Opus and Fable. Same pattern in their other benchmarks.
It didn't really do any better, the difference is well within the error bars for their benchmark.
I like their benchmarks. They ask the right, not easy questions: https://surgehq.ai/blog/sudo-l7
Has anyone verified their figures to see if they hold up?
I had Gemini help me with RMD calculations for an estate I manage. Advice was mostly good. Wrong about withholding for both US and CA. Missed deductions on US. Was accurate on MFJ and spousal death impacts.
I've found it so far more effective for preparation than for planning/advice.
I've got a plan to feed the laws and rules and instructions into the machine before the prompts - is RAG still cool? Maybe I should small-language this?
I wonder if you use opus 5.5 and let it research using the internet what you would get. I use opus 5.5 for legal stuff and it's generally great, but you do have to give it access to the web.
Not related but I hope QuickBooks learn something and make their software better.
I downloaded 1700 of my transactions and use AI to build an app for it, so I can analyze the transactions and make sure that everything is properly categorized. It used to take me whole day because of the Quickbooks. It took me less than an hour to go through all the transactions.
As a counterpoint, I ask the big llms every month the same set of tax questions about the last US tax bill and they still can't give me the correct answers.
I started my career fixing the mistakes that my clients' former tax advisers made. It looks like the next generation of tax lawyers will get to do the same except this time the former advisers will be llms.
What is human baseline? How it stacks against 10% LLM accuracy?..
Sharing all my financial info with LLM companies is not something that I like.
EDIT: good point that paying accountant does not ensure avoiding it
wonderful, you made a good point that make me dislike situation even more
Your accountant sure will.
I'm for sure reinventing the wheel here but I've had an LLM generate an app that walks me through my existing tax routine I've been running with for years. Thus no real numbers go to any LLM I just give it my situation and it fills out the form.
The plan now is to retrieve the current year's tax forms then have an LLM compare them to stored versions from last year, and look for variances (e.g. new field, new form requirements, etc) all of which an LLM would excel at (whilst basic math it historically did not).
Worst case scenario what does OpenAI do with your tax return?
- their data will be partially leaked in general, including my financial info
- someone who hates me personally will get access to it (see case where someone put their ex-spouse on no-fly-list etc)
- leak my private info via new version of LLM was trained on it
- sell it to everyone interested (maybe sold after their bankruptcy)
- give it to NSA and whoever else
from obvious ones
Useless benchmark as it seems to be evaluating literally only the weights.
i want to always have a fast way of reviewing my tax returns without having to hassle my accountant or hire another one.
In Germany, there are companies like Buhl data that let you do your tax return very easily and give a live estimate of your return based on the data you provided. From my experience over the past 5 years, its prediction error is less than 5%.
Article I hoped to read: Evaluating AI agents on 50 professional tax evasions
Ah, yes, the 'dude trust us' bench.
So are the accountants completely fucked yet?
Not yet, and not completely.
Some accountants need to make judgement calls. LLMs are not great at many judgement calls yet.
Kinda points to another thing... that our tax code is insanely bad.
[dead]
[flagged]
[flagged]
[flagged]
[flagged]
[flagged]
[dead]
[flagged]
I used Grok Bot and found out that I could offset some stock value going down and get back ~$300. I gave it my bank statements, email access (which had my stocks monthly report as attachment) and a few other things.
I'd have honestly never been able to do it myself - its just too tedious.