AI agents flunk real-world tax work: best model passes just 10.8% of professional assignments
Dayjob: Tax – evaluating AI agents on 50 professional tax assignments
Surge AI's DAYJOB: Tax benchmark puts AI agents through 50 realistic tax assignments across eight practice areas, complete with messy client files, handwritten receipts, and conflicting records. Claude Sonnet 5.5 leads at a 10.8% pass rate, but even top models miss consequential issues—like a deadline that could cost a homeowner a $250,000 tax break, or an unverified marital status that changes retirement account division. The benchmark tests professional judgment, not just calculation.
A tax adviser should recognize consequential issues even when the client doesn't know to ask about them.
- lexandstuff
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.
- RGS1811
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.
- waiwai933
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.)
- bluejay2387
Is this just giving a general-purpose agent a workspace full of documents and an assignment, then seeing what happens? I feel this doesn't really say anything. It appears they mostly dumped all of this in base models with little to nothing in the way of infrastructure to support it. If you just randomly grab some person off the street and say 'do taxes' they would probably mess up just as badly. The interesting question would be what the models could do in a well engineered harness focused on taxes. That would be reflective of the performance boundary and how serious we would have to take AI in that field.
- holdingunsteady
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.
- hintymad
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.
- licyeus
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...
- bodeadly
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.