This shouldn't be ignored in the discussion here:
The job performed by the humans was broader than what was requested of the model in this benchmark: humans also had to find the relevant invoices (searching through mailboxes, or requesting them from providers) and reason through any circumstances which cannot be inferred from the bank feed and invoices/receipts on their own. In the benchmark these circumstances are presented to the model as “user notes."
And I would be pretty nervous about asking any of the frontier LLMs to retrieve invoices: "cool, Claude logged that it found the May 6th bill from the paper supplier, I am sure it didn't just make something up arbitrary, then compound on the error by agentically iterating over the made-up invoice lurking in its reasoning traces. I checked the first 30 times and there were no problems!"
If and when a large number of companies blindly turn over their accounts payable workflow to some AI agent system, it'll be very interesting to see the "social engineer the LLM" methods that fraud people use to get money sent to them. Basically the same idea as the ancient "send a fax with a bill for an unsolicited delivery of copier toner to 30,000 businesses" but taken into the modern era.
edit: There's already a number of LLM which are intended for outgoing data loss protection to redact or prevent PII from escaping. Is anyone specifically working on a training set and agent that is specialized in reviewing "is this legit to pay", as a sub-task or filtering step in an AP workflow? I suppose it's a GIGO problem, as it would work best only if you have suppliers enrolled in some kind of existing db, with a specific contracted format for invoices, and correlating with project numbers/cost codes.
I think it would be absolutely insane to hand over a serious-sized company's books to an LLM.
As a small consultancy though, looking forward to my next filing, and having just moved to a new and better-specced jurisdiction, I'm sorely tempted to outsource to Claude.
I've had mixed experience with accountants in the past. No horror stories, but I often feel I'm not getting everything laid out clearly, and that I don't fully understand the process.
I've got plenty of reasons to dislike LLMs in my own work, but when dealing with well-scoped but professionally gatekept things like tax or property transactions, they're an absolute godsend.
i specialize in invoice related analytics and business processes. AI can be useful for data extraction and saves me some typing time. even so, i dont blindl trust it because it sometimes makes very reasonable mistakes because invoices, quotes, and POs are sometimes structured in very informal ways. people misuse the lines, sublines, totals, and other data fields. they are often technically incorrect but when you look at them you know what they mean. i am not sure how to hand that off to something that has plausible deniability to guess even if it doesnt know
sometimes details are just in notes at the bottom and are applied to selectively applied to line items. sometimes the charge doesnt exist anywhere on the bill but there is an understanding (due to a separate agreement) of additional charges to be paid as a result of the invoice.
taxes sometimes are or are not explicitly stated
tariffs sometimes are or are not explicitly stated
when things are not explicitly stated or line item'd, they will usually still appear in the invoice total. so you have item 1 - $500, item 2 $500. total: $1300
At the end of the day invoices are often part of an ongoing communication / conversation between two organizations and they are created with an assumption that a rational and reasonable human who is in the loop with that conversation is going to read it.
You can fix this simply by using normal controls.
That's why we have purchase orders that can only be entered by buyers. Product is received and approved by buyer. Invoice goes to accounting, who can't approve it unless there's a matching purchase order and receiver.
Yes, letting agents do whatever they want leads to disaster. But humans are gullible stochastic token generators as well. And that's why the problem is already solved.
Indeed so, a fairly mundane RFP, RFQ, buyer, receiver, accounts payable process will stop a lot of problems. If an agent is inserted at some stage in the process with a clear path to make a ticket/escalate to a human if it sees something it doesn't understand, the risk isn't absurdly high, in my opinion.
I've seen so many reports of humans with the authority/ability to execute an outgoing SWIFT transfer who've been social engineered into sending money to fraudsters... Or even just the basic low level "Hey I'm your boss sending you an SMS, please go buy some gift cards and scratch them off and send me the codes". No AI involved whatsoever.
The danger exists where some true believer AI evangelist type of management person tries to fully automate the entire purchasing and AP workflow, which I'm sure some people will attempt soon, with varying degrees of success.
I have seen the SWIFT thing happen for $100k. I think AI could actually be better for this, because it's often easier to implement hard rules for the AI.
With the SWIFT incident I saw, there was a rule that no payment can go to a vendor's bank that isn't a current, approved vendor. But the rule was not enforced in software: it was an internal accounting rule that humans were supposed to follow. The AP person "thought it had been approved" because there was a similar transaction with a different company that was a new vendor at a similar time. The other transaction was legitimate, the fraudulent spoofer wasn't. The wire got sent to a party in China.
With AI agents, if you approach it from the perspective that it will be gullible and trickable by fraudsters, you build in these hard guardrails. With humans, it's much easier to believe that "we trained Lucy on this procedure" will work in all circumstances, even if Lucy still has the technical ability to bypass the official procedure.
In these cases, it starts looking a lot more like traditional software, with your little AI chaos monkeys constrained in little boxes within the software chain.
Another thing that seems to be disturbingly common these days is some party involved in residential real estate transactions having a RAT on one or more PCs, and/or compromised email account, intercepting an email at the closing stages of a house purchase and sending someone information to send a wire transfer to a fraudulent location.
But the hard rules only work up to the point that there is an exception. See my other post above. Occasionally a very senior person (hopefully senior) has to approve a payment that is outside of the rules, because it is something the rules did not anticipate (and now they are hardcoded into software and can't be changed).
I run a team that includes people that do this kind of work using ocr software that matches invoices to po's. No AI needed. This is a solved problem. Why are there people involved? Because sometimes the invoice and po don't match. For instance, price on invoice is higher than po, refer to buyer. Buyer is sick, supplier puts you on hold, no parts for your factory, lose millions... Would you trust an AI to choose what to do next? This might get referred to me to resolve and make a decision, not just on the facts available, but on other facts I can discover, and years of experience. I might end up making an unauthorised payment, would you give an AI that power?
Sounds like Enron was just a bit early. They could have just blamed the AI system instead of becoming a meme with "Enron math"
With AI you can scale the protection against social engineering. Where with humans you have to start from scratch each time and they are more likely to mess up.
My wife (head of accounting for a small business) has been working on automating large parts of her job using AI.
It's not completely reliable and the human cannot be taken out of the loop, but the number of menial tasks she's been able to automate has been really cool. A lot of processing data that arrives in non-standard formats, generating documents based on that data, etc.
She still has to review everything, but her workload is way down, and when her assistant quit she automated away his whole position.
Hahaha non-deterministic accounting probably won’t fly well with the IRS
That made me laugh, thanks.
I remember talking to my accountant in the UK a long time ago when I was newly self-employed, asking if I could pass something off as a business expense that was sort-of-related, but I knew probably not really OK. Her reply has stuck with me ever since: "HMRC [the UK equivalent of the IRS] are interested in matters of fact, not interpretation."
As an accountant at a large Corp, I can tell you that there are cases daily where two senior accountants argue over how to interpret accounting and tax rules, and often each agrees that the others interpretation is valid. I regularly see rulings from area specialists, and have to challenge them, only to be told, 'well in that case, you would be right'
I sometimes ask an AI to comment, and usually their answer is couched is ifs and maybes, just like the humans.
I often consult with auditors who tell me x is wrong, and they go away agreeing that y and z would be valid too, and x is also fine. There are often second order effects that need to be thought through.
Good luck AI
This is the problem with the majority of AI push coming from devs. They look at something as complex as accounting and only think in terms of what they see on the surface and then say “oh that looks easy enough, there couldn’t possibly be more under the hood”.
Yeah, my query was an open-and-shut case, I know there are grey areas too. The Arctic Systems case in the UK was an interesting one for IT contractors.
Do you think human accountants are deterministic?
Get a large enough org and watch your accounting grow an error margin.
Accountant here. It suprised me when I joined the field, but profit is calculated using a range of accounting estimates. Each accountant will make different decisions. Not least about which accounting period something belongs. Imagine a factory with freight inwards. It is month end and I have a sheaf of bills from various freight companies, but which ones are missing, not received yet? I can't wait, I have 2 days to report, so I make an estimate...now imagine that I have perhaps 100 such things. I may have to justify my estimates, but how should I estimate it? Same as last month? Perhaps I know there is a lot of shipments so I make it at the higher end.
Now I have a product failure at a major customer that I may have to send free replacements for. Should I recognise that cost now, or not? The accounting standards say if I know about it, and I can measure it and there is a high degree of certainty then I should. But the method of choosing is up to me. £10m or 1m cost, and this year or next... and I get to decide.
We bought a £6m dollar machine to be depreciated over time, but how long. The machine lasts 10 years, but will we still be using it then? Do I capitalise the internal R&D work as well?
All of which is audited, but the auditor often only has the information I give them.
Accounting is not deterministic, within certain bounds it is very subjective.
Ha even if this was true (it’s not) you’re basically saying “humans will make some mistakes so let’s throw caution to the wind” which is probably the worst application of AI that I’ve heard yet.
>Ha even if this was true (it’s not)
You haven't met many accountants i see.
Regardless that's not what he is saying.
If there is an acceptable margin of error for humans, we should be able to measure it for AI, and once AI is within those margins then it should be feasible to replace the human.
Ah yes, because there’s no difference between a shifted decimal and entirely hallucinated credits and debits.
This was my CPA wife’s response btw.
The entire field is based on balance sheets and context that informs compliance.
Let me put it this way: it’s such a bad idea that even Intuit doesn’t let their AI replace a human.
Why would Intuit be the barometer here? They're old technology. Pilot.com's got AI bookkeepers and they love them.
Man I hope so because they’re the ones selling them haha.
The "beatings" delivered by IRS will be deterministic for sure!
The tax code is so complex that business taxes are already nondeterministic.
Hey, the author of the benchmark here.
The benchmark data was prepared in April 2026 (when I was manually doing our VAT return with my co-founder). The invoices were indeed found manually.
Currently we're using a custom "invoice searcher" built on Kimi 2.6 (in our testing several weeks ago it outperformed Opus 4.7; it was just more persistent).
Ultimately, I still verify everything manually after the model is finished fetching invoices for the month -- but it's a great help to have all the invoices already found (usually correctly).
adam, i'd like to get in touch and would love to run the benachmark with mixedbread as a search backend. we are doing this right now with a lot of compliance companies. would be very curious how it improves quality/cost e2e
Sure, I'm at adam@vineyard-finance.com
AI is a glorified calculator in this case. It frees up human to do other things. It shouldn't be viewed as human replacer.
AI helps automate things that didn't already have rigorous formatting and structures available as input... and that's really all it does (99% of the time).
Doesn't matter how many more nines you add, rigorous formatting is still required. In some cases, it has teeth with compliance standards. Those standards cannot be compromised because there are already a lot of other layers contributing inaccuracy. It all adds up.
In most situations, you could just hire a junior dev (or an intern! remember those?) write some CSV scripts and call it a day. Cheaper and auditable too. Those scripts can't change anyway until standards are revised.
I'm still not seeing the benefit outside of solopreneur efforts and shady businesses wanting to launder blame.
> doesn't matter how many more nines you add
I don't get this argument. People say the same about autonomous driving.
But humans also have some number of nines. If you can get it better than humans, that's better!
Manual data entry and other tedious chores are definitely unreliable. However, running a script that a human wrote according to committee specs is the most reliable part. You're conflating the different aspects of human work. We are much better at understanding our needs and arguing about them than doing the manual part.
So, I don't get your argument either. I hear yours often enough and so much louder that I feel it's a deliberate muddying of waters.
What cannot be obsoleted by automation becomes bureaucracy. To my ears, it sounds like you're afraid of ending the tech wild west. That bureaucracy was always the most valuable part, and the demand for experienced programmers over at that table is very high.
> And I would be pretty nervous about asking any of the frontier LLMs to retrieve invoices:
I watched an accountant YouTuber reviewing a new AI-driven personal finance app the other day (I really need to touch grass), and it started out just fine. He had seeded the account with a bunch of his data and was able to ask questions about which categories had the most spend, etc.
About half a dozen questions in, he asked it to calculate a certain segment of his spend (and being an accountant, he had his numbers memorized), and he immediately got back a calculation that he did not expect. So he asked for an itemized response and it hallucinated line items that never appeared in his account data, which he pointed out to viewers. He followed up with the chatbot with "where did line item X come from?" and the bot acknowledged that it wasn't legit. He immediately noped out after that, and who could blame him?
[flagged]