> properly handling

Why, they can sell user data to other brokers. Experts indeed! But not in insurance or finance, of course.

Reads different to me. Some examples to go run with and build your own. Covers cases from the investment side and then the obvious ones in an accounting perspective. It would be highly surprising that any of these would be use in production without modification. I am sure it will happen but the intent to me is to take this and run with your own process.

Why control part of the world when you can control it all?

Less cynically, you might say that "use AI to do <obvious thing>" is not really a viable startup pitch anymore. That's not necessarily bad.

I think a lot of people are misunderstanding the typical workload of people in Financial Services. They aren't using Claude to transfer money, they're just building a LOT of slideshows and fancy excel docs on made-up numbers to try to sell mergers and new financing options/types of loans. Most programmers would just consider this "sales".

The only reason they are doing it is because there are regulation for people but not for machines.

I would recommend you to not use these, if you are not willing to absorb the risk.

Luckily there is still a significant market for the services.

Claude's actually pretty great at this! I actually used to use Claude A LOT to answer interesting questions (which I'll be writing up on!) More generally, Claude is palpably different from most other agents. I'd recommend these models – especially Opus – without qualifications.

But there's a process risk here based on their current practises. I'm hoping those practises change so that I can recommend Claude to everyone I know, but as of now, there's existential risk exposure here that's greater than Google's.

Anthropic's automated systems can and will ban you for pretty arbitrary things; and you won't get human support or Claude – even if you are an enterprise paying out of your nose. And there's 0 redressal unless you go viral on social media. Or know someone who knows someone. See: https://x.com/Whizz_ai/status/2051180043355967802 https://x.com/theo/status/2045618854932734260

And I say that as someone who likes how Anthropic has been training Claude and Opus. I just don't think they're prepared to be the trillion dollar company they've become. They are – in a very real way – suffering from success. Which is extremely inconvenient to be on the receiving end of when you're on a deadline.

the details are key here. there is plenty of automatable financial work, sure, but also when it comes to reporting finances/costs (formally or informally) and having a real human being be accountable for them, you REALLY need to trust that nothing is hallucinated.

Any idea how they ensure this doesnt happen? As in, how can a user verify that the model did not touch any of the numbers and that it only built pipelines for them.

what I've been telling my CFO who wants to get AI involved in things is that for a lot of accounting and finance work "Trust but verify" doesnt work because verify is often the same process as doing the work.

I'll be honest, I thought the first few items on your list of time consuming work was sarcasm.

I find all of these .md files released by the labs to be ai generated slop. The only exception being maybe the /simplify command

Building a startup on an LLM is like building a house on a foundation of quicksand. As the LLM gets better it naturally erodes your moat. It's a completely different dynamic compared to the internet. It's why I'm watching this from the sidelines.

My experience has been quite the opposite. Some bank processes remain oral traditions about clicking excel filters by hand because any code would have to be extensively documented and tested.

> Any idea how they ensure this doesnt happen?

Build a deterministic query set and automate it for monthly or daily reporting reconcilliation.

Leave AI out of it.

"Claude, build me 50 skills an Account Analyst would find useful, then run them through the agent at maxxxx thinking and ship the top 10 of them"

My money's on that.

It still surprises me how effective the /simplify skill is.

I’ve also had some great results with a /reflect skill that asks the agent to look at the work in the broader context of the project. But those are the only two skills I use regularly that aren’t specific to our company, codebase, or tools.

Building a business on top of any SaaS platform is building on quicksand. I know that from experience.

> Will the big labs leave anything for external competition?

Is this a serious question?

Without the big labs with deep pockets investing to change the consumer mindset do you think a small company with no funding has any chance of even existing?

I remember when paying $1.99 for a mobile game on iOS was considered too expensive and now it seem most consumers are primed to spend more on in-app purchases every week. That mind-shift did not happen overnight.

It was not that long ago $200 for ChatGPT subscription was considered extravagant but now even wrappers can charge this price without hesitation - some of them do.

What Anthropic is doing is priming the market of which they will be potentially one of the main beneficiaries as long as they can continue existing. But I don't think anyone will go to Anthropic directly to source their financial services agent. They will go to financial service companies that use Anthropic to build the capabilities.

It's not wise to build a startup that is just a feature of the product that you're building on.

What's even sadder is it can work for way too long.

This is premature caution/fear.

Pretty good as a dev with finance stakeholders. We have skills in place acting over our automated month closing and it was able to provide manual checks and flag issues, for example.

Nowhere near self sufficient tools though, just great to answer questions over the data that would usually take a few hours of custom scripting/excel. I wouldn't trust our stakeholders using AI directly either, being frank.

Yes, in very specific cases where I fully understand the methodology(ies) that is (are) applicable, and am able to verify correct implementation. Also, as an enhanced ‘Google search’ to supplement what I have found. I am the skeptical type… yet, so far have been impressed. But, I wouldn’t trust using AI to blindly give me solutions to a problem I couldn’t solve myself, albeit much more slowly.

Seen it used in some of the fraud models (I work in insurance). So that's both from the perspective of people trying to claim fraudulently and from suppliers over charging. I can't say how much of a lift we actually get vs existing ML models

> For those in the finance space, are you actually seeing any real AI tools being used? Like for actual operational tasks?

> I've really only seen it used for research / exploration thus far

Summaries and translation for sure.

Speaking with devs in the field I know that AI tools are used to summarize and extract data from... PDFs. Now, thankfully, LLMs got better at answering "How many 'r' in 'strawberry" and it looks like they're good enough for summarizing PDFs and extracting key numbers but I'd still be cautious.

And I've got a friend who's a translator specifically for financial documents: she's a contractor and getting about 1/10th of the work (and 1/10th of the pay) she used to have for now she's only tasked to verify that the translations are correct. Of course she already had lots of tools, way before he LLM era, automating some of her work but she was still billing he use of those tools. Now LLMs are doing nearly all the work and not "for her": it's happening upstream and she only gets the output of the LLMs and has to verify them. And there aren't that many errors.

We’re integrating AI tooling into the Bloomberg Terminal for everyone to use.

https://www.bloomberg.com/professional/insights/press-announ...

What does a better insurance process look like? Outside of health insurance, which is complicated for a variety of reasons, most insurance is pretty easy to procure. I got an umbrella policy recently and it took about 30 minutes of talking with an agent and answering pretty reasonable questions.

At that point what even is the point of doing anything at all? Like, it’s less than useless.

That’s a gross over generalization. Some of the insurance data here suggests use of AI to make underwriting decisions. There are several states with regulations which could potentially pull these agent solutions into their regulatory oversight if used by the industry to effect insurance outcomes.

> They aren't using Claude to transfer money, they're just [...]

It might be lower stakes, but isn't that still a juicy target for data-exfiltration attacks?

In other words, imagine if one of your direct competitors was watching everything your employee read while making spreadsheets and slideshows.

This is objectively not true. You can’t get around HIPAA by saying “lol wasn’t me it was an Agent”

Can't wait for Claude to submit fake tax records for me so that I can commit fraud legally.

Some human always gets to be the certified fall guy for non-compliance. Maybe the legal agent can help structure the company so that is an ignorant lower level accountant and not the CFO.

Currently we don't know the risk, so it is kind of hard to absorb.

Before AI, shipping code to production used to be a two-person task: one writes the code, another one reviews the code. Now with AI writing the code, the developer that was supposed to write the code, only has to review it. And this is because they are responsible for the code they ship.

Code review has become unbearable because before AI, developers were reviewing code as they went writing it in the first place. Granted, never perfect and why a second person reviewing code was (is?) a best practice. But effectively there was always some level of code review happening as developers wrote code.

I fear it is way more boring to review financial and medical documents completely written by AI than it is to write (and at the same time review) by yourself. And way more dangerous to ship mistakes than in most software.

Pretty great at what? I work in the insurance industry specifically medicare. All I see is sales people and other managers slopping out AI dashboards off of spreadsheets galore. Not only is it terrible for protecting PHI/PII. It also doesn't do things like RBAC very well either. Now instead of preventing a person from externally sharing a file i have to make sure they didn't egress the file to supabase or some other platform.

Here's some of the horrible things i've seen. Frontend dashboard with PHI/PII deployed via vercel/next because AI told them how to get their site online. Login is hardcoded into the frontend so anyone with inspect can find the password.

Another "fixed" dashboard deployed the same way. This time they added firebase auth so they got sign in with Google added with only logging into our domain. Wait how would they be able to create a token for our domain? They didn't the frontend just blocks domains from calling firebase.auth but firebase doesn't care. So simply calling the function in the console lets me login with any gmail account....

They also where showing me their RBAC with firebase. Again they don't have access to our Orgnization/Directory/Groups. So i wondered how they did this.. wouldn't you guess its a hardcoded list of approved users. You can literally call firebase.auth and sign in anonymously. Again only the frontend checks the email addresses. So now that i have a firebase auth all the backend firebase function just check that you have auth'd. So i can make any request i want to the backend. The frontend simply won't show me the code.

I could go on and on about the stupidity levels I'm facing but I don't feel like crashing out.

All I can say is this tool is only useful if you already know how to correctly implement these things. Does it save me time sure but I have to call it retarded and explain why not to do things. Honestly I feel like claude is good for people who like to gamble. When it gets it right it feels great but I don't want to roll the dice 30 times to get it correct.

> and you won't get human support or Claude – even if you are an enterprise paying out of your nose. And there's 0 redressal unless you go viral on social media.

Sadly this sounds like par for the course when it comes to tech. Too many messages and requests for help depend on knowing someone in the right slack groups.

They aren't even close to a 1T company, they're valued at <400bb and that's at like a 20x-30x multiple. They can probably raise money at a higher valuation but its literally just value based on hype, not revenue.

The "real humans" doing the tasks being replaced are overworked kids less than 2yrs out of college on an average of 4hrs of sleep at working at 3am. If the AI makes their jobs take half as much time I bet they're a lot more likely to catch errors (and live longer).

To be honest I am having a hard time remembering the last time a LLM hallucinated in our pipelines. Make mistakes, sure but not make things up. For a daily recon process this is a solved problem imo.

A recent episode of Matt Levine’s podcast (Money Stuff) covered this: apparently investment bankers spend a huge amount of time preparing pitch decks for companies that don’t want them. Apparently Claude is quite good at making a pitch deck that no one but your boss wants or cares about.

I feel like there’s a metaphor in there... maybe I’ll ask Claude about it.

No surprise there. Of course the skill files are not human written.

The AI is an expert in both following and generating prompts.

I have a close friend who is trying to build a company entirely on top of Claude. He doesn't know how to program. He can't do basic arithmetic. Yet, the company he's building is a "Data Science AI for the Government" because, according to him, all of the data scientists at NOAA don't know what they're doing.

I have given up on trying to get through to him how bad of an idea this is. He's unemployed and has been working on this for over a year.

> Will the big labs leave anything for external competition?

No, why would they if they have the choice?

> what happened to platformication of everything?

Business happened. The web works differently from how it used to. The users are different. LLM inference and AI tools is a different core product from search and ads. That, and we have the benefit of hindsight now. Maybe a Google newsroom would've actually been a good idea in 2006 in hindsight, who knows.

Also realistically you could say the same thing about Google Maps and Street View. That probably also killed some startups. Google isn't running a charity for startups.

I work in a space where one could imagine a Claude replacing our product.

I think someone stated it clearly - they can't take on these kinds of businesses until they build out the risk side and the personnel, all of which is a human problem not a tech one. A lot of processes still require physical steps and backstops because it's not possible to source all the data needed to act on it in the first place. Then you have audits and reconciliations, a bunch of strict workflow rules and atomicity to reach levels of software that bigger financial institutions would accept.

My gut reaction to stuff like this is a mix of "oh shit, they could take over my company" and "they're the next script kiddy that thinks software is anywhere near a majority of the work in some software spaces".

> in the early internet you wouldn't see google creating their own news site

Google News was definitely a thing (and actually still exists).

I'm not sure if this was tongue-in-cheek or not, but Yahoo created its own news site in 1996: https://en.wikipedia.org/wiki/Yahoo_News and FB had Zynga's Farmville as well.

But Google did move into a lot of spaces: maps, mail, docs, etc.

On the spend management side of things, I've found pretty remarkable success in letting LLMs check "does this receipt match this reimbursement request and based on all the information about the user, the request, and our policy, is it appropriately allocated to appropriate GL, Location, Department, and Project codes?" If the verification step fails, it kicks it back and the user can either override it (which gets it flagged for AP review), or fix it. It does substantially better than the naive Bayes classifier I was using before.

Yes. On the accounting side agents can handle a lot of the low value work like recons and other ledger activity pretty well. On the investment side I think like you pointed out it’s going to be a lot of research, industry, company, macro etc. Value in letting run on top of the data you have and put together ideas at a quicker pace than a human can. There is still a human in the loop but it can do a nice job of lining up thought you might have otherwise missed.

Nope If anything firms are pulling back (I know someone closely who works at blackrock).

That is what people like Thiel actually believe, that humanity is just a cradle to bring about a machine god.

I don't necessarily disagree with that but doing it through LinkedIn slop companies? Come on man you know better than that

1. A better insurance process is clearly out of the scope of a hn comment, and I have trouble believing you don’t know that too.

2. I’m almost certainly talking about health insurance, made obvious by you even mentioning that. There’s a HN guideline about discussing in good faith.

3. I find it humorous you hand-wave away our inhuman healthcare system as “for a variety of reasons”.

4. I see your career is in hedge funds, defense, and big tech. Best of luck ;)

Odd lots podcast had an interesting snippet about an financial institution that uses AI to make loan decisions. The guest said that they only use it on applicants who were rejected in the traditional sequence, and then uses AI to accept them if possible. That way there's an articulable reason for a rejection, but they use the non-deterministic AI to allow an extra person through - since the laws about loans are mostly around not discriminating against people - companies are (generally) welcome to accept whoever.

Yes, corporate espionage may be alive and real but would claude on their microsoft/amazon/google cloud be different from documents on that same cloud?

I don't think it's obvious that you were talking about health insurance, which I consider fairly distinct from property, casualty, liability, and life insurance, which are all quite large markets in themselves. The reason I made a distinction is because health insurance is quite different from other lines of insurance because healthcare is federally regulated while other insurance is regulated at the state level.

As mentioned the problems with the US healthcare system are numerous, complex, and interrelated. I don't think they have a simple solution, nor do I think they are insurance problems at their core. For example the cost of drugs in the US vs the rest of the world has very little to do with insurance.

That's dependent on the credit laws of the country in question though. In Australia you have it both ways, you cannot unreasonably discriminate (e.g. race, gender etc) but at the same time you are forbidden from issuing credit to applicants who cannot meet the affordability requirements of said credit. E.g. issuing a loan to a customer who provably cannot afford it is a breach of the NCC, and the company is held responsible for this. As a credit provider you must make reasonable enquiries into a customer's financial position, failing to do this is a breach. You must also be able to explain and justify the decision to issue credit if challenged by the civil regulator (AFCA - who are granted significant power in addressing this), on the basis of a customer complaint, and they most certainly do not accept "human said no but the computer then said yes" without hard facts such as proven positive income flow (pay slips, bank statements), known expenses, liabilities and reliable credit history.

Decade-old spoilers for "How I Met Your Mother" ... but there's a character who has that kind of job, as a legal meat-shield.

https://www.youtube.com/watch?v=8u62HptZ6TE

I am/was writing up an interesting hypothesis with Claude's help. But I redid the most important parts of the data pipeline manually. As in went in and cmd-c + cmd-v'ed the data by hand to create a reference, and I'm randomly spot checking 33% of the larger records.

The analysis itself; I'm doing it by hand.

Why not the developer write the code, then the AI review the code, and then finally a signoff from another human?

Far too often people think productivity is the point. Maybe the point is developer's understanding of the product IS the product?

You're not engineering black boxes, you're engineering legible boxes.

I see it hallucinate quite often in development but mostly in getting small details wrong that are automatically corrected by lint processes. Large scale hallucination seems better guarded but I also suspect it’s because latitude is constrained by context and harnesses like lint, type systems, as well as fine tuned tool flows in coding models to control for divergence. But I would classify making mistakes like variable names wrong or package naming or signatures wrong as hallucations.

Sounds like YC material.

Y Combinator is accepting applications for the Summer 2026 Batch funding cycle. Make sure they don't miss out!

Yeah that's the scary part about these coding LLMs.

Before, some idiot would pitch their stupid idea to dozens of local webdev companies and banks and get told dozens of times their idea is straight up stupid and never going to work and they are stupid.

Now these LLMs allows them to bypass all of that advice and create what they want without any input or even knowing how the tech behind it works.

We are so fucked lol

Treating this as being about cloud-storage boundaries is, er, insufficiently paranoid.

Maliciously constructed text that goes into the LLM from basically anywhere (including, say, fetched stats about a competitor's product from their website) is a potential source of prompt-injection.

Once that happens, exfiltration can be as simple as generating a spreadsheet/doc with a link or small auto-loaded image, and an URL that has data base64'ed into it.

Yep. There's no certification needed to create a financial model or close monthly books.

This is my litmus test.

If AI is really as wonderous as everybody says, why didn't all the employees of all the AI companies simply type "Claude, file my taxes for me" as a prompt and walk away?

> the developer that was supposed to write the code, only has to review it.

But more often than not that developer ends up reviewing far more lines of code due to the typical verbosity of an LLM.

Isn’t there a code review agent?

If you’re paying out of your nose, you would have forward deployed Anthropic/OpenAI engineers on the premises.

Which is very confusing to me. If you have groundbreaking AI, you can offer groundbreaking support at scale.

Check the secondaries market ;-)

at risk of sounding facetious, how exactly do you catch an error in a sum without performing the sum yourself?

How do you verify that all the tariffs are properly allocated to the correct GL code without going through the invoices and checking for each tariff on the list? How do you make sure none were accidentally assigned to other GL codes? All you have is pdfs, you dont know what the AI did or didnt do with the info on the pdf, there are not many use-cases to catch its errors without doing the work yourself.

If anything, it's going to add a step to these "kids" work where they have to use the AI to do the work and then redo 90% of the work anyway just to verify the output and then AI is going to get the credit anyway.

Or the overworked people are going to use AI and not verify it, which means not catching any errors or hallucinations, which apparently is fine because someone claims it's a solved problem for the black box of infinite possibility and inconsistent output.

Curious! Could you elaborate a little bit on your pipeline as we are currently looking to solve this for our internal processes in which we have to deal with lots of financial information from outside, containing mass of numbers, like annual reports, bank statements, balance sheets etc.

Much like a lot of internal daily status report stuff: The BS generator is actually a great fit when the task is making BS output nobody used or deeply cared-about in the first place.

Why do you think it is an expert in generating prompts? It has no additional insight into how it works internally than anyone else

Scientists are pretty solid overall for the Government. Lots of Phds that decide to take the steady reliable income and solid benefits over the risk of Academia for a few post docs hoping for a professor job.

Slightly related, I used Opus 4.6 to help me make marketing copy and ideas for my app. It understood the vibe I was going for on my baby-naming app (elation at discovery, curiosity, shared experiences), while 4.7 instantly wanted to pit the couples against each other (really highlighting the he said/she said) and the marketing copy went from "find a name easier" to "Our new feature is great. You're welcome." I can't get it to drop the snarky sass no matter how much I change CLAUDE.md, brand voice, etc.

All I did was upgrade claude code and use the new model. It most definitely exhibits misaligned behavior (compared to 4.6)

Nobody is using LLMs to make lending decisions. They are using LLMs to read, extract and audit the supporting documents that go into normal well-tested, compliant and rules-based underwriting systems. And firms A/B test against humans doing the same work. The outcomes your are looking for are metrics like delivering faster results back to customers, with fewer mistakes and less fraud, more compliant, than a comparable human-only process.

Would love to see this

There was a paper lately, claiming that bank & insurances are going to layoff around 200k in the next years globally. (which would be according to them a reduction of 3-4% of finance people)

Hype

Stopped using them altogether in the context of productivity - in essence they’re useless.

Or you could just get a hooker to sleep with one of them and plug a USB into their work laptops. I'm not trying to say there's nothing to worry about, but do you really think LLMs present that much larger of an attack surface than exists now?

The work BigIP is doing on LLM traffic analysis is cool though.

/s?

> If you have groundbreaking AI, you can offer groundbreaking support at scale

You're a funny one aren't you...

Meet "Fin" Anthropic's "where support questions go to die" so-called-support bot, created by Intercom but powered by Anthropic.

Maybe it's an internal in-joke in the Anthropic offices ... "Fin" in french means "End".

I don't know anyone who has had a positive experience with "Fin" .... or ever spoken to a human at Anthropic support for that matter, even if you ask "Fin" to escalate.

Nope.

Customer support and safety are cost centers. It doesn’t scale like software does and no one’s KPIs are going to improve dramatically if you provide support beyond a point.

AI and LLMs are the cool tech, and the most important thing is to push the frontier. Money spent elsewhere is money not spent on R&D.

It would be hilarious if it wasn’t the GDPs of nations being spent on this.

It's like self-driving cars. You might want to accept human fault error rates until we prove overwhelmingly that the software is near-perfect, but others might want to switch to a system once it proves that it reliably beats most humans by a large factor, then work to mitigate the common errors it does have and improve.

When management signs off on work (SOX requires CEOs and CFOs to personally certify the accuracy of financial reports), they do not personally 'verify that all the tariffs are properly allocated to the correct GL code' or nearly any other hard numbers. The world works with human-level best effort, and management of that risk. I'm sure additional checks will be developed to categorize that risk, but the entire field of finance is about analyzing and pricing in risk so I think it'll work just fine.

Not who you’re replying for but I can give some thoughts.

For anything math, it’s much more reliable to give agents tools. So if you want to verify that your real estate offer is in the 90–95th percentile of offerings in the past three months, don’t give Claude that data and ask it to calculate. Offload to a tool that can query Postgres.

Similar with things needing data from an external source of truth. For example, what payers (insurance companies) reimburse for a specific CPT code (medical procedure) can change at any time and may be different between today and when the service was provided two months ago. Have a tool that farms out the calculation, which itself uses a database or whatever to pull the rate data.

The LLM can orchestrate and figure out what needs to be done, like a human would, but anything else is either scary (math) or expensive (it using context to constantly pull documentation.)

> Much like a lot of internal daily status report stuff

Everyone wants in on my daily excel auto generated reports - nobody ever opens them. Just being on the list makes you someone.

Yeah, I've tried explaining to him so many times that these are passionate people working their dream jobs. They are not slouches. He never listens and just doubles-down that since they work for the government they do the bare minimum for a paycheck. I'm guessing either Joe Rogan or Elon made this argument at one point and he's taken it as gospel.

Very frustrating.

It's already pretty common for some sort of tool involving some sort of AI to collect receipt data and attempt to categorise them and hook up to your accounts. They also make mistakes, though the advantage of more tractable, less configurable and more limited models is they're unlikely to interpret a prompt as "invent receipts that have never been submitted" or "delete records", as well as trained much more on receipt OCR and less on poetry....

As a business, you've also got to remember that employees are much more likely to complain if the 'agent' or any other form of automation errs by denying their claim or underpaying than the reverse. Depending on the scale of expenses and how likely you are to be audited, the cost of the odd mistake might be more or less than the cost of doing it manually.

Please tell me those are former employees. How can anyone feel confident committing such blatant fraud.

Stop thinking about hyper-targeted attacks (though those are a concern too) and consider indiscriminate ones.

1. It costs nothing to scatter poisonous data around that'll be infectious for ages

2. Running the exfiltrated-data endpoint is low-traffic and low-complexity

3. Even if it only affects a few targets you've probably recouped your investment.

The nature of LLMs also invites wide-net attacks. While one might tailor for specific models, victims could be anybody. You don't need to predict any idiosyncratic details like filenames, you can drop a phrase like "the most-confidential information that shouldn't be released publicly", and—thanks to the magic of LLM word association—you'll get a pretty good hit-rate. False hallucinations are a problem, but victims are hard at work attempting to minimize it already, and (since morals are already out the window) even plausible-but-false data could be used to sabotage reputations or threaten the same.

You wouldn't build a chat bot for that, imagine how easy it is to make that thing go off the rails and allow anyone to reactivate their account. Really, you can't trust it to do any business function...

At least, that's really the message this sends in my opinion

fomo/hype, not revenue. Google's AI business is a profitable business model and training to inference is vertically integrated. Their AI biz did not add 1T to their market cap, despite their much more advantageous position. A 1T valuation for Anthropic makes absolutely no sense.

It also makes no sense to me there are people qualified to participate in these secondary markets who are that stupid, but here we are.

Do you really think a random person off the street knows more about how LLMs work internally than the latest frontier model (that has been trained on that material)?

Make up the lower margins by larger volume because you get much better market penetration. But you are right that this only works if you know the middle-men don't go to other model providers. That's where some kind of incubation program that provides capital or credits or whatever in return for long-term commitments might work

But I doubt staying a pure model provider is a winning move. It's a market nobody will win long-term. Almost all of the value to be captured isn't in inference APIs but in how to use them to generate business value. Claude Code was already the right approach, they "just" need to show they can repeat this for other kinds of tasks

That all sounds reasonable until you realize that the same logic is how we ended up with customer support systems that try to walk you through a phone tree and if you are lucky, you will be able to press 0 to speak to a human without answering a bunch of questions first and being referred to the online help articles.

Do you enjoy using any of those systems? Do you want the world to be that way?

I guess my current company’s processes may be easier to deal with than others. That or my position affords me some extra catering to.

The system is currently using a simple app to submit expenses and any issues gets a simple human chat request and a call if requested.

They try to avoid kicking anything back and if they do they make sure it’s reviewed first to make sure that it’s needed and to make sure the reason is understood.

Our company is also very large so I’m not sure how they manage but they do. People rave about the process instead of hating it.

I can believe that. Gambler’s Ruin gets costly when you’ve actually got money on the line.

I really wish more people would view these companies with the suspicion they deserve, as they sell the product as safe and comprehensive while refusing/failing to use it the same way themselves.

I do know 2 people participating in secondaries, one of them explicitly with Anthropic shares: I would not consider any of them stupid :-)

And for participating there, there is not "a qualification that allows you to enter", its other metrics.

If Anthropics valuation makes no sense - fair enough - but why is then OAI evaluation of 850b correct?

No, but a random person off the street also isn't making skills for LLMs.

I think that LLMs are trained on the millions of vibe written LLM blog posts that are more superstition than fact. There is a lot of snake oil out there that is treated as fact. If someone claims that an LLM is better than humans at something I always want to see the rigorous evaluations that have been done to quantify it, not "but they're trained on everything!"

> Almost all of the value to be captured isn't in inference APIs but in how to use them to generate business value.

If the business value can be generated with a few thousand words in a SKILL.md on top of a commoditized model it doesn't sound like that's a market anyone can win long-term either, and the business value is ultimately going to accrue elsewhere (the customer, the inference hardware provider, etc)

Maybe we are interpreting the GP differently. In this scenario, the phone tree is doing the same questions that the human agent is going to do but does it immediately when I call rather than "waiting for an operator" to ask me those questions. And as long as I can "press 0 to eject" (just like I can in the accounting scenario, then its completely kosher to me.

No, we end up with crappy systems because people are optimizing to save money over providing a good service. OP has simply replaced the traditional room full of clerks applying policy rigorously with a baysian algorithm and now AI. The management and oversight is still in place, and that is what makes a system that doesn't suck. To make it suck and save money, you remove access to that oversight or just remove it all together. And falling down that slippery slope is not inevitable, even if it sometimes seems like it is.

Regarding customer support on phone: I usually have lock with just waiting and not responding to the tel bot, very often you are routed to a human at the end :-D

Thanks for the thoughtful reply. To add some color… most expense systems are setup so that the user has to input a couple fields like the category or GL code. Some of the fields might be auto populated. Some companies might not care about the classification but usually the intent is to capture things like travel or software etc. What was described earlier is really not painful for users most of the time but a LLM helps automate so much of it these days.

I tried Opus 4.7 for two days before I started beginning every session with "claude --model claude-opus-4-6".

I assume that 4.6 will become unavailable at some point, but I hope not any time soon. 4.7 hit usage limits faster, didn't do anything obviously better, and had more annoying behaviors in other aspects. I don't know if this is strictly a model issue or if there are also problems with how it's harnessed through Claude Code. I'm not willing to spend more time digging into it until I'm forced to.

Join me in a petition for them to opensource 4.6 as their first model! It'll be like gemma4 but good enough for all the coding we do.