Here is a small story about modern software, told in three acts: a premium product that silently didn't work, an investigation that fooled us before it enlightened us, and a stack of legal documents explaining — in beautifully lawyered prose — that none of this is anyone's problem but ours.
We think it's worth telling, because the punchline isn't "haha, bug." The punchline is the operating model of the entire AI industry right now, and it has consequences for how much of your business you should hand to it unsupervised.
Act I: The Product That Wasn't There
We run a good chunk of our internal tooling on Claude Code, Anthropic's coding agent. Anthropic recently released Fable 5 — positioned as its most capable model, the top of the line, the thing the credits we'd paid for were largely *for*. We opened the model picker, ready to use what we'd bought.
It wasn't there.
Not greyed out. Not "coming soon." Not accompanied by any error, banner, or log line. Simply absent — a paid-for capability that quietly declined to exist for us while, judging by changelogs and community threads, it cheerfully existed for other people.
Eventually, buried in one code path, we found the closest thing to an explanation the system would offer: the model was *"restricted by your organization's settings."*
Now, we *are* the organization. We hadn't restricted anything. There was no admin console with a checkbox, no policy we'd set, no documentation describing what this restriction was, where it lived, or how one might un-restrict it. As error messages go, "restricted by your organization's settings" ranks alongside "an error occurred" and "please contact your administrator" in the genre of sentences that are technically language but contain no information.
The real cause, when we finally dug it out, was two undocumented local mechanisms stacked on top of each other: a stale cached account state from *before* our premium credits had landed, and a local model allow-list that didn't include the new model. Neither is documented anywhere we could find. No official page says "if a model is missing from your picker, check these two things." The failure mode of a flagship product from a company at the frontier of machine intelligence was: silence, followed by a misleading sentence, followed by archaeology.
Act II: How We Fooled Ourselves (An Honest Interlude)
Here's the part most vendor-critical posts leave out, and the part we're most keen to keep in: while investigating, *we* got it wrong too — and we got it wrong in an instructive way.
Early on, someone had a hunch the missing model was a geographic restriction — available in some countries, not others. To test it, we spun up a cloud server in London and opened the model picker from there.
The model appeared.
From our normal location: gone. From London: there. Same account. It looked like proof. Clean, repeatable, satisfying proof. We wrote it down as proven. We even drafted an article about it.
It wasn't proof. It was a confounded experiment. The London machine didn't differ from our local one by *one* variable — it differed by several at once: different IP, yes, but also a fresh install with no stale cache and no old local config. We had changed three things, observed one outcome, and credited the variable that made the best story.
When we later went back and did it properly — isolating one variable at a time, on one machine — the geography theory collapsed. Clear the stale cache and fix the local allow-list *without* touching the network route, and the model appears right where you are. No London required. Our "proven" conclusion was false, and we disproved it ourselves.
Two lessons from that, and we'd rate them equally:
First, on us: a test that *feels* rigorous can still be worthless if you changed more than one thing. "It reproduced twice" is not the same as "the variable I blamed is the variable that mattered." We know this. Everyone knows this. We still did it, because a clean-looking A/B result is seductive. If your team uses AI tooling and debugs weird behavior, assume your first satisfying explanation is confounded until you've isolated it.
Second, on the vendor: we only *had* to run experiments on a paid product because the product offered nothing better. No meaningful error, no documentation of the gating layers, no diagnostic. When the observability of a premium tool is so poor that customers are reduced to renting servers on other continents to test folk theories about why it doesn't work, the customers will sometimes reach wrong conclusions. That is not a customer failure. That is a product externalizing its debugging costs onto its users — who, as we're about to see, have contractually agreed that this is fine.
Act III: The Fine Print, or: You Agreed to This
So: a premium product silently failed, gave a misleading non-explanation, and burned days of investigation. Surely, one thinks, there's some accountability mechanism. Some warranty. Some—
Let's just read the documents. These are real quotes, from the current public terms of the three biggest AI providers. We checked them so you don't have to (though you should).
Anthropic (Consumer Terms): the Services and Outputs are provided *"on an 'as is' and 'as available' basis"* and *"without warranties of any kind, whether express, implied, or statutory"* — explicitly disclaiming, among other things, *"fitness for a particular purpose, … accuracy, availability, reliability."* And in case that leaves any doubt about whose problem a malfunction is: *"your use of the Services, Materials, and Actions is solely at your own risk."* Total liability? Capped — on the order of what you paid in the preceding six months.
OpenAI (Terms of Use): *"OUR SERVICES ARE PROVIDED 'AS IS.'"* (The capital letters are theirs; in legal documents, shouting is how you make the important disclaimers extra binding.) They *"MAKE NO WARRANTIES"* and *"DO NOT WARRANT THAT THE SERVICES WILL BE UNINTERRUPTED, ACCURATE OR ERROR FREE."* Aggregate liability: the greater of what you paid in the last twelve months or one hundred dollars.
Google (Terms of Service + Generative AI terms): Google commits to providing services *"using reasonable skill and care"* — and then clarifies that this, plus service-specific terms, plus whatever the law refuses to let them waive, are *"the only commitments we make about our services (including … their reliability, availability, or ability to meet your needs)."* On the AI side: *"The Services may sometimes provide inaccurate or offensive content,"* so please *"use discretion before relying on, publishing, or otherwise using content provided by the Services."*
Read those again, slowly, and appreciate the architecture. These are companies valued in the hundreds of billions, selling subscription products marketed as capable enough to write your code, run your agents, and reshape your business. And the contractual promise attached to that marketing is, in substance: *this product is not warranted to work, be accurate, or be available; if it harms you, that's your risk; and if a court disagrees, our exposure is roughly a hundred bucks.*
To be scrupulously fair — and clearly labeled as fact, not opinion: this is standard software-industry boilerplate. Nearly every SaaS product ships with an "as is" clause. Lawyers would call it unremarkable.
And now, clearly labeled as *our opinion*: for this class of product, at this price point, with this level of marketed capability, it has become absurd. There is a widening gap between the front page and the terms page. The front page says: *the most capable model in the world, ready for your production workloads.* The terms page says: *not warranted to be accurate, error-free, or available; use at your own risk; liability capped at pocket change.* Both pages are written by the same company, about the same product, and only one of them is legally binding. Guess which.
When our paid-for model silently vanished behind an undocumented gate with a misleading error, every second of the cost landed on us — the investigation, the wrong turn, the cloud server rented to test a theory the product's own diagnostics should have made unnecessary. That allocation of cost isn't a bug in the terms. It *is* the terms.
The Conclusion We Actually Act On
We're not writing this to tell you to stop using AI tools. We use them daily, heavily, profitably. Claude Code, the very product in this story, does real work for us — most days, excellent work. That's precisely what makes the fine print worth staring at: the better these tools get, the more tempting it becomes to hand them things no one is contractually accountable for.
Put the three acts together and the shape is clear:
1. The product can silently fail, with no meaningful explanation, for reasons documented nowhere.
2. Even competent, motivated users investigating it can reach confidently wrong conclusions, because the system is a black box that offers no better path.
3. And when any of this costs you something, the provider's carefully written position is: *as is, no warranty, your risk, here's $100.*
That is why our operating rule — humans approve, agents execute — is not caution, and not technophobia. It's the only posture that matches the actual, contractual reality of these tools, as opposed to their marketed one. An autonomous system built on components that are (a) opaque, (b) undocumented at the failure boundaries, and (c) formally disclaimed by their makers is not a system you let approve its own actions. The vendor has told you, in writing and in capital letters, that responsibility lives with you. Believe them.
So we design accordingly. Agents do the work; humans hold the approval gates on anything irreversible, external, or expensive. Not because the agents are bad at their jobs — they're increasingly excellent — but because when something goes wrong at 2 a.m., the terms of service will not be joining the incident call.
Trust the tools to work. Verify that they did. And keep a human hand on every switch that matters — because the companies that built them, by their own signed admission, keep none.
---
*Facts in this piece — the error message, the cache/allow-list cause, the quoted terms — are real and were verified against the providers' published terms at the time of writing (July 2026). Characterizations like "absurd" are our opinion, and we've labeled them as such, which is more than the front pages do.*