In July, Jane Street lost about $15 billion in a single month. One of the most profitable trading firms in the world, first losing month in roughly a decade. The reporting from Reuters and the FT points at a concentrated position in AI-exposed assets that a lot of other people were holding at the same time. When it moved, it moved against everyone at once.
I'm not writing this to talk about hedge funds. Almost no Australian small business has a trading book or 400% leverage. But the shape of that failure is worth borrowing, because a lot of small businesses are quietly building the same shape into their operations right now. One vendor. One model. No fallback. No plan for the day the terms change.
You depend on more than you think
Ask an owner what the business runs on and you'll get app names. Xero. A CRM. Some AI tool that drafts quotes or summarises client calls.
The actual dependency usually sits a layer below that. It's the specific behaviour of a specific model, reached through a specific vendor's API, at pricing and availability that vendor controls completely. Nobody signed up for that. They signed up for a monthly subscription and a login.
Things that can change without asking you:
- The model gets retired. The one you tested and trusted stops being offered. The replacement behaves differently in ways nobody documents properly.
- Output quality drifts. Same prompt, same input, different answer six weeks later. This one is nasty because it's silent, and it's harder to catch than an outage. You usually find out through a customer.
- Pricing moves. It has mostly gone down over the last two years, which has been great for everyone. There's no law saying it only goes one direction.
- Rate limits or regional availability change. Suddenly the thing that ran fine at month-end queues up and times out.
- The vendor pivots or gets bought. Or decides your use case is now their product.
Model retirement notices are routine now. When Anthropic pulled Fable 5 overnight following a US government directive, customers who had built workflows around it had no fallback. If your quoting process runs through one model, a retirement notice isn't an email, it's an operational problem with a deadline attached.
Where the concentration actually hides
Three places, and only one of them is technical.
In the tooling. Everything calls one provider directly. Not inherently wrong, but if the vendor's SDK is scattered through forty different scripts, spreadsheets and Zapier steps, you don't have a system, you have a dependency with no seams in it.
In the workflow logic. This is the one I see most and the one people underrate. All the actual business knowledge, how you price a job, what a good client email sounds like, which exceptions get escalated, lives inside prompts typed into some SaaS product's web UI. That's your intellectual property sitting in a text box you can't export. Change vendors and you rebuild it from memory.
In one person's head. Usually the owner or one keen employee who set the whole thing up on a weekend. Nobody else knows how it works. That's a bigger single point of failure than any API.
Five things worth doing, cheapest first
1. Write down the manual fallback. For each AI-assisted process, answer one question: if this stops working on a Tuesday, what do we do for the rest of the week? Sometimes the honest answer is "nothing much happens, we do it by hand and it's annoying." Fine. That's a process you can leave alone. If the answer is "we can't invoice," that's the one to work on.
2. Get your prompts and rules out of the vendor's UI. Put them in a document you own. A Google Doc is enough. It costs an afternoon and it's the difference between switching vendors in a week and switching in a quarter.
3. Keep your data somewhere you control. If the transcripts, extracted fields and generated documents only exist inside the tool, you're not a customer, you're a hostage with a subscription. Check the export function actually works before you need it. Most people never test it.
4. Build a small test set. Twenty real examples with the answers you know are right. Invoices, emails, whatever the tool handles. When the vendor changes something, or when you're evaluating a replacement, run the twenty and compare. This is the single highest-value thing on this list and it takes about two hours to assemble. Without it you have no way to tell whether a model update quietly made your output worse.
5. Put one thin layer between your business and the provider. In practice this means the model gets called from one place in your code, not forty. Swapping providers becomes a config change instead of a rewrite. This is a build decision, so it's cheap if you do it at the start and expensive if you retrofit it.
What this doesn't mean
It doesn't mean running two vendors in parallel. For a business with twelve staff that's a waste of money and attention.
It also doesn't mean waiting. Every few months there's a headline about an AI selloff or a bubble, and somewhere an owner uses it as a reason to defer a project that would have paid for itself in four months. Market sentiment about AI company valuations tells you almost nothing about whether automating your reconciliation process is worth doing. Those are unrelated questions. A drop in CoreWeave's share price doesn't make your accounts payable process any less manual.
The point is narrower than that. Know what you're depending on, keep the exit cheap, and don't put anything you can't afford to lose behind a single vendor's roadmap.
The question to ask before you sign
When you're evaluating an AI tool, ask the vendor: what happens to my data and my configuration if I leave? Then ask for it in writing.
The answer tells you a lot. Vendors who have thought about it will show you an export button. Vendors who haven't will talk about how nobody ever leaves.
OrionX builds and integrates AI systems for Australian businesses, with a bias toward setups you can maintain and leave. If you're mid-way through an AI rollout and want a second opinion on what you're locked into, get in touch.
