AI is biased toward big companies

I built a feature recently and asked an agent to set up the deployment infrastructure for it. What came back was clean and sensible, and all of it was on AWS, the managed version of everything, the names you would put on a slide. Then I asked it to cost the design out for the load I actually expected. A thousand dollars a month.

For this feature, that was the end of the conversation. There was no margin left in a thousand a month.

So I kept going. Same load, same feature, a different mix of providers. Twenty dollars a month.

The first answer was not wrong. It was built for a company I am not. Ask an AI for the common way, or the best-practice way, and it gives you the answer that is best on average. On average means for the kind of company that produced most of what it read, and that company is big. It has a platform team, a procurement process, and a bill where a thousand dollars a month does not register. You may have none of those things.

This is worth understanding, not just working around. An agent’s default is the centre of mass of everything it has seen. On infrastructure, that centre of mass is large companies writing about how large companies operate, so “set up the deployment” resolves to the managed, defensible, expensive shape. The advice is sound. It is aimed at a reader who isn’t you.

Deployment is the easy example, because the gap shows up as a number. The same pull sits inside the answers that don’t. Ask for an architecture and you get one built to survive a scale you may never reach. Ask for a process and you get the ceremony that makes sense at two hundred engineers, not at six. Ask which tool to use and you get the one nobody was ever fired for choosing. None of it is wrong. All of it is sized for someone else.

The fix is not to trust the tool less. It is to stop asking for best practice and start handing over your constraints. The load you actually expect. The budget that has to leave margin. The size of your team, and the things you have decided you are not going to do. Give the same agent your numbers and it will find you the twenty-dollar answer, because now it is solving your problem instead of the average one.

The cheap answer is not a free lunch, and the honest read says so. Twenty dollars a month pulls operational work back onto your own desk. For a small team with more time than money, that is often the right trade. Later, when you have more money than time, it may be the wrong one. The point is to make the trade on purpose, looking at your own numbers, rather than inherit a default that already made it for a company you’re not.

So push past the first answer. The default is the median of everyone else’s situation, and a small team sits a long way from that median, which is exactly where the default gets expensive. Best practice is a real thing. Just make sure the one you ship is yours.