This is the first in a weekly series about putting AI to work in a small business. We are starting here because almost everyone follows the same arc, and the middle of that arc is where people quit. If you know the shape of it in advance, you can skip the part where you throw the whole thing out.
The spike
Everyone starts somewhere. You open Claude or ChatGPT, you ask it something real, and it answers well. Then you ask it something harder and it answers that too. Within about a week you have read enough headlines and seen enough demos to believe the strong version of the story: that this thing is a supercomputer that will shortly run everything.
That is the spike. We call it the AI high. It is not that the tool is bad, it is that the tool is extremely good at making you feel like you understand more than you do. It is confident in exactly the same tone whether it is right or wrong, and early on you have no way to tell those apart.
The most common version of this is software. People decide, sincerely, that AI has just made them a full stack developer. And you can get surprisingly far. You will scaffold an app in an afternoon and it will run. Then you reach the database layer, or the auth layer, or the part where real data has to survive a schema change, and the whole thing comes apart in a way you do not have the vocabulary to debug. That is not a failure of effort. It is the point where the fluff runs out.
What comes after the spike
After the spike you start seeing the holes, and there are real ones. This is the moment that decides whether AI becomes useful to your business or becomes a thing you tried once. Two things happen from here. Either you level out, keep using it, and get genuinely faster. Or you drop off, decide the whole category is overhyped, and go back to doing it by hand.
The people who level out are not the ones who found better prompts. They are the ones who learned where the tool is weak and stopped asking it to do those things.
The three places it falls over
Math
This is the big one, and it is the subject of next week's episode. AI is a creative tool. It is designed to be random, which is exactly what you want when you ask it for ten ways to phrase an email and exactly what you do not want when you ask it to reconcile a column of numbers. Randomness and arithmetic are opposites. There are ways to work around this, and they are worth doing, but the workaround is never “ask it more firmly.”
Ideas versus real systems
AI is genuinely excellent at ideas. You can sit with it for an hour and generate a hundred directions, and it will tell you how smart each one is along the way. That feels like work. It is not quite work. The gap shows up when you try to turn one of those ideas into something that runs every Tuesday whether or not you are paying attention. Ideas are cheap and the tool makes them cheaper. Systems are the expensive part and the tool helps less than you expect.
Context
Context is the least understood of the three. Every model has a fixed budget for how much it can hold at once, and everything you give it spends from that budget. The file you pasted, the earlier turns in the conversation, the instructions you gave twenty messages ago. When the budget runs out, things fall out of the back, and the model does not announce that it has forgotten. It just quietly starts answering as though the missing part never existed.
This is why a long conversation gets worse over time rather than better, and why the fix is usually to start a clean thread with a tight summary rather than to keep pushing.
The 70% rule
Here is the frame we will use for the rest of this series. Treat every answer as roughly 70% right.
You are working with something that is a genius and an idiot at the same time. It does not know your business. It is a tool.
Seventy percent is genuinely useful. Seventy percent of a first draft, seventy percent of a research pass, seventy percent of the boring parts of a document is an enormous amount of time back. Seventy percent of a payroll calculation is a disaster. The number does not change based on how important the task is, which means the judgment about where to point it has to come from you.
And it will not tell you which 30% is wrong. It will hand you the whole thing in the same even, competent voice. If you are not enough of an expert in the subject to catch the bad third, you will ship it.
So what do you actually do
Nothing dramatic. Keep using it, and change what you use it for.
- Point it at drafts, not decisions. First passes, summaries, rewrites, options. Anything where you are going to read the output carefully anyway.
- Stay the expert. Use it hardest in the areas where you would notice a mistake immediately. That is where the leverage is real.
- Assume the confident tone means nothing. It is a property of the tool, not a signal about the answer.
- Watch the length of your threads. When quality drops off, that is usually context, not the model having a bad day.
The people who get value out of this are not the ones who believed the most. They are the ones who got past the spike, kept a short list of things the tool is reliably good at, and built the boring structure around it. That structure is what the next few episodes are about.
Next week: why your spreadsheet is the worst possible thing to hand it, and what to do instead.
