How to size an AI project before you spend a dollar

You don’t need a business case with forty tabs. You need honest numbers on four things: cost, return, timeline and risk. Here’s how to put a real figure on each before anyone builds anything.

Cost

Three parts, and people forget two of them.

• Build. Setup, integration, whoever does the work.
• Run. The monthly bill, which scales with usage. Get a real per-use number and multiply by real volume.
• People. Someone has to check outputs, fix prompts, and keep it alive. That time is a cost even if no invoice shows up.

Return

Be honest here or the whole thing is theatre. Time saved times what that time is worth, minus the time you now spend checking the AI. If a task takes ten minutes and AI makes it two, but someone spends three minutes checking, you saved five, not eight. Count it that way.

If the return only works when you assume the AI is never wrong, you don’t have a return. You have a hope.

Timeline

Longer than the demo suggests. The demo is the easy 80 percent. The last 20 percent, making it reliable on your messy real-world data, is most of the work. Double your first guess and you’ll be closer.

Risk

Ask what happens when it’s wrong, because it will be. Some tasks that’s a shrug. Some tasks it’s a lawsuit. Also: where does your data go, will people actually use it, and what breaks if the vendor changes their price or their model next year.

Then do the maths

Put it together. If the return clears the cost with comfortable room, and the risk of being wrong is low, go. If it only works on best-case assumptions, don’t. And if the honest answer is “this saves half a person a day a week,” that might still be worth it. Small and real beats big and imaginary every time.

The number doesn’t have to be precise. It has to be honest. That alone puts you ahead of most people spending on AI right now.