# If AI Can Do Anything, Why Does an AI Project Cost Real Money? Source: https://metabinary.ie/blog/why-ai-projects-cost-money [← Back to blog](/blog) AI Reality Check # If AI Can Do Anything, Why Does an AI Project Cost Real Money? By [Joseph Rose](/about), founder of MetaBinary. 20+ years building enterprise software. Built the world's first instant online metal-casting costing engine. Published 2026-10-16 More than 80% of AI projects fail by some estimates, and when [RAND asked experienced AI engineers why](https://www.rand.org/pubs/research_reports/RRA2680-1.html), the top answer was that people misunderstood the problem the AI was meant to solve. That is where the money goes: picking a question worth answering, connecting your data, building the tools the AI uses, and checking it is right before anyone relies on it. Running the model itself costs little. AI is amazing, but it doesn’t know what you want. Teaching it what you mean is key. ## The model is the cheap part Running a big AI model is cheap, and it gets cheaper every year. One question and answer often costs less than a cent. So it makes sense to expect an AI project to be cheap too. The catch is what the model knows. It has read a huge part of the public internet. It has never seen your quotes, your job hours, or your supplier list. Ask it which jobs ran over budget last month and it can only guess. Sometimes the guess sounds right, which is worse than a guess that is plainly wrong. So a model on its own is a strong engine with nothing hooked up to it. The cost of an AI project is the work of hooking it up to your business and making sure what comes out is right. ## Where the effort goes Most AI builds follow the same six steps: 1. **Pick the question.** Find one question that costs you time or money to answer today. Agree what a good answer looks like. 2. **Reach the data.** Find where the answer lives: the ERP, a spreadsheet, an inbox. Get safe, read-only access to it. 3. **Build the tools.** Write the small programs the AI uses to fetch your records and work out real numbers. The AI asks for what it needs, and the tools do the exact work. 4. **Test on real records.** Check its answers against ones you already know. Count how often it gets them right. 5. **Fit it into the day.** Put it where your team already works, and show them how to use it. 6. **Keep it running.** Your systems change, your data changes, and the models get updated. Someone has to keep watch. Steps 2 to 5 take most of the time, and none of them is about the model. They are about your business: how your data is stored, what your people mean by a word like “late,” and which mistakes you can live with. ## Most failures start with the wrong question The RAND researchers talked to 65 experienced data scientists and engineers. The most common reason they gave for failure was that the people asking for the AI and the people building it did not agree on the problem. A build can work perfectly and still answer a question nobody needed answered. I learned this building the world’s first instant online metal-casting costing engine. Working out a price from a part’s shape and metal was the easy part. The hard part was getting years of estimators’ know-how into rules a computer could follow: which features drive cost, what they always check, and when they add margin and why. That took far longer than the maths. An AI build is the same kind of work, with your know-how instead. ## Demos are picked to work Sometimes AI gets it right on the first try. Sometimes it never gets it, however you ask. A demo shows you the first kind, because the person giving it picked a task that works. Your task could be either kind. You only find out by trying it on your own data. So I always build a small working version first, before anyone pays for a full build. ## “It should be cheap” and “it isn’t smart enough” are the same mistake People tend to say one of two things about AI. Some say it can do anything, so it should cost very little. Others say it isn’t as smart as their people, so why bother. Both look at the model and skip the work around it. The model is impressive, and it also knows nothing about your business until someone teaches it. Both are true at once. This series takes each side in turn. If you are thinking of building it in-house, start with [the story of the $640 toilet seat](/blog/640-toilet-seat-build-vs-buy). Other posts cover why a person stays in charge of every AI system, and why “AI” covers several very different kinds of tool. ## Start with one question The cheapest way to find out if AI can help you is to test it small. Pick one question your team spends hours answering every week. Build a small version on your real data, and try it. If it helps, build it properly. If it does not, stop there and keep what you learned about your data and your process. If the know-how behind that answer lives in one person’s head, write it down first. AI can only follow steps someone has written down, and [Process Capture](/automation) is how we do that. **Bring one question.** Book a call from the [AI Builds page](/ai). I’ll tell you if AI can answer it and what it would take. ## Common questions ### Why not just buy ChatGPT for everyone? General chat tools are good for writing and summing up. They do not know your jobs, prices, or suppliers until someone connects them to your data and builds the tools to read it safely. That connection work is what an AI project pays for. ### Do I need clean data before I start? No. Most businesses say their data is a mess. The first step checks whether your data is good enough for one question. Often it is. If it is not, you get a short list of what to fix first. ### What if the AI doesn't work for us? Then you find out early, on a small test. A working version on your real data shows how often it is right before you pay for a full build. If it cannot beat the way you do it today, I will tell you. ### How do I pick a first AI project? Pick a question your team asks every week that takes hours to answer. The data for it should already exist somewhere, and a wrong answer should be easy to spot. Leave anything where one mistake is costly until you trust the system. ## Where to go next - [You Could Build It Yourself. Should You? The Story of the $640 Toilet SeatIn 1985 the Pentagon paid $640 apiece for toilet seats. The fuller story explains the true cost of building AI yourself, and when you still should.](/blog/640-toilet-seat-build-vs-buy) - [Why Good Tools Beat a Smarter AIAI guesses when it lacks your data. Give it well-built tools that fetch and work out the real figures, and let the AI ask and explain.](/blog/good-tools-beat-smarter-ai) - [AI Isn't as Smart as My People. So Why Use It?Your experts know more than any AI. Use AI to clear the groundwork off their desks, so their hours go to the calls only they can make.](/blog/ai-not-as-smart-as-my-people) - [AI BuildsWhat we build, and how a build runs.](/ai)