Is All AI the Same Thing?

Ask ChatGPT to forecast next quarter’s demand and it will give you numbers, but it is the wrong kind of AI for the job. “AI” covers very different machines, the way “vehicle” covers a forklift, a van and a sports car. Business AI comes in three main kinds: AI for numbers, which forecasts from your history; AI for patterns, which spots the odd order or the slipping supplier; and AI for language, which reads, writes and answers in plain English. Pick the kind from the question you need answered.

One word covers very different machines

Most people met AI through a chat window. So when someone says “AI,” they picture ChatGPT or Copilot. That is one kind of AI, and it is the kind that is best with words.

The other kinds have been quietly at work for years. They sit inside stock systems, bank fraud checks and delivery planning. They don’t chat. They take in numbers and give back numbers.

Mixing them up is where a lot of AI money goes to waste, as I explain in why an AI project costs real money. A team tries the chat tool on a numbers job, gets a poor result, and decides AI doesn’t work for them. In fact they used a sports car to move pallets, and a forklift would have done the job.

AI for numbers learns from your own history

This kind answers “how many” and “how long.” How much of this product will we sell in March? How many weeks will this supplier take? How many hours will this job need?

It works by machine learning. That means you show the computer a long run of past examples, like three years of orders, and it learns the pattern. Then it uses that pattern to predict what comes next.

The test that matters is simple. You hide the last few months from the model, let it predict them, and check its guesses against what really happened. If it can’t beat the way you guess today, it isn’t ready. This is the kind behind the custom forecasting models on our AI Builds page.

AI for patterns spots what looks wrong

This kind answers “is this normal?” It learns what usual looks like in your data, then flags what doesn’t fit. People in the trade call this anomaly detection. In plain words, it is a watchdog for odd numbers.

In a factory that might be:

• A slipping supplier. One vendor’s deliveries creep later each month, before anyone has noticed.
• An odd order. A customer who always orders 200 suddenly orders 2,000.
• A scrap spike. One machine starts throwing out more bad parts than its usual run.

It points to where to look, early, and a person then works out why.

AI for language is great with words and weak at sums

This is the chatbot kind. It has read a huge amount of text, so it is very good at reading, writing, summing up and answering questions in plain English. For a follow-up email or a meeting summary, it is often all you need.

It is weaker at maths than most people expect. When Yuntian Deng, a researcher at the University of Waterloo, tested it, GPT-4o got less than 30% right on multiplication beyond four-digit by four-digit sums. It also knows nothing about your business. It has never seen your orders, your prices or your suppliers.

A chat tool will give a confident answer to a forecasting question, even though it has never seen your sales history. The fix is to give it tools: small programs that fetch your real records and do the exact maths. Good tools beat a smarter AI explains how that works.

Real builds often use more than one kind

The most useful systems mix them. Language goes on top, so your team can ask a question in plain English. Numbers and patterns sit underneath, so the answer comes from your real data.

Say a planner types, “Will we have enough stock of part 4410 for March?” The language model reads the question. It asks a forecasting model for March demand. It asks the stock system what is on hand and on order. Then it writes the answer in a sentence, with the figures and where they came from.

A price quote is a good example. It needs rules and exact maths, built from what experienced estimators know about shape, material and finish. An estimator doesn’t need something that writes nice sentences. They need a tool that gets the number right and shows how it got there, so the sums come off their desk and the judgment stays with them.

Some questions need no AI at all

There is a fourth answer, and it is often the right one. If the rule is fixed, like “reorder when stock drops below 10,” you need a formula. AI would only make it slower and less reliable. If you have no data on the question yet, no kind of AI can learn from it.

So this is not for everyone. If your problem is a fixed rule or missing records, start there. When AI is the wrong answer covers those cases.

“Superintelligence” will not help you choose

You will see headlines about one AI that will soon do everything. The idea is older than it looks. In 1965 the mathematician I. J. Good wrote about an “ultraintelligent machine” and called it “the last invention that man need ever make.” In 2014 Nick Bostrom’s book Superintelligence made the New York Times best-seller list for science books.

Those ideas make for good debate. They won’t tell you which tool forecasts your demand next quarter. The tools you can buy and build today are still separate kinds, each good at its own job, so start from the question you need answered.

Start with the question you need answered

Write down one question your team spends hours on each week. Then ask what kind of answer it needs. A number about the future points to AI for numbers. A warning about something odd points to AI for patterns. A plain-English answer from your own records points to AI for language with tools underneath. If you already pay for ChatGPT or Copilot, why not just use ChatGPT covers where those seats fit.

Once you have the question written down, book a call from the AI Builds page. I’ll tell you which kind of AI fits it, or if it needs AI at all.