Why Didn't AI Get It Right the First Time?

Ask AI to write a follow-up email for a quote and it nails it first time. Ask it to price a job the way your estimator does and it may never get there. The difference is what it knows. AI is amazing, but it doesn’t know what you want, and the rules it needs are often in someone’s head: the old price for that customer, the machine you never quote under a certain size. Write down what “right” looks like, test against real answers, and results improve fast.

Some jobs need only general knowledge

A lot of everyday work only needs what the AI already knows. It has read a huge amount of public writing. So it is good at drafting an email, summing up a long report, tidying a rough note, or turning bullet points into a letter.

For jobs like these, there is no hidden rule to learn. A good email to a customer looks much the same in every business. If the first draft is off, you can see why at a glance and fix it in a few seconds. That is why the first try with AI often feels like magic.

Those early wins set a high bar. So when the next job goes badly, it feels like the AI has broken, when usually the job is the thing that changed. Closing that gap is where most of the work in an AI project goes, as I explain in why an AI project costs real money.

Other jobs need your rules

Now ask it to price a job, plan next week’s work, or pick which order ships first. Each of these runs on rules that belong to your business. Many were never written down.

Think about what your estimator knows. This customer always gets the old rate. That machine is never used for small runs. A certain metal adds a day, because the supplier is slow. None of this is on the internet, so the AI has never seen it. It will still give you an answer, and the answer will look neat and sure of itself.

This is where people get burned, because the AI fills the gap with a guess. A 2025 OpenAI research paper explains why: “Like students facing hard exam questions, large language models sometimes guess when uncertain.” The guess is often wrong in small ways that only an expert would spot.

AI can only follow steps someone wrote down. If the know-how lives in one person’s head, the first job is to get it out and onto paper. That is what Process Capture does, and it pays off even if you never use AI.

RAND found the same thing in failed projects

Big AI projects hit the same wall. In 2024, the research group RAND interviewed 65 experienced data scientists and engineers about why AI projects fail. The report notes that by some estimates more than 80 percent of AI projects fail.

The top cause was a gap in understanding. In the report’s words, “Misunderstandings and miscommunications about the intent and purpose of the project cause more AI projects to fail than any other factor.” The builders did not fully know what the business needed, and the business did not fully explain it. The AI did exactly what it was told, and what it was told was not quite right.

The report’s advice was to make sure the technical team understands the purpose of the project and how the business works. In plain terms, someone has to explain the rules.

Describe “right” with real examples

I learned this building an instant costing engine for metal castings. The maths was the easy part. The slow part was getting the rules out of the estimators’ heads. I tell that story in full in why AI isn’t as smart as your people.

An AI build works the same way. Before you judge it, write down what a good answer looks like. The best way is a set of real examples:

1. Real questions with known answers. Pull 30 or 40 jobs your team has already handled, along with what they decided. These are your test.
2. The tricky cases. Add the odd ones: the rush job, the customer with special terms, the part nobody wants to quote.
3. What a wrong answer looks like. Note the mistakes that would cost you money, so you can check for those first.

Then run the AI on every example and count how many it gets right. Add a missing rule and run it again. When the score climbs, you can see progress in numbers instead of gut feeling. Most of the gain comes from the rules and tools you give it. A newer model helps less than people expect, a point I make in why good tools beat a smarter AI.

A person still checks the answers, especially at the start. That is how your team learns where to trust it. I explain why that check matters in why AI isn’t as smart as your people, and why to use it anyway.

Some jobs never get there, and that is useful to know

The same RAND report lists a fifth cause of failure: some problems are too hard for AI to solve. Its authors put it bluntly: “AI is not a magic wand that can make any challenging problem disappear.”

So give it a fair test. Use your real data, write down the rules, and build the tools it needs. If the score still stalls well below what your team does today, stop. You have not wasted the effort. You now have your rules written down, a test set of real answers, and a clear reason not to spend more. I go through the warning signs in when AI is the wrong answer.

This is also who AI is not for yet. If the rules for a job change every week, or nobody can say what a good answer looks like, AI will struggle. Sort out the process first. If a wrong answer could hurt someone or cost you a large customer, keep a person making that call.

If you want to know where your job sits, the AI Builds page explains how I test an idea small on your own data before anything bigger.

Start with one job it got wrong

A bad first try with AI usually means it was missing your rules. Pick one job it got wrong, and sit down with the person who knows how to do it right. Their rules and a handful of real past answers are enough to start a fair test.

Bring the job it got wrong. Book a call from the AI Builds page, and I’ll tell you if AI can learn to do it and what it would take.