AI Doesn't Replace Your Thinking. It Raises the Bar for It.

AI rarely fails on its own. Start with the business problem, provide context, review the output, and keep people in the loop until it earns trust.

I've used AI tools almost every day since November 2021. I was one of the original seed trainers for Google Bard. Over the last few years, I've coached colleagues and business professionals on how to get real value from these tools.

Here's what I've learned. The tool is rarely the problem. The problem is usually what we ask of it and how little we question its output.

AI doesn't make critical thinking less important. It makes it the skill that separates good results from bad ones.

Here are the habits I coach most often.

1. Start with the business problem

Before I open an AI tool, I ask one question. What business problem are we trying to solve?

Everything I write shows how a solution makes life simpler for a developer, engineer, architect, or business professional. AI work should start the same way. Not with the tool. With the problem and the person who has it.

This is where I see many AI efforts go off track. Teams point AI at a task and make it faster without asking whether the task matters. Automating a bad process just gets you bad results faster.

2. Most bad results start with missing context

A technology journalist colleague recently told me his LLM wasn't giving him what he expected. He's been writing for 30 years. He wanted new articles that sounded like the ones he'd always written. What came back didn't.

I have to believe the fix was simple. If he had shared 10 of his articles with the model, it would have picked up his style.

An LLM doesn't know what's in your head. It doesn't know your audience, your standards, or your past work. Give it one line, and you get a generic answer. That's not the tool failing. That's the tool guessing.

Most of us aren't training these models. We're giving them context. And context is something you control. Before you ask, tell the model:

• Who you are and what role you're in

• Who the output is for and what they already know

• What good looks like, with examples of your own work

• What to avoid, including words, tone, or assumptions

• What it can't see, like the data, history, or constraints behind the request
  

I do this every day. I write for developers, engineers, and architects. I also write for tech-savvy business professionals. Each audience and each website gets its own context: who's reading, what they care about, and how I sound there. So my LLM writes differently for each one. That's not magic. It's setup.

The critical thinking part is noticing when a bad answer is really a bad question. Before you blame the tool, read your prompt back and ask: would a smart new hire know what I want from this?

3. Decide what good looks like before you ask

Developers call this writing evals. Teachers call it an answer key. Most business users skip it.

If you don't know what a good answer looks like, you can't judge the one you get. You end up accepting whatever sounds confident. And AI always sounds confident.

So before I ask for anything important, I write down what the answer has to get right. For an article, that's the business problem it addresses, the facts I already know, and the reader I'm writing for.

It takes a minute. And it turns "this seems fine" into "this meets the bar" or "it doesn't."

4. Treat the first answer as a draft

The biggest mistake I see is treating AI output as a conclusion. It's a starting point.

I've interviewed more than 4,000 IT executives. You quickly learn not to print the first thing someone tells you without checking. The same rule applies to AI.

I've caught made-up numbers. I've caught made-up quotes. When I do, I ask the LLM to cite its source. When I call it out, it admits it made them up. That should give anyone publishing AI output without checking it pause.

Here's what I do with an answer that matters:

• Check the sources. If it cites a stat or a quote, I ask where it came from and find the original.

• Ask what it left out. "What would have to be true for this to be wrong?" is one of my most useful follow-ups.

• Ask for alternatives. One answer is an opinion. Three answers show you the trade-offs.

• Get a second opinion. Run the same question through a different model and compare.

None of this is hard. It just has to become a habit.

5. Decide where agents work and where people do

Agents can now run real parts of business processes. But not every step should be agent-driven.

I coach teams to break a process into small steps and look at each one. Which steps are repeatable, rules-based, and low-risk? Those are good candidates for agents. Which steps need judgment, context, or accountability? Those stay with people, at least for now.

Work in small increments. You learn more from handing one step to an agent and watching it closely than from handing over the whole process and hoping for the best.

6. Keep people in the loop until AI earns your trust

Keep people in the loop until you've seen enough use cases and edge cases to be confident AI handles them the way you would.

The easy cases don't prove much. The edge cases do. Confidence comes from watching AI handle real, messy situations correctly, again and again. Until then, a person reviews the work.

AI can draft, summarize, and suggest. It can't own the outcome. You do.

The bottom line

If AI isn't giving you what you want, start with four questions. What problem am I solving? Did I give it enough context? Did I know what a good answer looked like? Did I check what came back?

Most of the time, the fix is in one of those four. And all of them are within your control.