I want to be careful about how I open this chapter, because it would be easy to read what follows as a man who has been in technology too long complaining about the newest thing to arrive. That is not what this is.
I have spent over thirty years in information technology and security. Every modern cybersecurity platform on my desk has artificial intelligence built into it. I use these tools daily. I use them to talk out my own thinking and help me shape it into something usable. I am not standing outside this technology throwing rocks at it.
But I have watched something happen over the last few years that concerns me, and it concerns me precisely because of everything else in this book.
Here is the shortest version I can give you: AI is a tool. The moment it becomes a crutch, you have lost the thing that made you worth listening to in the first place.
What Most People Do Not Understand About These Tools
Most of the people I watch using AI at work have no working understanding of what it actually is. They do not know the difference between traditional AI (the kind that has been sorting spam, flagging fraudulent transactions, and scoring risk for two decades) and generative AI, the kind that writes sentences back to you. That distinction is not academic. The two fail in completely different ways, and if you do not know which one you are holding, you cannot know how it is about to let you down.
Traditional AI generally fails quietly and statistically. It miscategorizes something and you find out later.
Generative AI fails confidently. It will hand you a well-formed, articulate, entirely wrong answer in the same tone of voice it uses for the correct ones. There is no tell. The prose does not get shakier when the machine is guessing.
That confidence is the whole problem, and it produces two distinct failure modes I see constantly.
The first failure is voice. People accept the output as written. They take what came back, paste it into an email, and send it. It reads competently. It also reads like nobody. Everything that made the message theirs, the phrasing they always use, the joke they would have made, the sentence that would have told the recipient this is David, and David is not actually angry, is gone. The words are fine. The person is missing.
The second failure is fact. People treat the output as verified because it sounds verified. In my line of work that is not an inconvenience, it is a finding.
The Verify Spine
Let me give you a concrete example from my own work, because I want you to see exactly where the line sits.
I will hand an AI system a set of our policies and ask it to identify the related government codes, state or federal, that those policies touch. It comes back and cites a specific privacy act. Chapter and verse. Confident.
Then I go and read that act myself.
Sometimes it is exactly right, and it has saved me two hours of searching. Sometimes it is dead wrong, the act exists, but it does not say remotely what the system claimed it said. Either way, my behavior is identical. I read the source.
That is the rule, and it is the only rule in this chapter I would call non-negotiable.
Trust the tool to find you the thread. Never trust it to be the final word. Verify. Don’t verify and you’re doomed.
Notice what that rule does not say. It does not say avoid the tool. The tool found the thread. Finding threads in a haystack of regulation is genuinely what these systems are extraordinary at, and refusing to use them out of suspicion is its own kind of professional malpractice. The rule only governs what you do with the thread once you have it.
The deeper danger is not any single wrong citation. It is the slow handoff. You verify the first ten answers and all ten check out. So you verify the eleventh a little less carefully. By the fiftieth, you are not verifying at all, you are approving. And somewhere in there, without any single moment you could point to, you stopped being the person who understands the subject and became the person who forwards what the machine said about it.
The moment you hand the final decision to the tool instead of keeping it in your own hands, you have lost the ability to actually understand the thing you are deciding about. Not the ability to appear to understand it. That survives a long time. The ability to actually understand it goes first, and quietly.
Does AI Make the Human Factor More Important, or Less?
I have been asked this a few times now, and the honest answer is that it can go both ways depending entirely on the person holding it.
But if you want my real answer, unhedged: left unchecked, it subtracts.
Everything this book has been about (reading how the other person is receiving you, catching the moment their posture changes, adjusting mid-sentence because you saw something land wrong) none of that survives the trip through a text box. AI is words. Written words. And the written word, however well constructed, is a narrow pipe.
Think about what actually gets stripped out.
Tone goes first. A sentence you meant as a light correction arrives as a reprimand. You know this already, because it has happened to you, and it has happened to you in messages you wrote yourself. Now hand the composition to a machine that has never met the recipient and has no idea that the two of you have history.
Passion goes next. Not theatrics, the ordinary human signal that you actually care about the outcome. In writing it either vanishes or, worse, it comes out as intensity, which reads as anger.
And the win-win read goes entirely. The core of the human factor is understanding how the other person is interpreting you, and remembering that you are there to find the compromise where everybody wins. You cannot do that from a text box. The bare written word cannot relay it and cannot receive it. You are negotiating blind.
The human factor only survives contact with these tools if you deliberately put it back. That means rewriting. Taking what came back and running it through your own voice and your own emotion until it sounds like a person who knows the recipient wrote it, because one did. If you are not editing, you are not communicating. You are forwarding.
When It Actually Matters, Walk Down the Hall
I can write a perfectly on-point email. Thirty years of practice will do that for you. Structure, tone, the right amount of context, the ask stated clearly at the top.
And when something genuinely matters, I do not send it.
I walk down to the person’s office, or I pick up the phone.
Because the written word is just the written word. When you speak, it comes straight out of your mind, through your passion and your emotion, out through your tone, and into the other person’s ear. There is no translation layer. They are not reconstructing what you meant from punctuation. They are hearing you mean it.
That is how people actually want to be talked to. And it explains something you have probably felt without naming.
Most people feel emails are a lecture.
They do. Even good ones. Even yours. A paragraph of text arriving in someone’s inbox has an inherent one-directional quality to it (here is my position, delivered) and no amount of softening language fully removes it. The recipient cannot interrupt you. They cannot ask the clarifying question at the moment it occurs to them. They can only receive, and then compose their own lecture in reply.
The most valuable thing I can tell you about AI and communication is not about AI at all. It is this: the tool is very good at helping you produce the thing you should probably not be sending.
Using AI Without Losing Yourself
None of this is an argument for avoidance. It is an argument for keeping your hands on the wheel. What that looks like in practice:
Use it to think, not to decide. Talking a problem out with one of these systems is genuinely useful, it forces you to articulate what you actually believe, and it will surface angles you had not considered. That is the good use. The output is raw material, not a verdict.
Never send the first draft. If the words came back from a machine, they are not yours yet. Rewrite until they are. If you cannot be bothered to rewrite it, ask yourself honestly whether the message was worth sending.
Verify every factual claim before you repeat it. Every one. Especially the ones that confirm what you already hoped was true.
Escalate off the page when it heats up. The moment a written exchange starts to spiral, no tool will save it. Nothing composed can fix what needs to be spoken. Pick up the phone.
Know which kind of tool you are holding. Take twenty minutes and learn the difference between the systems that classify and the systems that generate. You do not need the mathematics. You need to know how each one fails.
The Part That Cannot Be Delegated
Every hard thing in this book (the empathy, the listening, the ego management, the willingness to sit across from someone who does not want what you want and find the ground you both can stand on) every one of those is a thing a person does.
The tool can help you prepare for the conversation. It can help you find the regulation, structure the argument, anticipate the objection. Use it for all of that.
It cannot have the conversation. And on the day it convinces you that it can, you will have outsourced the only part of your work that was ever really yours.
Exercises
Take an honest inventory of how these tools have already changed the way you communicate:
1. Find the last three substantive messages you sent that were drafted with AI assistance. Read them aloud. Mark every sentence you would not have said out loud. That is your drift, and it is usually larger than people expect.
2. The next time a tool hands you a factual claim you intend to repeat (a statute, a statistic, a citation, a vendor capability) go read the primary source before you repeat it. Note how long it actually took. That is the real price of being right, and it is almost always lower than you assumed.
3. For one week, before sending any message that carries emotional weight, ask whether you could say it in person or by phone in under five minutes. If yes, do that instead, and note what happened differently.
Then write down, specifically, the decisions in your work you will never delegate to a tool. Be concrete about it. Post it somewhere you will still see it in six months, which is roughly when the slow handoff starts happening without your noticing.
“The machine can help you find the words. It cannot mean them. Meaning them is still your job.”
