AI turned the lights on your content

The AI agents now reading your content have no instincts and no patience. Leave anything to interpretation and they’ll probably get it wrong.

Picture a kind little error message. “Oops, something went wrong. Try again or contact support.” Warm and free of blame. I’d ship it any day…minus the “oops.”

Now hand that line to an AI agent that’s paying a bill for someone. It reads the message and has no idea what to do. No reason, no next step…just a friendly nudge to try again. So it tries again. And again. Ten times, until the bank’s fraud limit kicks in and freezes the customer’s payment.

Now you’ve got a bad experience, angry stakeholders, and everyone’s real quick to blame the AI. But the AI did its job. So did the content designer, for that matter. The problem is nobody had decided what should happen when the payment failed. And the nice, warm sentence was covering for that.

Should the customer retry, or use a different card? Or does another attempt trip the fraud limit? These are all things a human might guess. An agent won’t. It has no instinct to fall back on. It only has the words we gave it.

Vague content used to hide the missing pieces. Now AI agents shine a flashlight on them.

The agent might be reading your content first

A year ago I wrote that AI had started reading our content before people did. Summarizing notifications. Rewriting emails into previews. That felt big at the time. Little did we know reading was just the start.

Now there are AI agents coming in droves. And they don’t just read the words. They act on them. Then they tell your customer what happened…or what they think happened.

Imagine someone opening a savings account. They used to search, land on your page, compare, and decide who to trust. Every word we wrote was aimed at that person. We had personas plastered across the walls to remind us who our customer was.

Now that person might never see your content but their AI agents will. In June, Cloudflare reported that bots and AI agents now generate more web traffic than people do, a crossover its own CEO didn’t expect until 2027. And Salesforce says 37% of consumers are already comfortable with AI agents creating more personalized or useful content for them. 

And why wouldn’t they? It’s easy. Now a customer can just ask Claude. “Find me a high-yield savings account. No monthly fees. And I need direct deposit.” The agent reads the pages, compares the rates, checks the fine print, picks one, opens the account. Your page never gets a human visitor, meaning it never had a shot.

That means your content isn’t about marketing anymore. The same agent meets your words on more than one surface, and the job changes with each. On a marketing page, it’s sizing you up to decide whether to recommend you. In a product flow, it’s acting for the person, moving the money. The tricky part is the agent reads nothing like we do on either one. A person reads between the lines. An agent only reads the lines.

To further complicate things, the agent speaks for you, too. When an agent reports back to the customer, it puts your content in its own words. Write clearly and it describes you the way you’d want. Write vaguely and it fills the blank with a guess, says that guess to your customer in a calm, confident voice, and you never even see it happen. If your pricing and limits are vague, the agent might not describe you wrong. It might not mention you at all.

This isn’t a future problem. We’re already seeing it in the news. Last year, Cursor users started getting logged out when they switched machines. Support told them it was expected under a new one-device policy, signed off on by a company rep named Sam. Great. Except there was no policy. There was no Sam. An AI support agent had invented a rule that sounded right, and customers believed it.

From persuasion to instruction

For decades we wrote to persuade. Now we also write to be carried out. From persuasion to instruction. That one shift changes our idea of what “good” even means.

And it keeps landing in the same place. Every spot where a machine stumbles, you’ll find a decision nobody made. What to say when the answer isn’t known. A person fills the blank on instinct and moves on. But the AI agent can’t. Or, worse, it thinks it can and gets it all wrong.

This means a content audit isn’t a copy audit anymore. It’s a decision audit. The vague sentence is just where the missing decision shows up. The real bug is the decision behind it.

Only the real warmth stays

One thing caught me off guard when I started really digging into this. I was going to have to let go of what was one of my favorite parts of the job…nailing that perfect phrase. Now the copy that felt cold to me, the version stripped of the warmth I used to spend hours getting right, was the clearest one for the AI agent. There was no feeling to wade through. And the reassurance I was proud of was the exact thing an AI agent just couldn’t get past.

In the new world, a lot of the craft I defended was decoration. For years, it was probably already losing the humans who skim. But there was no doubt when it came to the agents.

The catch is knowing who’s on the other end. Errors, statuses, fees, limits: those go machine-first now. Plain and exact…every number named. A welcome screen or a thank-you, the moments a person actually reads, keep their warmth. You tell which by what breaks when it’s misread. A wrong reading that triggers a wrong action is machine-first. One that just lands flat belongs to the person. When a screen is both, the facts lead and the warmth follows.

I can’t talk about this without addressing the fear underneath it, because I’ve felt it. It’s not “will AI take my job.” It’s much quieter. “Will I still get to make the work good?” The warmth, the little turn of phrase that made a flat screen feel human. The zhuzh. The worry is we’re about to become spec writers for robots.

But the warmth that was ever real didn’t die. It got harder to place. Spreading tone evenly across a flow felt like craft. It was mostly habit. You’re not losing the good zhuzh. You’re just being asked to aim it.

I won’t miss one part of the old way. The hours I spent arguing with my friends in marketing about how much zhuzh is too much zhuzh. Nobody won those. It was all taste. The agent settles the bet now. Too much on an error screen and it can’t act. Finally, a machine has ended the great zhuzh debate.

What this looks like in the real world

We have a rule about exclamation points on my team. Two of them. Not per page. Per designer, per year. Too much enthusiasm reads as noise, so we ration it hard. Warmth you ration is warmth that lands. That’s a rule for humans.

Lately we’ve been writing a different rulebook. Hundreds of content rules now, built into the AI assistants my team works with, and more every week. We wrote them so clearly any agent can count on them as rules, not vibes.

Five ways product copy can break a good agent

The rules are even more important when it comes to the product-flow surface. On a marketing page, a vague word costs you a recommendation. In a product flow, it makes the AI agent do the wrong thing on someone’s behalf. You don’t fix that by reading more carefully. You fix it by knowing exactly what trips an agent, then writing a rule so it never ships again. These are the five to really watch out for. Failure first, then the rule to stop it.

The orphaned pronoun. “We’ve got it. It’ll show up in 3 to 5 days.” The agent’s immediately confused. It, what? The refund, or the email confirming it? This copy gives a human just enough info to assume what “it” is. But an AI agent tracking that object into the next step guesses, and it can guess wrong. That loose pronoun is how it calls the wrong tool. Write the noun and the number first: “Your $47.32 refund is on the way. The money reaches your bank in 3 to 5 business days.” You earn “it” only after you’ve said the thing once.

Rule: no pronoun without a named antecedent in the same string.

The dead-end error. The one we opened with. “Something went wrong. Try again.” No cause, no branch. The only move the AI agent has is to repeat the action, which is how a declined card becomes ten retries and a frozen account. The fix: “Your payment didn’t go through. Your bank declined the card. Use a different card.” Cause first, then the one next step. And, yes, I know. That copy’s cold as ice.

Rule: every error names what failed and what to do about it. “Try again” can never be the only instruction.

The overloaded button. A control that says “Continue” on the screen that actually sends $2,000. “Continue” tells the agent nothing about the stakes, so it can’t judge whether to act for the person. In its eyes, “Continue” and “send $2,000” are the same word. Don’t leave it to chance. Put the consequence on the button: “Send $2,000 to Maria Lopez.”

Rule: no “Continue,” “Submit,” or “Done” on a screen that moves money or changes account state. The label clearly says the action and the object.

The buried number. “You’re all set. Heads up, new accounts may have some limits while we get to know you.” The real limit isn’t there, so the agent reports no limits and the person hits a wall mid-transfer. Write: “Your account is open. For the first 30 days, you can send up to $500 a week.”

Rule: if a limit exists, state the number. “Some limits” is not a limit.

The screen with no memory. “Almost done. Confirm to finish.” Finish what? The agent doesn’t carry the last screen in its head the way a person does. Put the decision on the screen that asks for it: “Confirm autopay. $112.40 on the 1st of each month, from checking.”

Rule: every screen restates what’s being decided. Never borrow the antecedent from the screen before.

Notice what these share. Not one is a wording problem. Each one is a missing fact or a missing decision, and the warm sentence was standing in for it. The fluff your marketing partner added to “make it pop” is usually the exact thing the AI agent chokes on.

So don’t fix the screen. Write the rule. Every one of these should live in your own AI assistant as a check that runs before any string ships. Fix one error message and you’ve fixed one error message. Write the rule and you’ve fixed every error message your team writes next, and every one the AI agent generates on its own. That’s the difference between editing copy and owning the content system that produces it.

Now, put the warmth back. These rewrites look cold on purpose. I stripped them to show the skeleton. You don’t have to ship them bare.

It comes down to order. Lead with the cause and the next step, the part the agent acts on, then be as human as your brand wants. The agent reads in sequence, so once the fact and the action are there and in the right order, the warmth around them is free.

Take the payment error. Stripped, it’s “Your payment didn’t go through. Your bank declined the card. Use a different card.” Warm it and you get “We couldn’t process your payment. Your bank declined the card, so try a different one.” Same cause, same next step, so the agent still acts. The voice just got warmer for the person.

The version that breaks is the one that leads with feelings. “We’re so sorry, that didn’t work. Don’t worry, these things happen. Give it another try.” The apology’s first, the cause is gone, and “give it another try” walks the agent right back into the loop. Warm and useless to both readers.

Warmth after the fact is voice. Warmth in place of the fact is a liability. And the instruction never bends to tone. “Try a different card” can’t soften into “you might want to try again,” or you’ve swapped a clear action for a vague one.

That’s what people miss when they hear “write for machines.” It doesn’t mean strip the humanity out. Clear words serve both readers. You’re not choosing between the person and the agent. You’re deciding what they read first.

The fastest gut check I’ve got, before you start: paste a screen into Claude and ask it to say the message back in one sentence. If the summary drops the point or flips it, the words aren’t ready. Front-load the fact and run it again.

You’ll find a sixth thing the rules can’t fix, because it isn’t copy at all. It’s every decision nobody made, and that’s where your real backlog is found.

Surfacing the decision isn’t the same as making it

You can prove nobody decided what happens when a payment fails. But you won’t always be the one to decide it.

The audit hands you a list of unmade decisions, and you’ll quickly find most aren’t yours to make. Whether a customer can retry is product’s call. What a fee can say is legal’s. And so on.

So the new job isn’t deciding. It’s refusing to let the decision stay unmade. You bring the decision to the room, you name who owns it, and you don’t let the flow ship until someone answers. The string is just where that answer gets written down last.

And yes, I know that won’t be as easy as it sounds. But it’s my hope for the content designers of the future.

Why content designers are in a good place

The fear I hear everywhere is that AI replaces content design. It’s quite the opposite. As AI becomes how people reach a product, the words behind it carry more weight, not less. The AI agent only knows what your content tells it. Vague words in, vague agent out. Clear words in, clear agent out. It’s that simple.

From inside a team that uses AI every day, I can tell you the work isn’t shrinking at all. It’s growing. Every flow has a second reader now. Every rule we write down surfaces ten more we never had time to define. Drafts come faster, sure. But a faster draft just gets you to the real work sooner: deciding what’s true and what happens when it breaks.

If you lead a content team, this is the year to reread your operational copy with no instincts and no patience, just the way an AI agent will. You’ll find the fluff. You’ll also find the decisions your company has been quietly avoiding for years. Shine the light on them.

We spent years asking people to take words seriously. Turns out, the machines already do.