One story. Gone deep.

There is always a meeting where a bet gets started. Somebody champions it, somebody funds it, somebody builds a demo over a weekend and it looks great on a Tuesday. Nobody schedules the meeting where a bet gets killed. That meeting has no owner, no budget line, and nobody whose year gets better for calling it. So it doesn't happen. The bet doesn't die. It just keeps not dying.

You have one of these right now. The AI customer support idea that's been on the leadership agenda for three straight quarters. Never approved. Never killed. Two different people have built demos on their own time. A vendor has been "in conversation" since February. Every quarter it comes back around, everybody agrees it matters, and nothing gets decided.

It costs nothing, because it never shipped. That's what you tell yourself.

What it costs is the attention of the three people who keep re-litigating it, and the credibility of whoever championed it first and has now raised it twice with nothing to show. The bet isn't running in production. It's running in everyone's head. That is not free.

And I'm not going to tell you whether that support project was a good idea. That's the point. You don't know either, and it isn't because the call is hard. It's because nothing in your company is allowed to end it.

"No" and "not yet" are different answers, and only one of them is usually a decision.

"No" closes it. The bet is dead, everyone stops spending attention on it, and the person who championed it keeps their reputation. What happened was a decision, not a person being wrong.

"Not yet" is a real answer too. Often it's the right one. But only when it can say what it's waiting for. A "not yet" that can't name the thing it's waiting for isn't a decision. It's a mood with a calendar invite.

Then there's the third thing, and it's what most companies actually run. Not a no. Not a not-yet. No decision at all. From the inside it feels like "not yet." It behaves like "yes, eventually, by accident."

Sometimes the undecided bet ships anyway. Not because anyone approved it, but because nobody had the standing to stop it and somebody had the budget to start it. So it goes live. It runs for a while. And then something happens on a Thursday and it comes back out.

The rollback is the no. Same decision. Free at the start, and nobody made it. So you made it after the wiring, after the training, after a customer saw it, and after somebody's name got attached to a failure instead of a call.

So here's the question, and I'd want the answer fast.

Name the last AI idea this company said no to.

Not paused. Not still evaluating. Not "we're keeping an eye on it." Said no to, on a date, out loud, and everybody knew it was over.

Most people can't name one. And the reason is never that every idea was good.

A company that has never killed an AI bet before launch doesn't have a great track record. It has no way of saying no at all. Everything that arrives gets adopted, or gets postponed until it arrives anyway.

That's the work I've spent this whole run building toward. Next week I'll show you what it looks like in practice.

What else matters this week.

On July 9, OpenAI said it's shutting down ChatGPT Atlas. The browser goes dark on August 9, one week after you read this. Atlas launched on October 21, 2025.

OpenAI's own reason, from the announcement: the capabilities were built on "what we learned from Atlas users who took a leap of faith on a new browser." The browsing work moves into the ChatGPT desktop app and Codex.

The reporting has been less generous. Futurism and TechTimes pointed at prompt injection exposure, an agent slow enough to be a punchline, and a product that never made it off macOS. Keep those separate from what OpenAI actually said. Two different claims, and only one has OpenAI's name on it.

Here's why it's in this issue. The best funded, most all-in AI company on the planet shipped a flagship bet and ended it inside a year. Not quietly starved. Not left to rot behind a login. It named a date, told everyone, and moved the good parts somewhere else.

If OpenAI can end an AI bet on the record, your company's inability to do it isn't about AI.

On July 21 a serious flaw turned up in a small piece of free software called fastjson. You've never heard of it, and that's the point. It sits deep inside other software, the kind your vendors build with and never put on an invoice. It let an attacker run commands on a server without a password. Attacks started the next day, landing almost entirely on US companies, according to Imperva.

Every writeup that week said the same thing. No patch. BleepingComputer, SecurityWeek, The Hacker News. They were right: the version everyone ran was the end of the line. It was finished, and its maintainers had moved on.

Then on July 29, Alibaba shipped a fix anyway, on something they'd already closed.

So for eight days, flaw to fix, a lot of companies had no patch and no plan. What saved them wasn't a decision any of them made. It was a stranger reopening a door they'd already closed.

Somewhere in the software you bought is a component nobody maintains. You didn't pick it, you can't see it, and moving off it has never been on anyone's agenda. That's fine right up until it isn't, and then somebody else's goodwill is doing the job your upgrade plan was supposed to do.

You can't budget for that twice.

Not every signal needs action.

New flavor this month, and it's the mirror image of what I've spent the summer arguing with. That one said move faster. This one says the whole thing has failed, so sit it out.

The number is real. Sinch surveyed 2,527 senior decision makers across ten countries at the start of this year and found 74 percent had rolled back or shut down an AI customer communications agent after deploying it. Sinch sells customer communications infrastructure, which is worth knowing before you read their conclusions. And it measures that category specifically, not all enterprise AI.

Now the number nobody quotes. The rollback rate climbs to 81 percent at the organizations with the most mature governance.

Read that in the direction it actually points. The companies best at watching their agents pulled the most agents. That isn't a failure rate. It's a detection rate. The firms reporting no rollbacks aren't the ones getting it right. They're the ones who wouldn't know.

Rolling back isn't the embarrassing part of that survey. Never having a way to say no before launch is.

And "AI is finished, we're sitting this one out" is the same non-decision as shipping everything. It just dresses better.

No answer. Just the question.

When was the last time somebody here said "not yet" and could actually name what they were waiting for? Not a quarter. Not more data. The specific thing that would have changed the answer. And if that has never once happened, what exactly have you been calling a decision?

This run has worked the half of AI adoption nobody puts in the sales deck. Two weeks ago: not all of those reaches are equal, and the first cut is whether you survive the ones that go wrong. Last week: the ones that survived still need a name on them and a limit around them. This week: most companies never had a way to say no in the first place, so the no shows up late, wearing a rollback.

If you want a plain read on which of your own AI bets nobody has actually decided, that's what the free assessment is built to surface. Fifteen minutes, nothing owed.

→ echocyber.io/assessment

Next week: what this work actually looks like.

Prefer audio? Jane reads every Pulse edition on the Signal vs. Noise podcast. Five minutes, same signal. Find it wherever you listen.

Michael Faas is a fractional CTO/CISO who translates technical complexity into business decisions. echocyber.io