One story. Gone deep.

In the next six months, every consultant your inbox knows is going to hand you the same thing. An AI roadmap. Where to point it, what to automate, which agent lands in which corner of the business. It will look like insight: confident, specific, naming your own workflows back to you.

Here is what that map is actually worth. Almost nothing. Not because it is wrong, but because everyone has it now. The roadmap is becoming a commodity, and the roadmap is the easy half.

Finding where AI goes is easy. The hard half is the question almost nobody is selling you: which of these bets can your business actually survive getting wrong?

One question asks where AI could help. The other asks what happens the day it does not. Only the second decides whether adopting AI makes you stronger or quietly hands you a failure you cannot trace.

Here is the part the roadmap skips. An AI agent is not a tool you bolt onto the side of the business. A tool sits in its own box. It does one thing, and when it breaks, it breaks in that box, you notice, and you fix it. An agent is not that. The moment you drop one into a real workflow, it becomes a dependency. It holds a credential. It reads your data. It writes to systems other functions rely on. You did not add a feature. You wired a new node into identity, into data, into workflow, into trust, all at once. And a node does not fail in its own box. When it fails, it fails outward.

Make it concrete. The agent that drafts your customer replies also holds your CRM credentials, so when it starts inventing commitments you never authorized, that is not a support glitch. It is a legal exposure, a data-integrity problem, and a trust problem with every customer who believed the reply. The agent that reconciles your invoices touches financial records, vendor data, and an approval step three people downstream assume is clean. It does not fail quietly in the finance corner. It takes four other functions with it, and the org chart never showed you the wire connecting them.

Here is why the roadmap cannot see it. A roadmap treats your business like a complicated system: break it into parts, optimize each part, add up the wins. Point AI at support, save hours. Point it at finance, save more. Sum the hours and there is your return. That math works on a jet engine. It does not work on your business, because your business is not complicated. It is complex. The parts are wired to each other, and an AI agent adds wires faster than anything you have ever adopted. Complicated systems can be optimized. Complex systems must be governed. An AI agent does not just make your business more efficient. It makes it more complex. And almost everyone is measuring only the efficiency.

For six weeks I wrote about a discipline built for exactly this. Which decisions are expensive, which are hard to reverse, which are genuinely uncertain. Which ones earn a model, and which you can just decide. I pointed it at vulnerabilities, the easiest place to show the move. This is me pointing it somewhere else. A founder dropping an AI agent into a live workflow, with real credentials and real data, is expensive if it goes wrong, nearly impossible to cleanly reverse once four functions lean on it, and genuinely uncertain in how it fails. That is not a roadmap decision. That is a decision that earns a model.

So take the roadmap. Let someone map every place AI could go. That was always the easy half.

The map tells you where AI can go. It will never tell you which one takes the other four down with it. That question is not on the map. It is the whole job.

What else matters this week.

You turned on Microsoft 365 Copilot for the same reason every founder does: it saves the team hours. Nobody sat down first to scope what it could actually reach, because that is not how you buy a productivity tool.

Here is what that scope turned out to be. Varonis Threat Labs found a flaw in Copilot's Enterprise Search, nicknamed SearchLeak, CVE-2026-42824, published to the U.S. National Vulnerability Database on June 4. A single crafted search link could make Copilot hand over the user's data. It worked by chaining a prompt injection, hidden instructions buried in the search query, with a few other bugs. And the data in reach was not just emails and indexed files. It was one-time MFA codes and password-reset links. The keys to everything else.

Microsoft fixed it on their own servers, so there was no patch for you to apply. That is the good news and the whole point at once. You never touched this, and you could not have. The assistant you switched on to save time is not a self-contained chatbot in its own box. It holds a credential and sits wired into your identity and your data, so one link reached the codes and reset links that unlock every other account. You did not add a feature to your business. You added a door, and you did not get to decide where it opened.

Not every signal needs action.

The headline making the rounds this month is built to make you feel late. Cisco replaced around 4,000 workers with AI agents. If a company that size is handing jobs to software, the takeaway writes itself. Move now or get left behind.

Except it did not happen. That number is two unrelated stories welded into one scary sentence.

Story one, mid-May. Cisco announced it was cutting fewer than 4,000 jobs, under 5% of about 86,000 people, and shifting that capital toward AI infrastructure: chips, optics, security. Cisco's own CFO said on the record it was "not savings-driven." A reallocation, not software replacing those workers. Same stretch, the company posted record revenue of $15.8 billion, up 12% year over year. Not a company in trouble.

Story two, around July 1, entirely separate news. Cisco is rolling out an internal AI agent to roughly 90,000 of its own employees. A productivity tool for the people who still work there. Not the reason anyone left in May.

Neither of those is "AI took 4,000 jobs." The scary number is a misread of a timeline.

So here is the turn, same as every week. This is not "do not adopt AI." Adopt it. But you are being sold speed built on a headline that is not even true, about a company that looks nothing like yours. Speed is the noise. Survivability is the signal.

You are not Cisco. Do not let a misread of a giant set your timeline. Adopt what you can survive, at the pace you can govern. That is not caution. That is the whole discipline.

No answer. Just the question.

Think of the one AI agent you are closest to putting into a real workflow. Do not ask whether it will save time. Assume it will. Ask the other question instead: the day it fails, how many functions does it quietly take with it? And do you actually know the answer, or are you guessing and calling the guess a plan?

This arc spends the next several weeks on the half of AI adoption nobody is selling you: not where it goes, but which bets you can afford to get wrong. It starts with seeing the wires, the dependencies an agent gets bolted into that never show up on the roadmap.

If you want a plain-English read on where your own program is measuring only the easy, visible things and missing the connections underneath, that is what the free assessment is built to surface. Fifteen minutes, nothing required from me.

→ echocyber.io/assessment

Finding where AI goes was always the easy half. Next week we stay on the hard one.

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