Why Almost Everyone Chasing AI Is Looking in the Wrong Direction

I want to open this one differently. Most weeks I can point to a single stat or a single headline and build the whole issue around it. This week didn't hand me one thread, it handed me seven, scattered across five days, about wildly different things: a healthcare data study, a gripe about the LinkedIn algorithm, my own AI certification, a stat about training budgets, a running list of the ways I use a personal AI assistant, a governance controversy at a frontier lab, and a networking conversation in Houston about broken lead handoffs. On the surface, none of those belong in the same room together.

But I kept circling back to a phrase I used twice this week without planning to: the unsexy work. It showed up when I was talking about API permissions and read/write syncs. It showed up again when I was talking about diagnosing a broken process before automating it. Once I noticed I'd said it twice, I realized it was the actual subject of everything I'd written all week, whether the post was about healthcare regulation or a product leader walking away from a frontier lab. This issue is my attempt to explain why the boring, invisible, hard-to-post-about work is the whole game right now, and why almost everyone chasing AI is looking in exactly the wrong direction for it.

90% of Healthcare Leaders Say Their Data Isn't AI-Ready. Most Small Businesses Have the Same Problem.

Regulation forced healthcare to fix its processes before adding AI. Most other companies skip straight to buying the tool and never find out where their own process actually breaks.

A new agentic AI study covering healthcare, manufacturing, retail, and banking put a number on something I see constantly in the field: 90% of healthcare leaders say 20% or less of their data is ready for an agent to act on. Hospitals can't bolt an agent onto EHR systems that don't talk to each other, so, for now, they've mostly stopped trying. Instead, they narrow the agent's scope, keep a human on the final step, and fix the process underneath first, not because they wanted to be disciplined about it, but because regulators and decades-old legacy systems left them no other option.

Tech companies, without that same outside pressure, have mostly gone AI-curious instead. The tools get bought. The underlying work still runs exactly like it did before, just with a chat window bolted onto the side of it.

I was at a local networking event in Houston earlier this month, talking with business owners across a totally different set of industries: HVAC, landscaping, law, dental. None of them cared about the technology once we got two minutes into the conversation. "AI" is technical jargon to most small business owners. What they wanted was more leads and more revenue, plainly stated. But when I asked where their leads actually get stuck today, most of them couldn't answer me. That gap, between wanting the outcome and being able to name where the process actually breaks, is the exact same gap the healthcare numbers are describing, just at hospital-system scale instead of a five-person shop.

Here's the analogy I keep coming back to. Adding AI is a purchase order. You buy the tool, point it at the workflow as it runs today, and hope it quietly sorts the mess out on its own. Being ready for AI means the operating system underneath the tool can actually hold it up. Take a single website as the example, since it's the anchor for most companies' GTM effort: Marketing drives traffic to it, Product Marketing uses it for messaging, Product uses it for self-serve conversion, Sales leans on the pricing and enablement pages, and Customer Success uses the live chat and case studies. If nobody can say which of those handoffs is actually leaking revenue, an agent sitting on top of all of it doesn't fix anything. It just does the same leaking work at machine speed.

This is also the theme that came out of a GTM call I had recently with a partner who builds "second-brain" systems for executives, preserving the decisions, SOPs, context, and next steps so the why stays in the building after a person leaves. My work tends to end at finding the leak and fixing it. His picks up right where mine ends, making that fix stick as context the rest of the team can actually reuse. Different lens, same order every time: diagnose, fix, write it down, then hand it to AI. Skip any of the first three steps, and you're not saving time, you're just paying to make your existing problem run faster.

Related: How to run an AI readiness diagnostic

Reach Isn't the Signal That Matters in Content Marketing. Engagement Quality Is.

Chasing whatever topic is currently rewarded doesn't necessarily shrink your reach. It can quietly wreck content's ROI as an acquisition channel by filling your inbound with the wrong audience.

I noticed a real pattern in my own posts this week. Content that leans into the current AI-hype wave reaches people outside my existing network. Content that makes a less convenient, more substantive point, like the fact that most teams still can't prove ROI from AI, tends to stay closer to the people who already follow me. My first instinct was to read that as the algorithm punishing substance. The more useful question isn't reach, though. It's who the reach is actually made of.

Content marketing is supposed to be one of the cheapest acquisition channels a company has: no media spend, just time and expertise. That cost advantage disappears the moment the audience you're reaching isn't your buyer. A post that rides an approved trend can genuinely pull in more impressions and more profile views. That says nothing about whether those impressions turn into a qualified conversation. If the tradeoff is more reach in exchange for an audience that was never going to become pipeline, that's not a free channel anymore. It's an expensive one wearing a free one's clothes.

This is the exact trap a lot of operators fall into chasing whatever the loudest approved topic of the month happens to be (a couple of months ago, everyone on this platform was suddenly a GTM Engineer). Optimizing purely for reach volume pulls attention toward content that performs and away from content that actually resonates with the specific people who'd buy from you. The real diagnostic was never "how many people saw this." It's "how many of the people who saw this were the right people," and that's a harder number to track, but the only one that actually predicts whether content marketing is lowering your CAC or quietly raising it.

Related: How to reduce CAC without cutting spend

What Should an Agent Never Be Allowed to Do?

Whatever the answer is, it needs to exist in writing before the agent ships, not after it does something nobody can explain.

I've spent a long time being the cautious one in the AI room, and it hasn't historically been a popular seat to hold. My focus has stayed on educating teams on governance: human-in-the-loop checkpoints, permissions, and audit trails that can actually be replayed. So, it's been a strange week reading the headlines: a cybersecurity incident at a frontier lab, a senior product leader walking away from a major lab specifically over where the current AI trajectory is heading, and reports of agents built to talk their way past a human approval step.

Here's what I find almost oxymoronic about all of it. The general pattern is frontier labs acting surprised that models built to override human judgment might, in fact, override human judgment. Governance-minded operators have been saying exactly this for a while, usually to polite nods in a room and then a change of subject. My honest read, and it's only my read, is that at least some of this alarm from the labs doubles as a PR play, one that keeps the market convinced the only responsible move is to fund the race even faster.

What I'm actually watching right now is a pendulum swinging from both directions at once. Teams that were already hesitant read one scary headline and freeze completely. Nothing ships, and the broken process underneath stays exactly as broken as it was before AI ever entered the conversation. Meanwhile, teams that went full throttle into the race are pulling back hard, starting to realize the importance of the unsexy work a lot of technical operators have been quietly doing the entire time: read versus write permissions on an API, one-way versus bidirectional syncs, which data should be connected and which should stay firmly walled off. A lot of teams simply forgot you can't AI your way out of that work. The ones who went full throttle are now asking who approved each agent, what credentials it's actually running on, and how they'd roll it back if they needed to. Agent sprawl turns out to be worse than tool sprawl in one specific way: there's usually no paper trail at all.

In both versions of this story, fearful and reckless, a board member or investor eventually asks what the agents in the business are actually doing, and almost nobody in either camp can give a straight answer yet. Fear and recklessness are the same mistake wearing different marketing.

The teams I'd actually bet on right now sit in the deliberately boring middle: one agent at a time, a written boundary on what it can never be allowed to do (apparently you now have to monitor for an agent attempting to talk its way around that boundary too), a human approving anything that touches revenue, and a log you can actually replay when the agent does something odd. None of that is exciting to write about, which is exactly the point from the last section. What's missing on both sides of this pendulum isn't more caution or more speed. It's education, plainly, and that's the work I keep leaning into.

Related: AI agent governance

73% of Teams Need Training and Implementation Help. Here's Why.

Buying the tool was never the hard part. Most of the 80% of teams seeing no ROI from AI don't need a third platform, they need someone to teach them how to use the one they already bought.

People ask me fairly often how I keep up with how fast the AI landscape moves. My honest answer is always the same: my own academy courses. I built Stack Finder Academy to teach this material, but I'm also very much still a student of it. Staying sharp, apparently, means re-training constantly, even on your own material.

There's a reason I built it the way I did. I've watched a lot of people land a "GTM Engineer" title with essentially zero real operator experience behind it, and that gap is exactly why the Academy leans on hands-on learning pulled from actual operator experience, not theory lifted from a blog post or a generic course shell. The track covers what most AI rollouts skip entirely: the everyday use cases that actually earn their keep in a workplace, the real basics of working with prompts, and how you move from a raw idea to something polished enough to ship.

I bring this up not to sell the Academy, it's free to start, so there isn't much to sell, but because of a stat I ran into this week that made the whole thing click into place: 73% of teams say they need training and implementation help with AI. Line that up against last week's number, that only about 20% of revenue teams are seeing measurable ROI, and you get a clean diagnosis: most of the 80% coming up empty don't need to buy a third platform. They need someone to teach them how to use the one they already bought.

Since I wrote about that ROI number, most of my calls with founders and CROs from the other 80% open the exact same way: someone reading me the list of AI tools they're already paying for. Almost nobody on those calls asks me which model is better. The actual questions I hear are far more basic: where do we even start, is our data clean enough for this, who owns it once it's live, and how do we prove to the CFO that it worked? None of those are software questions. There's no tool license you can buy that answers any of them. Somebody has to actually sit inside the workflow, get Marketing, Sales, CS, and Finance to agree on what success even means, and stay long enough to make the first version of that agreement stick.

That's the shift I'm watching play out in real time heading into the fourth quarter of this year. The tool was always the cheap part. The market is now paying for the people who can show a team how to actually use the one they already bought, which is exactly why education, training, and implementation help are starting to show up as their own line items in AI budgets, in the spot where a second platform used to sit.

Related: Who belongs on an AI implementation team

Why Does a Narrowly-Scoped AI Assistant Outperform an All-in-One Platform?

Depth beats breadth. A tool scoped to one job done well quietly lowers the labor cost baked into your CAC instead of just adding another subscription on top of it.

In the spirit of practicing what I keep preaching about narrow scope, I want to talk about the one AI tool in my own stack I'd genuinely struggle to give up. I've been using a relationship and networking assistant for a while now, and I keep finding new jobs to hand it. At this point it's sourcing inbound leads, partners, and qualified opportunities; acting as a thought partner when I want to pressure-test an idea against a wider community; helping me tell the difference between an actual commercial path and someone who just wants to talk; finding new channels worth breaking into, podcasts, speaking slots, funding conversations; managing warm introductions over email; filtering recruiting and opportunity requests, full-time or fractional; scheduling meetings, prepping the agenda, and taking the notes; debriefing my calls afterward with an honest read on next steps; sending a daily summary every morning with my calendar, relevant news, and a few thought-leadership angles worth considering; and drafting the follow-up emails, resources, reminders, and even LinkedIn post drafts so opportunities don't quietly go stale.

I call it my Chief of Staff, because that's genuinely the closest job title for what it does. I was asked to become an advisor to the product back in June, and I said yes mostly because I was already a heavy user by that point, not the other way around.

Here's the part that actually belongs in a newsletter about CAC. Every one of those jobs, sourcing, scheduling, note-taking, follow-up drafting, daily briefings, used to cost real operator hours, mine or someone I'd otherwise have to hire. Fully-loaded CAC isn't just ad spend and commission. It's every hour a human spends getting a deal from first contact to closed, including the unsexy admin work nobody puts in a deck. A narrowly-scoped tool that actually gets used replaces a meaningful chunk of those hours instead of sitting there as one more line item nobody adopted. That's the real difference between this tool and most of the AI rollouts I watch fail: this one is measurably making me cheaper to operate, not just busier.

Related: Cost per decision

What Is Time to Value, and Why Does It Matter as Much as CAC?

Time to Value is how long it takes something you deployed to actually produce the outcome you bought it for. Every shortcut in this issue quietly stretches that number instead of shrinking it.

There's no single universal formula for Time to Value the way there is for CAC, because "value" gets defined differently depending on what's being measured. For a tool, it might be the point where a team is using it without hand-holding and can point to a specific result. For a new hire, it's the point where their work starts moving a real number instead of just building context. For an AI initiative specifically, it's the gap between go-live and the moment the hypothesis from earlier in this newsletter series actually gets proven or disproven.

Here's why this belongs in an issue about the unsexy work specifically. Every shortcut described above stretches this number without anyone noticing it happening in real time. A company that buys a tool without fixing the process underneath it doesn't get to value faster, it just gets to disappointment faster, while the real Time to Value clock keeps running underneath. A team that skips training because it feels like the slow option doesn't actually save time, it just moves the time cost from before launch to after launch, usually in the form of a struggling rollout nobody wants to admit is struggling. A team that freezes out of fear, or barrels ahead recklessly and has to retrofit governance after the fact, both add real time back onto a number they thought they were racing to shrink.

Time to Value and CAC Payback Period are close cousins. A slower Time to Value doesn't just feel frustrating, it directly delays the point where the revenue or efficiency an initiative was supposed to produce actually starts showing up, which pushes payback further out on the same timeline. The unsexy work, diagnosing, fixing, documenting, training, isn't a delay bolted onto the front of the timeline. It's the only thing that's ever actually shortened it.

Full breakdown: Time to value

How Does Each of These Actually Impact Your CAC?

Here's the direct mechanism behind each pattern above: not just that it costs money, but specifically where in your CAC math it shows up.

• Skipping process readiness: An agent placed on an undiagnosed process automates the existing leak instead of closing it. CAC impact: acquisition spend keeps funding the same leaky funnel, so cost per closed deal rises even though nothing about the top of funnel changed.

• Optimizing content for reach over audience quality: Chasing whatever topic the algorithm currently rewards can grow impressions while filling the funnel with people who were never your buyer. CAC impact: content marketing, normally one of the cheapest acquisition channels available, starts costing more per qualified lead because more of the traffic it generates never converts.

• Treating governance as optional: Skipping a written boundary on what an agent can do, whether from fear or recklessness, leaves no answer when a board or customer asks what happened. CAC impact: the cleanup, the trust repair, and the compliance response after an ungoverned incident get absorbed into the same budget that's supposed to be funding new customer acquisition.

• Treating training as optional: Buying a tool without budgeting to teach a team how to use it doesn't save money; it moves the cost into slow adoption and quiet churn. CAC impact: the AI line item keeps growing every quarter while the tool sits underused, inflating the fully-loaded cost behind every deal it was supposed to make cheaper.

• Using broad, unused tools instead of narrow, adopted ones: A tool that tries to do everything but gets used for nothing doesn't replace any labor cost, it just adds a subscription on top of the labor you're still paying for. CAC impact: the hours a narrowly-scoped, well-adopted tool actually replaces are hours no longer baked into your fully-loaded acquisition cost.

This Week's Move

Pick the one process at your company that would get worse, and faster, if you automated it tomorrow morning. Diagnose it. Fix it. Write it down somewhere the next person can find it. Only then hand it to AI. Skip any of the first three steps, and you're just paying to make your existing problem faster.

Have a workflow, metric, or mistake you want broken down in a future issue? Reply to this email or send me a DM on LinkedIn.