Videos — Page 5

Lightly edited for clarity.

When people think about AI reading their website, they usually expect that the first thing that changes is traffic or rankings, but it’s not really. What usually breaks, um, what starts to break earlier than that is accuracy. And, like, not on the webpage itself, because the page is still the same, right? But the way that AI pulls that content, the way that it’s understood, and the way it’s reused once it’s combined with other sources.

So an AI tool doesn’t read your website the same way a person does, and it’s not moving from top to bottom and thinking about intent or filling in gaps based on experience. Well, actually, LLMs do fill in gaps, but not based on experience. The LLMs extract fragments from your content, and then they use those fragments as new material to generate a response. And if those fragments are not complete, or they’re loosely defined, or they’re inconsistent across different parts of the site, then the system doesn’t stop and ask which one of those you meant. It keeps going, and it’ll fill in whatever is missing with what’s statistically probable, based on the rest of what it has in its context.

And so that’s how marketing language turns into very ambiguous input. Assumptions become these implied facts, and small inconsistencies that are on different pages, so they were never meant to be read together, start to get merged into a single explanation that no one on your team would ever consciously write.

And the difficult part is that none of this looks like a clean failure, and, in fact, most companies don’t even realize that it’s happening. You don’t see a broken page. There’s not an obvious drop in performance. What disappears instead is something that’s a lot harder to notice unless you’re actively measuring for it. It’s control over how your business is described when someone else, or something else in this case, is doing the explaining.

Lightly edited for clarity.

Most teams haven’t changed the way they build their website, even though the way that content is being consumed has already changed and will likely continue to change.

Websites are still organized around pages, product pages, landing pages, about pages, funnels, because that model works really well when a reader is going to your website and moving through the site.

But AI systems don’t experience your website in the same way. They pull information in pieces, often from multiple URLs at once, and then try to reconcile those pieces into a single explanation.

So when the definition lives on one page and the positioning language lives on another, and constraints or edge cases are buried somewhere else, the system doesn’t know which one is supposed to anchor the others. It just tries to make them all fit together.

And that’s where problems begin. Not because the pages are inherently bad, and not because funnels don’t matter, but because content is now being consumed in fragments while it’s still being written as if someone’s gonna read the entire page.

So when you’re organizing your content on your site, your concepts need to be clearly defined. You need to consistently describe them and scope them explicitly wherever they appear, because that’s what allows those fragments to be used and recombined without losing their meaning.

Lightly edited for clarity.

The more I work with AI, the less convinced I am that the intelligence of the model is currently the bottleneck.

The models have improved incredibly quickly, right? A few years ago, it felt kind of like a high school student, and today they feel more like pretty capable professionals. And we can debate how smart the models are, but it’s obvious how quickly they’ve improved. And it makes me wonder how much additional intelligence will actually change the equation.

If I think about why experienced people are valuable, it’s easy to say experience, but what does that really mean? Some of it lives in people’s heads. A lot of it lives in people’s heads. But a surprising amount of it actually gets written down, like architecture decisions and design documents, emails, meeting notes, code, project plans, lessons learned, and retros.

People naturally accumulate knowledge by doing the work, and AI doesn’t. But that doesn’t mean that AI has to learn it the same way that we do.

If every design decision, trade-off, meeting, project success story, and failure, if all those things become searchable and referenceable, how much smarter do the models actually need to be?

The work we’re gonna be doing over the next decade, the integration of AI into business processes and into people’s lives, is taking decades of accumulated organizational knowledge and making it available to models, not just making the models smarter and smarter.

Lightly edited for clarity.

This weekend I was working on some stuff, and I exceeded my GitHub Copilot token limit.

I was mostly finished, you know, the old way, before AI. And I haven’t forgotten how to code. I haven’t forgotten how to write software. But after spending the last couple of years mostly working with AI every day, it felt really slow.

And it reminded me of something. Writing code is not really the most difficult part of software development, but it traditionally has been the most time-consuming. The hard part’s always been figuring out what to build and exactly how to build it. Understanding the requirements, finding the edge cases, realizing business rules contradict each other, or figuring out that one small change here breaks things elsewhere. And that’s where experienced engineers spend most of their time.

AI has fundamentally changed the economics of software development. Writing code is dramatically faster than it used to be, but it doesn’t eliminate the need for good decisions. And in fact, it does the opposite. The easier it becomes to build something, the more important it becomes to know exactly what should be built.

You can’t hand AI a prompt that says, “Build me a website. It should be cool.” You have to tell it exactly what success looks like.

Lightly edited for clarity.

One of the weird things I’ve noticed about AI is that it sometimes creates work faster than it automates it, and I’ve run into this three times recently.

The first was a website performance project. We wanted to reduce cumulative layout shift, which is when you load a web page, the pieces and images move around. That’s called layout shift. So the way to fix that in a CMS-driven website is to specify the correct dimensions for the images. And a few years ago, that would have meant a lot of manual testing and manual review. Instead, I generated the code changes with AI. I built a crawler with AI, I generated screenshots of the entire site, and then I compared the before and after results with AI, and then fixed the outliers, again with AI.

The second was an SEO project. We wanted every URL on the site to use the correct canonical format, and it required updates to both code and content. So again, instead of reviewing pages manually, we crawled the entire site, generated a report of every issue, fed that report back into AI, found every place URLs were rendered, and then fixed them systematically.

And I see the same thing happening in marketing. What used to be one campaign is now becoming several: checking accounts for young adults and retirees and small businesses, and for different geographies.

I realized these are all the same story. AI makes generation really cheap, and when generation becomes cheap, you do more of it. And that creates more code to test, more content to review, more campaigns to validate, more experiments to measure.

AI is automating a lot of tasks in jobs right now, but it is also creating more work to do.

Lightly edited for clarity.

I spend all day, like all day, working with AI, and I still haven’t seen an AI that could replace any of my team members.

What I’ve instead seen is people becoming dramatically more capable because of AI. Not replaced, but amplified.

And the conversation around AI often gets framed as humans versus AI, but there’s a much more reasonable possibility, which is AI and humans integrating and working together.

I haven’t seen AI replace developers or architects or project managers, or business analysts, marketers, any role that I work with regularly.

But what I have seen is people becoming more productive, more capable, taking on larger problems, expanding their skill sets.

The future isn’t AI taking over everything that people do. The future is people becoming much more capable because they use AI effectively.

Lightly edited for clarity.

So I started this week talking about how quickly AI is getting smarter, and I’m ending the week with how intelligent people really are.

I spent a lot of time this past week with my team on multiple client projects and software delivery challenges. And it’s not about how much AI can do, but it’s about how much judgment people apply subtly without even really thinking about it.

AI is incredible at generating code and content and synthesizing information and finding patterns in things. But this week, as I watched people recognize incomplete requirements, spot contradictions before they became problems, balance competing priorities, navigate organizational politics, and ask questions that completely changed the direction of a conversation, I noticed that none of those things looked particularly remarkable while they were happening.

The more experienced someone is, the more invisible their expertise becomes.

At the beginning of this week, I wondered whether AI was approaching the capability of a mid-career professional. And by the end of this week, I wasn’t thinking nearly as much about AI. I was thinking about how much of professional expertise is really judgment.

The more I work with AI, the less impressed I am with memorized knowledge, and the more impressed I am with human judgment and nuance and understanding. And those things are a lot harder to automate than people think.

Lightly edited for clarity.

Last week, my team at Sagepath Reply and I ran an AI enablement workshop with one of our clients using Optimizely Opal. Opal is Optimizely’s agentic marketing platform that sits on top of the Optimizely ecosystem and integrates really well with their tools.

We don’t start by talking about AI at all. We start by opening a FigJam board and listing what the teams responsible for campaigns, trade shows, email marketing, website content, and product launches do. Then we break those responsibilities down into smaller, individual tasks underneath them, like writing emails, building landing pages, or preparing for trade shows.

Even within those, there are smaller tasks. If you’re building a landing page, you need to create copy and assets, consider the audience, personalize, and analyze click-through rate and marketing uplift.

Once that’s mostly mapped out, we start asking questions. Which of these things are repetitive? Which happen constantly? Which are tedious? Those become our first candidates for automation because nobody in the room is trying to replace marketers. We’re trying to look at the process of marketing, break it down, analyze it, and then rebuild that process in a more efficient way.

Lightly edited for clarity.

So when we’re running these AI enablement sessions, once we’ve identified the first candidates for automation, we start building Opal agents to handle them. And things like landing pages, campaign emails, website updates, they’re good candidates because they’re pretty structured, they’re repeatable, and they never really stop coming.

And so as we start talking about automating those tasks, people don’t really ask, okay, well what work is left for me? It’s pretty obvious what work is left for them. What they start asking is, what work’s never really been possible before, that is now?

So what if competitor research wasn’t something you did once a quarter, or once a year, or never? What if voice of customer analysis was always current, instead of based on the last survey that you did, which was who knows how long ago? What if every campaign started with current market research, instead of whatever you were able to get together, in the schedule that was allotted for it?

Before AI, a lot of that work wasn’t economical. You couldn’t justify spending days researching every competitor before every campaign or continuously analyzing customer feedback. The time just wasn’t there.

AI changes that equation. Work that used to be too expensive or too slow, or too tedious becomes practical.

And so, AI is not just helping marketing teams finish the same work faster. It’s changing what a marketing team is capable of doing.

Lightly edited for clarity.

So by the end of the workshop, we’d already started building some automations, and what stuck with me was how the conversation changed. Instead of asking, how do we save time, people started asking, what should we do with the time we get back? And the answer is not really work less, sorry. It was run more campaigns, experiment more often, spend more time understanding your customers, continuously analyze competitors instead of just reviewing them every few months, invest more time in figuring out what’s actually working instead of just simply producing more marketing assets.

And it’s a completely different way of running a marketing organization. AI changes the economics of work, and as that happens, the work that creates the most value changes too. And that’s why I keep saying AI automates tasks, not roles. The tasks change, the role gets redesigned, and I think that’s one of the biggest opportunities that AI creates.

Lightly edited for clarity.

My dad tells me, you can do anything you want, but you can’t do everything you want. And I swear I repeat that to myself every day.

Uh man, I’ve had a lot of really good conversations over the past couple weeks, and with really smart, interesting people. And almost every conversation created another possible direction, you know, someone I could introduce, or a project we could explore together, or another conversation we could have, or an idea I’d like to spend more time thinking about. And I genuinely would like to pursue a lot of them.

But every time you say yes to something, you’re deciding where your time and attention go. And if you try to keep every possibility alive, eventually, you make commitments that you can’t give enough attention to.

And sometimes a good conversation can just be a good conversation. You can appreciate the person and stay connected, and leave room for the right reason to work together later.

So yeah, you can do anything you want, but you have to choose because you can’t do everything you want.

Lightly edited for clarity.

Someone I talked with this past week described an interesting use case involving feature flags. They have this recurring process where several different feature flags, that are split across a few different systems, need to be turned on or off. And accuracy matters because if you miss some of them, it means the process wasn’t completed correctly.

So he built a tool to handle it, and nobody’s job disappeared. The person responsible for the task still owns the outcome. The process just became faster, more reliable, and easier to maintain.

And I think that’s a useful way to look for AI use cases. You find a point of friction. You look for a task that happens repeatedly, follows understandable rules, and produces a result you can verify. And then you automate that task. You make the person doing the job more capable, you learn from it, and then you find the next point of friction.

You don’t have to start by transforming your entire company. You can start by making one frustrating part of someone’s job work better.

Lightly edited for clarity.

It’s my birthday today. I’m 50. And maybe because I’m 50, I was just thinking about life and, you know, what I spent 50 years doing. And when you look at that, it’s really easy to see the mistakes, you know, all the relationship problems that were your fault, and all the times you were mean to somebody that didn’t deserve it, all the times you lied when you should have told the truth, all the decisions that you just didn’t make.

And it’s really easy to see the things that really weren’t your fault too. You know, maybe you trusted somebody that took advantage of you, or, you know, you had an accident or medical problems, or just tough situations that, you know, you just have to deal with. They suck.

And it’s also really easy to see your successes, you know, all the good relationships that you have and the things that you built that you’re proud of and the people you helped. And I look at my life and I see all of that. And it’s a curious thing to realize that all of that is what makes you who you are.

And I think about 40-year-old Mike, and I’m a much different person than I was then, which, of course, I am. I’ve had 10 years of life experience that 40-year-old Mike didn’t have. And it’s important to understand that 10 years from now, you’ll be a much different person too, because you’ll have had 10 years of life experience that you currently don’t have.

But you have the opportunity to have a say in what those life experiences are gonna be. And you can’t control all of them, but you can definitely direct them. You can choose what kind of experiences you wanna have.

And, uh, I’m exhausted. I have so much to do, but I’m excited to do it because I’m choosing to curate the experiences that I think I need to have. And I need to learn from to become the kind of person that I wanna be, to become better and more capable. So I’m gonna go get started.

Lightly edited for clarity.

AI exposes how poorly documented most organizations actually are. A lot of what makes a company work lives in people’s heads. It’s in old conversations, Slack messages, emails, decisions that everybody remembers, but nobody ever wrote down.

And then we give a task to AI and expect it to understand how our organization works, but it can’t read your mind. If you don’t give it the goal, the relevant history, your standards, your terminology, examples of good work, constraints it needs to follow, it’s gonna fill in those gaps, and you’re probably not gonna like what it fills them in with.

The same documentation that gives AI useful context also makes work easier to teach, transfer, maintain, and improve.

So when AI produces something that doesn’t fit your organization, it’s likely not the model that’s the problem. You should ask whether your organization ever documented what good actually looks like.

Lightly edited for clarity.

I met something like 50 new people in the last few weeks. It’s been really amazing. Last week I went to a meetup, and I met a bunch of solo founders, people that are building small businesses.

And it’s interesting how many of them start with domain expertise, not really technical know-how. They understand a specific industry, and they’ve worked inside a particular process. They’ve experienced a problem enough times, know exactly what was broken and what a useful solution would be.

And even a year ago, turning an idea like that into a software business might have involved raising money, finding a technical co-founder, or hiring a development team before you can really even find out whether or not that idea worked.

And AI’s lowered the barrier dramatically. People who understand the problem can now build a product, and test it with real users, and start a business in way less time and with way less money than was practical even a year ago.

And we’re gonna see a lot more businesses created this way. The person who deeply understands the problem can increasingly become the person who builds a solution.