In the previous post I wrote about watching someone build working software for his own business without being a developer.
That meetup didn't hand me a new question. It lined up with one I'd already been carrying since Drupal Pivot.
Companies are building things themselves now, and a lot of them feel like they don't need us. That's not a new question for me, and it isn't new for anybody else in this business either. We've had two answers already, and I believed both of them at the time. That's the part I want to be straight about.
The first answer was that AI would do everything.
That was the mood when the models got good enough to surprise us. Whatever the work was, you'd describe it and the machine would produce it, and the job became knowing what to ask for. I spent real time in that phase. I built prompts and frameworks and got genuinely useful results, and I also watched confident output turn out to be wrong often enough that I stopped trusting it on its own.
That answer didn't survive contact with anything complicated. Not because the models were bad. Because a description of what you want is not the same as a system that runs, and the gap between those two things turned out to be most of the job.
The second answer was agentic workflows.
This one made more sense, and I think it's still mostly right. Instead of one model doing everything, you connect things. A workflow tool in the middle, agents doing specific steps, systems talking to each other, a human somewhere in the loop. It works. What I build is a version of that for my own work, plugins and skills rather than automation for clients.
But it stopped being a safe place to stand, for two reasons.
The first is that it's being deployed much faster than anyone is governing it. Gartner expects more than 40% of agentic AI projects to be canceled by the end of 2027, mostly for reasons that have nothing to do with how good the model is. And when Retool surveyed 307 CTOs, CIOs and CISOs this year, 8% of them rated their AI governance as strong. Eight.
That isn't the technology failing. It's that shipping an agent and being able to answer for it turned out to be two different jobs, and most people only budgeted for the first one.
The second reason is more uncomfortable, and it's the one that actually stops me.
That road is already full.
Automation workflows are cheap to learn. The templates are free. The tooling is genuinely good and getting better, which means the barrier that used to protect this kind of work is mostly gone. There are now guides telling you how to start an AI automation agency and reach ten thousand a month, and when a service category has get-rich-quick content around it, the margin left in it is already thin.
So being able to build agentic workflows isn't a differentiator anymore. Everyone has it, and everyone got it fast.
Which is the same thing that happened to building, one layer up. Creation got easy. Selling got hard. We solved for the part that stopped being the problem.
Creation got easy. Selling got hard. We solved for the part that stopped being the problem.
At Drupal Pivot, everyone agreed on what's ending. Time and materials is finished. Fixed price means committing to an outcome instead of to effort. Nobody answered the other half. If that's what's ending, what are we selling?
I don't think that was avoidance. I think it's genuinely unresolved, and the industry's own numbers say so. When Promethean Research surveyed agency leaders about the year ahead, the top three concerns were sales, margins, and lead generation. Not delivery. Not talent. Not technology. Selling.
That's the whole thing, in the industry's own words. We are not short of capability. We're short of something people will pay for.
We are not short of capability. We're short of something people will pay for.
I've heard people take this further and say SaaS is dead, or that products are over, and I don't buy it. Plenty of software still gets sold. What's changed is that finding the right thing to build has become far harder than building it.
So here's where I actually am, coming out of these two waves.
The first answer, that AI does everything, was wrong about how much of the work is the description. The second answer, that we sell agentic workflows, isn't wrong. It's just crowded, and crowded is its own kind of wrong when you need to make a living.
Which leaves me looking for something that stayed difficult. Not because the tools aren't there yet, since that kind of difficulty keeps expiring. Difficult in a way that doesn't get easier when the next model ships.
I think I found one, and I found it in a conversation about hotels. That's what I want to write about next.