
I’ve been building software for more than twenty years. Most days I’m the skeptic in the room, not an early adopter. I don’t trust most of the marketing. So that makes it a little awkward to admit that most of what I build now starts as a prompt.
This isn’t an AI hype piece, but it isn’t a warning, either. It’s a view from the middle, which is where most of us are working right now.
Here’s a scene you may recognize. Your team used AI. You got faster. Everything was going great. Then something went wrong, and the question shifted from “what did the AI do?” to the more uncomfortable “who here understands this well enough to own it?”
That question is the whole story. AI made generating code, emails, or even entire presentations cheap. It didn’t make getting it right cheap. The hard work moved downstream, into judgment, review, troubleshooting, and accountability.
I see this from two sides every day. At Laivly, my team uses AI to build a platform, and that platform puts AI to work in contact centers drafting responses and case notes alongside agents or resolving customer requests on its own. On both sides, the lesson is the same: generation is cheap. Judgment isn’t.
Review Is the New Bottleneck
“Generation is cheap” describes time and effort, not dollars. The token bill is real. (Trust me, I’ve seen it.) But the bigger cost shows up when faster output lands on the same team with the same backlog. That’s speed without direction.
As time gets less scarce, value matters more. “Can we build this?” is increasingly a yes. “Should we build this?” is a question no token budget can answer.
None of this is entirely new. Tech has long talked about the “10x engineer,” the developer who gets ten times more done than everyone else. But that was never about typing speed. It was about framing the problem, making good design calls, and knowing what not to build. I once compared code metrics with a junior developer and found I’d deleted more code than either of us had actually written. I’m still a little proud of that.
But now AI hands everyone the speed. If everyone has a 10x multiplier, the multiplier becomes the floor. Judgment becomes the ceiling.
Meanwhile, output piles up faster than anyone can review it. AI removed the effort that used to keep low-quality contributions rare. The maintainers of cURL, a free piece of software that runs on billions of devices, used to pay people who reported security flaws. After six years, its small team shut the program down. AI-generated reports looked convincing but were mostly wrong, and checking them all left no time for the real ones. The review queue is now the constraint.
And that’s true well beyond code. Anywhere AI writes more, someone still has to check the work. In a contact center, that’s agents, QA, and supervisors. Their capacity doesn’t grow just because the drafts got faster.
Feeling Fast Isn’t Being Fast
One of my favorite studies from last year is a controlled trial by METR. Experienced open-source developers predicted AI would make them 24% faster. Afterward, they felt about 20% faster. When it was measured, they were 19% slower.
Feeling faster is not the same as being faster. I think about that whenever a number on the dashboard starts moving in our favor.
That gap gets widest when something breaks. To be clear, AI is a real help in an outage. It can gather information and suggest where to look faster than any of us. What it doesn’t have is instinct. Fixing a real problem means keeping the whole system in your head while the business is on fire, and knowing which leads are worth chasing. During a recent outage, AI did the legwork well, but at one point it reinforced a bad assumption and sent us the wrong way. Someone who knows the system can often spot that and get to the answer faster. Information isn’t the same as understanding.
That’s what worries me over the long term, because understanding comes from practice. AI handles the easy first pass well, but easy cases don’t build instinct. In a 2026 study from Anthropic, developers who learned a new tool with AI help scored 17% lower on comprehension, and the biggest gap was in troubleshooting. You can’t practice a skill you’ve delegated, whether you’re tracing a system failure or solving a customer’s problem.
Don’t Confuse Assistance with Authority
The line I care most about is between assistance and authority.
A tool that flags a problem for a person to act on is assistance. A tool that acts on its own has authority. Think of AI that suggests a next step to an agent versus AI that resolves a customer’s request end to end. Both can be the right call. The difference is where the trust sits, and who answers for the outcome.
When AI assists, the person who acts on its suggestion owns the result. When AI acts on its own, that ownership doesn’t disappear. It has to be designated: who sets the limits, who watches the results, and who steps in when something goes wrong. The question changes from “can the AI do this?” to “who owns the risk when it’s wrong?”
Owning that risk means having people who can tell when something is off, often before anything looks wrong. That kind of pattern recognition comes from years of hands-on work, inheriting messy systems, and learning the hard way. Those people take a long time to develop. We should be deliberate about whether we’re still creating the conditions for the next ones.
Judgment Is the Career Now
I recently helped a student prepare for co-op applications. Using AI for a first pass on a resume was fine. The more important conversation was about how they were learning the tech stack, and whether they were getting enough practice on their own to understand and own the work.
That’s the real question for anyone starting out. Using AI is a given. Learning the underlying job is the part that still takes effort.
What AI is doing to a lot of jobs right now feels familiar to me. When I moved from implementation into leadership, my work shifted from doing to directing: reviewing, guiding, and helping other people do their best work. What used to happen over the span of a career is now happening to whole teams in a few years. That doesn’t have to mean getting less hands-on. Instead, it should mean getting better at direction, judgment, and owning the result.
If you’ve spent years getting good at something AI now does faster, those years weren’t wasted. They gave you the judgment the model doesn’t have.
One warning for leaders. AI can reduce burnout when it takes repetitive work off people’s plates. It can also create a new kind of exhaustion when people spend their days supervising systems they don’t fully trust. If every minute saved becomes a minute of more work, a productivity gain turns into a burnout engine. That’s as true for a team of agents as it is for a team of engineers.
Being Intentional with AI
I’m obviously not against AI. I spend my days building it. But I’m not interested in using it on autopilot, either. I’m trying to be intentional about it, which is harder and less exciting than hype or fear.
What’s useful now is figuring out where you still need human judgment and clear ownership. Here’s where I’d start:
- Individuals: keep doing the hard parts yourself. Let AI take the first pass, but stay hands-on with the tricky cases. That’s where instinct comes from.
- Teams: check the thinking, not just the output. Before anything goes out, make sure someone understands why it’s right and owns the result.
- Leaders: protect time for judgment. Don’t turn every minute AI saves into more volume.
- Organizations: decide where AI acts on its own. Be clear about what it can handle alone, what’s at stake if it’s wrong, and who steps in.
So here’s the question I keep coming back to. Good AI tools make your team faster and take on more of the work. That’s the point. But if you set them aside for a day, would your people still know how to handle the hard cases? Would they recognize a wrong answer when they saw one? Could they explain why a decision was made?
If any of those answers makes you a little uncomfortable, good. That’s the part worth working on. Get it right and AI doesn’t replace your team’s judgment. It gives them more room to use it.
Craig Shyjak is VP of Architecture for Laivly. This post is adapted from his keynote at Prairie Dev Con 2026.







