Squeezed Between the Hype and the Sceptics
A 5 min read
If you lead a software team right now you might be squeezed between two kinds of certainty. Above you, the hype train: AI can do anything, so why is your team so slow. Below you, the sceptics: it's a toy, it makes garbage. Both got there honestly, and both are probably overdue a second look.
I want to talk about why that happens, and the only approach I've found that actually works.
Opinions set like stone
Most people didn't form their opinion about AI tooling from nowhere. They tried it. Maybe a year ago, maybe two, they gave one of these tools a real problem and watched it make garbage. That experience was real and the conclusion was fair, at the time. The problem is what happens next: the opinion sets, and human nature makes it difficult for us to change our mind on something. We tend to be set in our opinions.
The same thing happens in the other direction. Someone watches an impressive demo and comes away just as certain, only the other way round. That opinion sets just as hard as the sceptic's.
You'll know the stuck sceptic when you hear one. They're still on "it doesn't know how many r's are in strawberry" and they can't get past it.
There is no lightning bolt
The thing that makes AI opinions go stale faster than most is that the capability doesn't jump, it creeps. I never had a big bang moment where the tooling suddenly became worth it, and I was watching closely. What I saw was a long series of small improvements that took agentic tools from frustrating, a lot of time wasting messing about, probably not worth it, to some real acceleration, and then more and more acceleration as time went on.
Early on it felt like it was '10 steps forward, 9 steps backward', then repeat. That one step of progress is easy to miss if you've already decided the whole thing is a waste of time. But the ratio kept moving, and the net gain went from barely worth it to impossible to ignore. No announcement, no version number that made it official. You were waiting for a moment that forced everyone to re-evaluate, but it never came. Opinions formed on last year's tools, or frankly even two months ago's, just quietly expire, and nothing tells you to check.
Confront opinions with ground truth
So here's the principle. You are unlikely to talk someone out of an opinion they've set. Not with benchmarks, not with enthusiasm. I've tried, and I suspect you have too.
What works is making people confront their beliefs with ground truth. For the sceptic, that means making the capability impossible to ignore: real work, done in front of them, on problems they care about. For the believer in magic, it means letting them get close enough to the details to see limits for themselves, the security questions, the integration mess, the compliance realities. You don't have to say much at all.
It's the same lever in both directions, which is the part I find satisfying.
Check the friction before you diagnose the belief
A warning though: some scepticism isn't a belief problem at all. If someone's experience of these tools comes through bad prompting, not enough tokens, a weak harness, poor model selection, then their frustration is accurate. They're not wrong about what they experienced. They're wrong about what's available.
The fix for that isn't persuasion, and it isn't a hackathon or an innovation day either. It's proper tooling with enough capacity, inside their real day to day work, and then getting out of the way. Space to evaluate for themselves converts more people than any presentation ever will.
Ground truth is slower, and that's the price
I won't pretend this approach is free. Letting reality make the point takes longer than winning a meeting would, and sometimes it means watching people, including senior people, walk towards a wall you can see perfectly well from where you're standing. Patience is the cost. You pay it because the alternative, pulling rank, doesn't actually change anyone's mind. They just stop saying it near you.
This was never about AI
The job is changing under all of us. Developers spend more of their time reviewing than writing now. The question has shifted from whether these tools can do the work at all to how much of the work you can hand over. That change will suit some people and not others, and whatever calibration you have today will be stale before long anyway.
Which is fine, because the principle was never about AI. Making people confront their opinions with ground truth is a very basic principled view, and it applies to many things in life. AI adoption just happens to be the place where, right now, opinions expire faster than anywhere else.