A consensus is forming, and I think it’s right.
AI can produce code, prototypes, entire applications, faster than most people expected. So the hard part of software has moved. It isn’t production anymore. It’s confidence — whether the thing can be verified, governed, still understood two years from now, changed without fear. A recent Thoughtworks report put it neatly: the bottleneck has moved from creation to confidence.
What usually follows is a list of practices. Design systems that stay understandable. Maintain architectural integrity. Build good feedback loops between humans and AI. Make sure engineers exercise judgment rather than rubber-stamping machine output.
I agree with every item on that list. My problem is the phrase “make sure.”
Because you can’t make sure. Not by requiring it, through process or authority. And — this is the part I’ve come to believe strongly — not by paying for it either.
Incentives are short
I don’t believe people can really be incentivized. Or more precisely: incentives are real, but they are usually short-lived. Most of what we call an incentive is a one-time purchase of a behavior. Pay a bonus for a metric and you will get the metric, for a while, often at the expense of the thing the metric was standing in for. And even if the bonus keeps coming, the person underneath is unchanged. They haven’t been transformed by the incentive — they’ve simply learned to expect more of it.
What does change people is environment.
Environments come in two kinds. The soft environment is culture: mission, values, the way an organization actually behaves as opposed to the way it describes itself. The hard environment is structure: org design, cost model, the processes that stay fixed whether anyone likes them or not. A person inside a company is changed by both. The hard environment works faster and more directly — but it exists in service of the soft one, or is built to make the soft one real. And the longer someone stays, the deeper and more permanent the change becomes.
That belief is why our company is built the way it is.
70/10/20 is not an incentive scheme
We publish our cost structure: 70% of what a client pays goes to the engineer building their product, 10% to the Account Owner who owns the relationship, 20% to the platform.
Almost everyone reads that as an incentive scheme — a generous one, designed to motivate developers. Our own HR team read it that way for years. So did I, at the beginning.
It isn’t one. It’s a structure.
An incentive sits on top of an organization and pushes behavior around inside it. A structure decides what the organization looks like: it serves the mission and the values, and it determines how the business actually runs. Change an incentive and behavior shifts. Change the structure and you’ve changed the environment — and the environment changes people.
So look at what 70/10/20 actually determines. At 70% to the engineer, there is no money left to fund an internal management layer. The engineer answers directly to the client, with no internal “boss” above them. And there is no line in the allocation where engineers with work carry engineers without it — so people put everything into making the project succeed, idle time falls, and it falls low enough that the platform’s 20% can absorb what remains without touching the 70% that goes to an engineer serving a client. That isn’t a perk we grant anyone. It’s arithmetic.
Nobody here gets a bonus for understanding a client’s business. We have no internal KPI system at all — the Account Owner’s 10% isn’t a performance score either; it’s the same structure seen from a different seat. And yet our engineers routinely become the people who understand a client’s business best — sometimes better than the client’s own staff, because they’ve often been on that product longer. Not because we rewarded it. Because after five or ten years in that seat, it isn’t possible not to.
What an environment produces
Consider what that seat is actually like. You work far from the client. Nobody hands you your week — you plan it. Nobody manages you — you manage yourself. And you have to stay honest, because you’re the only source of truth the client has, and one concealment breaks the thing the whole relationship rests on.
Spend years there and something happens to a person. They develop ownership. They start thinking from the whole picture rather than their assigned slice, because nobody else is holding the whole picture for them. They make judgment calls, and they carry the consequences.
A lot of people assume engineers like this don’t exist — that the few with those qualities were promoted into management long ago and stopped writing code. If we found such people by filtering for them, then our value would be finding them, and we’d be a recruitment firm. The truth is we have no ability to detect business judgment in an interview. What we have is an environment that produces it, and people who — having grown up inside it — choose to stay.
Both halves of that sentence matter: the environment shapes the person, and the person chooses. It’s also why our engineers are rarely poached successfully. It isn’t a portable trait sitting inside an individual. It’s what those individuals became, in a particular environment, together.
Why this can’t be copied
An incentive scheme can be copied in an afternoon. Any competitor could announce a revenue share or a new bonus plan tomorrow morning, and some will.
A structure can’t be copied that way, because copying ours means dismantling the internal logic their business runs on: the management hierarchy, and the cost model in which margin from billable engineers carries everyone who isn’t billable. That is not a decision about generosity. It’s a decision about whether to become a different company.
It’s the same point I made a few weeks ago about AI-native being a structural claim rather than a tooling one. Tools copy-paste. Structure doesn’t. The question that will matter in this AI transition was never which firm adopted which model. It’s which firms rebuilt themselves around what AI made possible — and which kept the hierarchy, kept the cost model, and added a label.
Back to judgment
So: judgment is the bottleneck now. Everyone agrees.
You won’t get it from a practices document. You won’t get it from a bonus. You get it from people who have spent years in a seat where the outcome was unambiguously theirs — where there was no one to hand the problem to and no one to hide behind — and who, having become that kind of engineer, chose to stay.
That is the only mechanism I know of that reliably produces judgment. It’s slow, it can’t be bought, and it is the whole reason we look the way we do.
Jiwei Zhang is the founder of Core70 — a Shinetech company. The 70/10/20 model and full pricing are published at core70.com/pricing.