AI Moved Software's Humans Upstream

The dominant emotion in software right now is a quiet dread, and it has a specific shape: the machine is coming for the work. If an AI can write the code, what is the human for? It's a reasonable fear to feel and, we think, a misreading of what's actually happening. AI is not removing humans from software. It is moving them to a different part of it — and the part it's moving them to is the part that was always the hardest and most consequential.

Where the humans used to be

For most of the industry's history, human effort concentrated at the point of construction. The keystroke. The implementation. You took a more-or-less understood intention and did the difficult, skilled labour of turning it into a working system. This was where developers spent their days, where their expertise was measured, and where the bottleneck sat. So naturally, when we imagine AI replacing developers, we imagine it replacing that — the construction. And it does. That part is real.

But construction was never where software succeeded or failed. Software fails upstream of construction, in the murky territory of figuring out what to build, for whom, and whether everyone involved actually means the same thing when they say they agree. It fails in the gap between what the client asked for and what they needed, between what the business specified and what the engineers understood, between the feature as imagined and the feature as it would actually be used. Construction was the visible labour. Alignment was the invisible determinant.

The bottleneck moved, and the humans moved with it

When construction was the bottleneck, optimising it was rational. Now construction is nearly free, and the bottleneck has snapped upstream to exactly where the real difficulty always lived: deciding, correctly and collectively, what the software should be. That is now the gating constraint on delivery. Not how fast you can build — how well you can agree.

And in an enterprise, agreeing is not one person thinking clearly. It's many parts of an organisation reconciling genuinely different positions — the sponsor's deadline against architecture's standard, security's constraint against the delivery team's estimate, what legal will permit against what the business wants to ship. None of that is automatable, because none of it is a generation problem. A model can produce any feature you can specify. It cannot tell you whether the feature is worth specifying, navigate the tension between a stakeholder's stated want and their actual need, or hold the political reality of an organisation where three departments each believe something different. Those are human acts, and AI has made them the centre of the job rather than its preamble.

This is a promotion, not a replacement

The developer of the AI-native era is not being made obsolete. They are being moved to higher-leverage work — from implementing decisions to making them, from writing the spec's consequences to authoring the spec itself, from being measured on output to being measured on judgment. The keystroke is automated. The decision is elevated. For people who got into software to solve problems rather than to type, this is the work finally arriving at the part that was always the point.

But it comes with a demand, and the demand is uncomfortable. If the human contribution is now the agreement — the upstream decision about what's right, across an entire organisation — then that agreement has to be treated with the seriousness the work now carries. You cannot move the humans upstream and then leave the upstream ungoverned. You cannot make the decision the most valuable act in the delivery and then capture it in a meeting nobody minuted and a chat thread lost by Thursday. The value moved upstream; the instrumentation has to move with it.

Why we built for the upstream

Propel is built for where the humans went. Not for the construction — there are excellent tools for that, and more arriving constantly. For the part above it: the agreement, the decisions, the commitments, the record of which part of the organisation decided what and whether the build still honours it. The upstream is where software is now won and lost, and it has been the least instrumented part of the entire lifecycle. We moved all the humans there and gave them almost no system to work in. That's the gap Propel fills.

The anxious story says the machine is taking the work. The truer story is that the machine took the part of the work that was mechanical, and handed us back the part that was always human — the deciding, the agreeing, the being accountable for what we chose to build. That part was never going to be automated, because it was never really about software. It was about people, and about organisations, agreeing. AI just moved them to where they always mattered most.

Where the humans went, tool by tool

• Construction. Handled by Cursor, Claude Code, Codex and Copilot, with people supervising rather than typing.
• Verification. AI code review — CodeRabbit and its peers — reads every diff faster than a human can.
• Regulated review. Harvey and Legora moved legal work upstream the same way, for the same reason: the judgement was always the expensive part.
• Agreement. Deciding what to build, and recording who signed off, is now the least instrumented step in the lifecycle. That is the step Propel exists for.