AI Didn't Eliminate Software Cost. It Moved It.

If you are the person who has to model what AI does to your software budget, you have almost certainly built the spreadsheet that everyone builds. Engineering throughput goes up. Cost per feature goes down. The lines cross somewhere pleasant, and the conclusion writes itself: software is about to get dramatically cheaper. The logic feels airtight because its first premise is simply true — code generation really has collapsed in cost.

The spreadsheet is still wrong, and it's wrong in a way that will quietly blow your forecast, because it models the cost that fell and has no row for the cost that moved.

Cost doesn't vanish. It relocates.

Here is the assumption hiding inside the model: that the cost of software lived in construction, so cheaper construction means cheaper software. The first half is right. Construction cost has genuinely cratered. But software cost was never only construction cost — it was construction plus coordination: the expense of getting an organisation to agree on what to build, reconcile competing requirements, resolve what each department actually needs, and converge on a single version of the truth. For decades, construction was so expensive that it dominated the total and coordination hid inside its shadow. We called the whole thing "the cost of building software" and attributed it, loosely, to engineering.

AI collapsed the construction half. It did nothing to the coordination half — if anything it made coordination more pressing, because now the building waits on the agreeing instead of the other way around. So the cost didn't leave the system. It relocated. It moved out of the part of the budget you were watching, construction, and into a part of the budget most organisations don't even have a line for: the cost of reaching, capturing, and maintaining agreement across the business.

Why the moved cost is harder to see than the cost that fell

Construction cost was beautifully legible. It had invoices, headcount, story points, a cloud bill. You could see it, so you could manage it, so you could celebrate cutting it. Coordination cost has almost none of that legibility. It shows up as rework that gets booked as "engineering effort." As the feature built three times that appears in the metrics as three features' worth of throughput. As the delay everyone blames on technical complexity when the actual cause was three departments quietly disagreeing for a month. The coordination cost is real and often larger than the construction cost ever was — but it is scattered, misattributed, and invisible to the very spreadsheet that's modelling the savings.

This is why organisations adopt AI, watch their construction cost fall exactly as predicted, and then find that total delivery cost stubbornly refuses to drop the way the model promised. The saving was real. It was just offset by a cost that was always there, that AI made more acute, and that nobody put a row for. You cut the visible half and the invisible half grew to meet it.

The cost you can't see is the one you can't cut

There's a hard corollary here for anyone holding the budget. You cannot manage a cost you don't measure, and you certainly can't cut it. As long as coordination cost stays disguised as engineering effort, every efficiency initiative will keep aiming at construction — which is already cheap and getting cheaper — and keep missing the place the money now actually goes. You'll optimise the half that fell and never touch the half that grew, and you'll wonder why the forecast never materialises.

Making that cost visible is the precondition for doing anything about it, and making it visible means treating the agreement as a real, trackable artifact rather than an informal activity that dissolves into engineering time. When the wrong build is attributed to its actual cause — an agreement that was never reconciled — coordination cost stops hiding inside the construction line and becomes a thing you can finally see, budget, and reduce. That's a large part of what Propel is for: to move the agreement out of the invisible part of the ledger and into the part you can manage. You reduce coordination cost the way you reduced construction cost — by measuring it first.

Rebuild the spreadsheet

If you take one thing from this, make it a correction to the model. AI did not eliminate the cost of software. It eliminated most of the cost of building it and left the cost of agreeing what to build entirely intact — then raised the pressure on it. Your budget didn't get smaller. Its centre of gravity moved, from a line you watch closely to a line you probably don't have. The organisations that win the AI-native era will be the ones that notice the cost moved and start managing it where it actually went, instead of celebrating savings on the half that was never the problem.

Where the money actually went

• Construction: down sharply. Cursor, Claude Code, Codex and Copilot are the visible saving, and the easy one to celebrate.
• Coordination: unchanged, and now the larger share. Meetings, clarifications, sign-off, and the arguments about what was meant.
• Rework: rising. Wrong requirements now reach production faster than they ever could before.
• Net: the budget did not shrink. Its centre of gravity moved to the line most finance models do not have.