AI May Make Mission-Critical Software More Valuable, Not Less
The common reading of cheap software is that serious software companies are finished. The more defensible reading is that the scarce part was never the code.
The loose version of the AI argument ends somewhere near “software is finished”. It is worth stating plainly that this does not follow, and that the opposite conclusion is at least as defensible.
Code and system value are different things
When people say software is becoming cheap, they mean code is becoming cheap. Those are not the same claim, and conflating them is the source of most of the confusion.
Nobody buys a mission-critical system for its code. They buy the thing that has run correctly for eleven years inside a process that cannot pause, that the auditors already accept, that the machine on the shop floor is wired into, that eight hundred people are trained on, and that somebody will answer the phone about at two in the morning.
A frontier model can reproduce the functionality. It cannot reproduce the eleven years.
Code becoming abundant does not make reliable systems abundant.
What actually makes a system hard to replace
The list is not mysterious, and it has almost nothing to do with the quality of the implementation.
- Installed base, and the operational context that comes with it
- Deep integration into a workflow that would have to be renegotiated, not merely reconnected
- Operational reliability demonstrated across years of edge cases nobody documented
- Domain knowledge encoded as behaviour, most of which nobody could restate as a requirement
- Proprietary data produced by operating, including its corrections and exceptions
- Physical integration with equipment, and the approval cycles attached to it
- Security posture and auditability that have already survived an audit
- Long lifecycle support: someone answerable for a system that must run for a decade
- Trust, which in high-consequence environments is a purchasing criterion rather than a sentiment
Every one of those is an accumulation. None of them is generated. And in a world where generation is nearly free, accumulation is precisely the thing that becomes relatively scarcer.
The distance between plausible and trustworthy
AI compresses the cost of producing something that demonstrates well. It compresses far less the cost of producing something an operation can depend on — because that second cost is dominated by the work of finding out what is wrong, in an environment where being wrong is expensive.
That distance was always the real barrier to entry in serious software. Cheap generation does not shorten it; it lengthens the queue of things that have crossed the first threshold and not the second.
Where the value does compress
None of this is an argument that nothing changes. Two things compress hard, and pretending otherwise would be the mirror error.
Commodity functionality
Anything a customer could plausibly have generated, or would accept from a cheaper source, loses its price. Feature comparisons stop being a defensible basis for a premium.
Software sold beside the work
A tool used alongside a process, rather than one the process runs through, has always been the easier thing to replace. Cheap generation makes replacement cheaper still.
The practical consequence is a widening spread rather than a general decline: reproducible functionality falls in value while dependable systems hold or gain, and a great many products sit somewhere in between and will have to choose.
What this means for a buyer
The diligence question is not whether the code is good. It is which of the accumulations above the company actually has, and how long each would take a funded challenger to build.
That is a question with concrete answers. An installed base is countable. Certifications have queue times. Workflow embedment can be tested by asking a customer what a migration would cost them in weeks of disruption rather than in licence fees. Operational history either exists or does not.
Where those answers are strong, cheap creation is a tailwind: the incumbent gets the same productivity gain and keeps the assets the gain cannot reproduce. Where they are weak, the 80% rebuild test will find it quickly.
Is this company selling functionality, or is it selling the fact that something works?
The second is not becoming abundant. If anything, the flood of things that merely demonstrate well is what makes it valuable.