AGI Won’t Make Software Companies Interchangeable
I expect AI to make software dramatically cheaper to produce. I also expect software companies to take responsibility for much more of their customers’ work. The prediction that AGI will commoditize the entire software industry misses how those two developments fit together.
The argument usually begins with something plausible: if anyone can ask an AI to build an application, building that application becomes less valuable. But then it holds the application constant. It imagines tomorrow’s customers buying today’s products, with the only change being that anyone can manufacture them.
Competition changes what customers expect to buy. The company that once sold a narrow tool has to offer more as that tool becomes easier to reproduce. Its competitors have the same technology, and its customers have new alternatives. The standard rises.
This is where specialization matters. Repeatedly working on a particular class of problems lets an organization improve its methods, discover failure modes and spread the cost of what it learns across many customers. That can produce better work at a lower cost. The gains accumulate in processes, data, relationships and working systems.
Now grant the strongest version of the AGI premise: a broadly available intelligence that can outperform every human specialist, at a lower cost. The scarcity of individual expertise would collapse across many tasks. A skill that took someone twenty years to acquire might become available to anyone with access to a model.
That still leaves a distinction between an individual’s abilities and a firm’s accumulated capital.
A company carries knowledge of what its customers need, records of what has gone wrong, access to particular systems, established distribution, contracts and operational routines. Some of that knowledge is written down. Some is embedded in the product. Some exists in the way the company measures results and corrects mistakes. Having access to an intelligence capable of understanding all of this does not automatically give a competitor possession of it.
An exceptionally capable model can reason about a customer’s business. It still needs the relevant information, permission to act and a way to observe whether its actions worked. Intelligence can accelerate the acquisition of those assets. It does not make every firm start with the same assets.
There is an established economic idea here. In The Productivity J-Curve, Erik Brynjolfsson, Daniel Rock and Chad Syverson describe how general purpose technologies are accompanied by investments in new processes, products, business models and human capital. Those investments create valuable intangible assets within firms. The research concerns historical technology adoption; the extension to AGI is an argument we still have to make. But it gives us a reason to take organizational capital seriously when thinking about technological change.
I expect that capital to change considerably. Old processes will become obsolete. Knowledge that once required years of experience will become easy to reproduce. Some firms will discover that the asset they thought they owned was mostly an expensive workaround for a limitation that no longer exists. Others will use the new technology to accumulate more useful knowledge and build better systems. Specialization can move from people to the organizations and systems applying the intelligence.
The build-versus-buy argument also has to account for both sides using AI. A customer may be able to build a satisfactory internal application far more cheaply. A specialized vendor may be able to build, operate and improve it more cheaply too, while spreading common costs across thousands of customers. The relevant comparison is between those two improved alternatives.
Being capable of doing something yourself does not make doing it yourself the cheapest choice. Even a company full of superhuman agents would have to decide how to allocate compute, acquire information and maintain systems. As long as those activities have costs, there is room for suppliers that perform particular work more efficiently. The supplier might have very few employees. It would still be a specialized firm.
Take the 10,000-fold productivity premise seriously. Suppose AGI allows a software company to produce 10,000 times as much useful work from the same budget. A company that uses that gain simply to keep selling its current product at its current price invites competition. Customers will expect lower prices, more capability, better results, or some combination of the three.
The number is a thought experiment, not an exchange rate between productivity and customer value. Ten thousand times as much code is not ten thousand times as useful. Demand, implementation costs and competition determine how the gains are divided. But the direction of pressure is clear: the amount a customer expects for a given budget rises.
That pressure gives SaaS companies a reason to expand the job they do.
Consider accounting software. A product might begin by recording transactions and helping someone reconcile the books. With much greater capabilities, a supplier could take on reconciliation, investigate discrepancies, collect missing documents, prepare filings and coordinate resolution of exceptions. Each additional responsibility involves information, integrations and evidence that the work was completed correctly. Customers can benefit from having those responsibilities handled together.
The valuable expansion would be measured in work the customer no longer has to do. A product could become much more capable while presenting fewer screens and requiring fewer decisions. Firms with deep knowledge of a workflow would have a reason to extend into adjacent work where that knowledge remains useful.
There is a limit to what this argument establishes. Specialization alone does not prevent commodification; wheat producers specialize too. Software firms remain differentiated only where their accumulated assets produce differences customers care about, and where competitors cannot instantly reproduce those differences. Generic tools, standard components and some entire product categories could become commodities. Some software companies will disappear.
Nor does a larger offering guarantee a larger profit. Customers may capture much of the productivity gain. Prices and margins can fall while products improve. This is a case for the continued economic role of differentiated software firms, not a guarantee that today’s vendors keep their valuations or their customers.
For firm-level commodification to follow from AGI, the argument needs more than universally available intelligence. It needs the relevant information, access, operating capabilities and commercial relationships to become equally available, or to stop mattering. Perhaps sufficiently advanced technology erodes many of those differences. That is a further claim, with further assumptions. It does not follow merely from a model becoming smarter than every person in the company.
My expectation is that software firms will compete to take responsibility for larger portions of their customers’ work. Their boundaries will change. They will employ different numbers of people, charge in different ways and absorb work that currently sits outside the product. The firms that succeed will keep accumulating useful knowledge and applying it across a growing range of problems.
If AGI makes today’s software trivial to produce, today’s software becomes a poor description of what a customer will pay for. The subscription has to earn its place against everything else that has become possible. I expect that to make software companies deliver much more.