When Code Gets Cheaper, Who Keeps the Savings?
Cheaper software creates value, but competition decides who keeps it. The strategic question is which scarce capability turns a shared cost reduction into a private advantage.
The attractive thing about cheaper software is that your company can afford to build more of it. The awkward thing is that so can everybody else. An investment committee can celebrate a lower development bill while quietly approving the elimination of the very scarcity on which its business depended.
That is the first strategic question I would ask about AI: after the cost falls, who keeps the savings? The answer may be the company deploying it. It may also be the customer, the distributor, or the supplier that owns the part of the process nobody can replace. A productivity gain comes with no forwarding address for the profit.
Run the competitor’s spreadsheet
Imagine a distributor building an assistant that recommends purchasing orders. This is an illustrative business, not a reported implementation. Generating the interface and connecting the catalog become cheaper. The first calculation is straightforward: less development spending for a useful internal tool.
The second calculation is more interesting. If rival distributors can build comparable assistants, they can also reduce their costs. Some may pass the saving into prices to win accounts. The original company could end up with a more efficient operation and the same margin, or a smaller one. Adoption was necessary to remain competitive; it was insufficient to become more profitable.
Now change one fact. Suppose this distributor has timely information about demand, dependable supplier commitments, and the ability to replenish a customer’s stock before an interruption becomes expensive. The assistant makes those capabilities easier to combine. A competitor can copy the interface without obtaining the information or making the delivery.
That difference offers a possible basis for keeping some of the value. The customer is buying fewer stockouts, not an impressive conversation with a catalog. Whether the distributor can charge for that result still depends on alternatives and bargaining power. Useful does not automatically mean profitable.
David Teece makes this more specific in Profiting from Innovation in the Digital Economy (2016), section II.B, on complementary assets. He identifies bottleneck assets as places where profits can accumulate, while stressing that the bottleneck can change over time. That is a better starting point than assuming the company adopting the technology will keep the return. My application is that cheaper AI can shift the scarce complement without protecting the adopter from imitation.
There is evidence for separating productivity from strategic advantage. In their April 2023 working paper Generative AI at Work, Erik Brynjolfsson, Danielle Li, and Lindsey Raymond studied a staggered rollout across 5,179 customer-support agents. The authors describe their central productivity result as follows:
Access to the tool increases productivity, as measured by issues resolved per hour, by 14% on average
They also report a 34% improvement for novice and lower-skilled workers, with little effect on experienced, highly skilled workers. Those are results in that support setting, not a forecast for every workforce. And the reported productivity measure does not establish who retained the surplus after competition. A real improvement in production and a defensible margin remain separate claims.
The bottleneck sends an invoice
There is also a less flattering possibility. The distributor’s records are contradictory, exceptions live in someone’s head, and every proposed order waits for one senior buyer. Faster recommendations make the waiting room bigger.
Management might call this an adoption problem. It is an operating-design problem. The buyer is being asked to reconstruct the meaning of each recommendation because the system has produced an answer without producing the evidence needed to accept it.
Another agent can generate more answers. It cannot grant purchasing authority, settle an ambiguous supplier obligation, or decide which customer deserves scarce stock. Those decisions require rules, accountable owners, and occasionally an argument. None is included in the token price.
The tempting response is to declare workflow knowledge a permanent moat. That would be convenient, particularly for incumbents with untidy workflows. But knowledge becomes less protective if a supplier can package it into a standard product, employees can take it elsewhere, or customers can achieve the same outcome without the distributor. Complexity is sometimes accumulated expertise. Sometimes it is just accumulated complexity.
The useful question is which knowledge changes an outcome and how difficult it is for a rival to acquire an equivalent. Being hard to understand is not the same as being hard to compete with.
Separate the saving from the advantage
I would divide the investment case into three claims, each with its own evidence. The tool makes a task cheaper. The operating system turns that task into a better customer outcome. The business can retain some of the resulting value after competitors respond.
A demonstration may establish the first claim. A trial inside the real workflow can begin to establish the second. The third requires a view of customers, distribution, substitutes, and pricing. Passing one test does not waive the others.
The same distinction applies to generating your own tokens. Controlled infrastructure may lower cost or preserve a useful alternative to a supplier. Neither proves differentiation. If the hardware supports a capability competitors cannot readily reproduce, the strategic case is stronger. If it merely manufactures an increasingly available input, the investment has to survive on economics and continuity, without a moat painted on the casing.
Before approving the next AI project, run the competitor’s spreadsheet alongside your own. Assume comparable model access and cheaper implementation for everyone. Then ask what your business can still do better, what the customer will pay for, and which party can demand a larger share.
If the answer disappears when the technology becomes widely available, you have identified an efficiency project. That can be worth doing. It should be funded as one.