Instead of paying an AI company simply to install technology, the Pentagon is testing a model where the vendor gets paid from the costs its automation actually eliminates.
WHAT’S HAPPENING
The Pentagon’s Chief Digital and Artificial Intelligence Office has entered a one-year agreement with Red Cell Partners worth up to $100 million to test a shared-savings contracting model.
Under the arrangement, Red Cell portfolio companies can identify government workflows that could be improved with AI and automation.
But there is an unusual condition:
Red Cell covers the upfront cost.
The company is then paid a percentage of the savings its technology actually generates.
If the promised savings do not materialize, Red Cell gets nothing.
The agreement is structured as an Other Transaction Agreement, a contracting mechanism that gives the Pentagon more flexibility than traditional federal procurement.
WHY IT MATTERS
Most technology contracts begin with a familiar question:
How much does the technology cost?
This model begins with another:
How much existing cost can the technology make disappear?
That changes the economics of AI adoption.
Instead of government agencies paying for licenses, consultants, implementation hours or promises of future productivity, an AI vendor can potentially make money only after proving that a workflow became cheaper.
It also creates a very different incentive for the vendor.
The faster AI removes unnecessary work, duplicated processes or administrative expense, the more valuable the AI becomes.
That turns automation from a feature into the financial product itself.
WHO BENEFITS
Government agencies could gain access to AI modernization without absorbing all of the upfront financial risk.
AI companies gain a new route into enormous organizations where traditional procurement can be slow and difficult.
Taxpayers could benefit if contractors are rewarded for reducing spending rather than expanding the size and duration of government projects.
And companies capable of measuring real productivity improvements could gain an advantage over vendors selling AI primarily through demonstrations and promises.
WHO LOSES
Traditional contractors that earn revenue through large teams, long implementation cycles and billable hours could face a fundamentally different incentive structure.
Some administrative work could also become a direct target for automation because every eliminated cost potentially creates revenue for the company doing the automating.
There is another challenge:
Savings must be measured accurately.
A system that eliminates an expense but creates another one somewhere else has not necessarily produced a true saving.
WHAT HAPPENS NEXT
The first targets are expected to be ordinary government business processes rather than battlefield systems — the repetitive administrative work where AI agents may be able to automate tasks, combine disconnected information or help employees work faster.
But the larger experiment is economic.
If this model succeeds, other government agencies — and eventually private companies — could ask AI vendors to make the same offer:
Don’t tell us how productive your AI is. Prove what it saves us.
That would create a very different AI economy.
The product would no longer just be artificial intelligence.
The product would be the cost that disappears after you deploy it.
