Michael Garland, a federal procurement expert, explains why agencies should take steps now to avoid “emotional and cognitive incumbency” of AI tools.
In August 2025, the entire federal executive branch gained enterprise access to OpenAI’s ChatGPT and Anthropic’s Claude for $1 per agency, and Google’s Gemini for 47 cents per agency, through a series of General Services Administration OneGov deals. Other AI companies offered similar effectively “free,” too-good-to-be-true “buy-in” deals.
Four months earlier, in April 2025, the Department of Veterans Affairs signed a contract for enterprise use of Microsoft software worth $4.65 billion, or $930 million a year. That is what a $1 deal might turn into after 30 years of accumulated incumbency. The $1 AI pricing may prove equally scandalous if it produces the same kind of lock-in Microsoft currently demonstrates. High pricing is the visible symptom. Vendor lock-in is the underlying disease.
These are early AI days. Think of this as the DOS computing infancy era of AI. Roughly 100,000 IBM PCs shipped in 1981; by the end of the decade some 60 million PCs and compatibles had shipped worldwide. But the speed of AI adoption makes that first computing wave look glacial by comparison. ChatGPT achieved 100 million users in its first two months. The technology works; it’s quickly infiltrating our lives, yet almost everything about what AI eventually evolves to remains unknown. Speculation is all we have. Nonetheless, the procurement decisions made today will shape or limit what is possible later.
This moment differs from the first compute wave in at least one important way. We can reflect upon the lock-in pattern that produced vendor lock because we lived through it. The new AI wave is a chance to apply what the first wave taught us, but only if we choose to apply it. Without careful planning, I fear we are walking into an AI vendor-lock that might make Microsoft’s entrenchment and resultant pricing look like a bad rehearsal. Microsoft is the natural comparison because it is, by an enormous margin, the largest software company in the world, with fiscal 2025 revenue of $281.7 billion, roughly five times its nearest enterprise software competitor Oracle. In the government space it rarely faces a genuine product-centric competition; it is almost always simply renewed, year after year.
Earn CPE credit: The latest webinar from the Billington CyberSecurity Cyber and AI Outlook Series will focus on the real-world risks facing AI deployments across the federal landscape. Register now!
The lock-in concern is not hypothetical. As Federal News Network reported in August 2025, Nicolas Chaillan, the former chief software officer for the Air Force and Space Force, filed protests at the Government Accountability Office against the OneGov deals, arguing they were “100% designed to get people locked in.” His point is worth underscoring. The $1 and 47-cent offerings cover access to the vendors’ user interfaces. They generally do not include application programming interfaces (APIs) and model access, which is what agencies need to build platform-agnostic automation across providers. The capability that would allow even minimal vendor-neutral deployment is the thing being held back from the giveaway. GAO dismissed the protests in December on jurisdictional grounds. The lock-in argument was never addressed.
Yet the OneGov AI deals are working as anticipated. GSA announced at an industry conference in May that 3.4 million government users have gained access to AI through OneGov offerings, across more than 120 agency orders. For scale, the 3.4 million figure somewhat exceeds the estimated size of the entire federal workforce.
Consider what the AI pricing trajectory might soon look like. In September 2025, the consulting firm Bain looked at the current global market for software, the current investment dollars flooding into AI and made the following stark prediction. To fund the compute infrastructure the dominant AI firms are already building, the industry will need to generate $2 trillion in annual revenue by 2030. That is roughly twice the size of the entire current global enterprise software market, which Gartner estimates at over $1 trillion a year.
If that revenue requirement is met proportionally across the global software-buying population, the federal government’s roughly $20 billion in annual commercial software spending would double.
Some of the current AI investments will fail, of course, which reduces the total revenue burden across the industry. But a major shake-out will not destroy the market. Consolidation typically works against competitive pressures and prices will likely increase even faster. The VA’s current $930 million annual Microsoft bill is probably an unpleasant harbinger of future AI expenditures.
The government has at least begun to recognize one element of the gap. Earlier this year, GSA proposed a new contract clause addressing AI procurement through a wide set of requirements including data ownership, data handling restrictions, disclosure obligations, American sourcing rules and the override of commercial terms of service. These are useful provisions. But even if this legal language survives, there is a deeper problem the clause does not address.
Owning the data is not the same as being able to use it. Raw text, chats and documents can move between platforms easily enough. But the artifacts that matter for serious AI deployment, the fine-tuned models, the embeddings that power agency-specific search indexes, the custom agents configured against a specific platform’s tooling, are tied to the vendor’s infrastructure. A fine-tuned deployment built on one company’s platform cannot run on another’s. Think of it as a bucket of car parts that fits only one chassis. The parts are yours; you own them. But they will not work in another car.
Sign up for our daily newsletter so you never miss a beat on all things federal
Interoperability is the obvious answer, but actual AI interoperability does not exist and there is no realistic timeline for when it might. Chaillan’s protest identifies a narrow version of this problem. The disallowed API access would at least give agencies some optionality to route prompts between AI vendors, assuming they had the in-house capacity to operate such a layer. But this is only the prompting level. The kind of interoperability that would let a fine-tuned deployment move from one vendor to another does not exist at all. Nobody has built that capability and it’s unclear that the knowledge currently exists.
Anyone who has tried to open a complex Word document in a competing suite, or sync Outlook with Google Calendar, or share an Excel file with macros across platforms knows this limitation. Email’s interoperability is a useful exception that illuminates the current problem. Simple Mail Transfer Protocol (SMTP) was established as an open standard in 1982, before any commercial email vendor reached the scale to prevent it. This is why your email message can find its way to every email platform: Gmail can talk to Outlook which can even talk to an antiquated AOL account. Meanwhile, cross-platform interoperability in office productivity took roughly two decades to achieve and remains uneven and far from perfect.
Horizontal AI interoperability, the kind that flows between competitors, currently seems to be off the table. The dominant vendors have every commercial reason to keep their platforms walled, no regulator is forcing the issue and the technical standards do not exist. What interoperability between AI platforms would actually mean is itself an open question. Nobody has defined what it would look like to move a yet-to-be-built AI implementation from one vendor to another. Designing for true portability now is like designing a freight elevator for a building with an unknown number of floors and unknown cargo burden. We cannot accurately specify interoperability contract requirements now because no one has been able to accurately articulate what that means.
When an agency tries to leave an integrated AI deployment, it will leave behind everything the system has been built to do. Call this cognitive incumbency. The product has something analogous to cognition, accumulated through use, that evaporates when the product is replaced. Its cognitive capacity is buried in the platform.
AI tools also invite something that functions like a personal relationship. Their sycophancy-tuned algorithms are designed to build user connection. When the renewal cliff arrives, will agencies be dealing with a procurement decision or a divorce? This is emotional incumbency, and, I predict, it will be a significant hidden source of vendor lock.
Buyers and agency technology leaders need to think about what protections might work. Three come to mind immediately.
Commercial buyers running serious deployments will find out first what AI actually costs. The federal government will find out next. What it does with that information will depend on whether agency leaders have positioned their organizations to act on it. The risk is that the government reacts slowly once the real price is known and then slow walks itself, once again, into another massive vendor-lock that cannot be mitigated for decades.
Michael Garland is a government procurement lawyer who has written extensively about vendor lock-in. His clients include IT companies and the federal government. He also provides litigation support and expert witness testimony in disputes involving IT and government contracts.
Read more: Commentary
Copyright © 2026 Federal News Network. All rights reserved. This website is not intended for users located within the European Economic Area.
Leave a Reply