Acquisition essay · Mission Delivery Notes
Your data outlives every contract.
AI is cutting the cost of building software every year. An agency that keeps control of its data, can modify its code, and keeps a working exit can take advantage of that. An agency that signs away those rights for a decade cannot. Don't pick a winner for the next ten years. Keep the ability to choose.
Nobody can pick the winner
Try to name the best AI model, the best coding agent, or the cheapest way to run either one twelve months from now. You can't. Neither can I, and neither can the people selling you a five-year commitment.
The price of AI keeps falling, fast and unevenly. Epoch AI found that the price of reaching a fixed level of model performance has fallen anywhere from 9 to 900 times per year depending on the task, and 40 times per year on PhD-level science questions. Stanford's 2025 AI Index found that inference cost at GPT-3.5's level dropped more than 280-fold in about two years. Those are inference prices, not the full cost of building software, and nobody promises the trend will hold. But the direction is not subtle.
When the cost of building software is falling that quickly, a long contract that locks in today's price, today's architecture, and today's vendor isn't a safe choice. It is a bet that the future will look like the present. That bet has never paid off in technology, and it is a worse bet now than at any time I can remember.
Picking a winner for the next 12 months is optimistic. Picking one for the next five years is foolish. Ten is a wish.
The data outlives the code
Mike Acton put the core idea plainly in 2014: "The purpose of all programs, and all parts of those programs, is to transform data from one form to another." And: "If you don't understand the data you don't understand the problem."
Martin Kleppmann, in Designing Data-Intensive Applications, names the consequence. Data written years ago is still sitting there in its original form after the application around it has been replaced. In his words, "data outlives code."
Every agency has lived this. The benefits system, the case management tool, and the grants portal have been rebuilt, re-platformed, and re-competed. The records underneath are the same records. The people, cases, payments, and decisions they describe are the mission. The application is just the current way of working with them.
So the most important thing an agency can control is not this year's application. It is the data, the definitions, and the rules that will still be there when this year's application is gone.
The code is the context
There is a second reason this matters now, and it is new.
AI is about to let government staff do far more with their own systems than they ever could. An analyst can ask an agent to build a report the vendor never prioritized. A program office can prototype a workflow change in an afternoon. A small internal team can fix a bug, add a check, or write a migration script that would have been a contract modification a few years ago.
To do that well, the AI needs context, and the best context for how an agency's work actually happens is the software that runs it. The code shows the validation rules, the edge cases, the state transitions, and the business logic that nobody wrote down anywhere else. An agent that can read the source can work with the system. An agent that can only see screens and a PDF of requirements is guessing.
If the government can't see or modify the code, its own staff can't point AI at their own systems. That is the real cost of a closed contract now. It isn't only the price of switching vendors later. It's the work you can't do yourself today.
Lock-in has a price, and GAO keeps measuring it
This isn't theoretical. GAO has been documenting what happens when the government doesn't control its data and code.
Weapon systems. In 2025, GAO found that selected programs, including the F-35, F/A-18, Littoral Combat Ship, Stryker, and Virginia-class submarine, "experienced vendor lock—or reliance on a single supplier—due to data rights shortfalls," which "can drive up costs and lengthen repair timeframes." GAO added that options to fix it "are limited once programs enter sustainment."
The F-35. GAO put the program's life-cycle cost at more than $1.7 trillion, about $1.3 trillion of it for operating and sustaining the aircraft, and found DOD had not "identified and obtained the technical data needed" while relying heavily on its contractor.
Legacy IT. In 2025, GAO reported that agencies spend about 80 percent of their IT budgets on operating and maintaining existing systems. The critical legacy systems it reviewed were 23 to 60 years old, and eight of eleven used outdated languages.
The Army said it bluntly in its 2018 intellectual property policy: negotiating IP rights late "may lock us into long-term, costly, sole-source arrangements."
Lock-in is not an accident that happens to a few programs. It is what you get by default when the government doesn't ask for the rights up front.
The rush to lock agencies in
Here is the blunt part.
Some contractors see exactly where this is going. AI is making it possible for agencies to do far more themselves, with smaller teams and less dependence on any single vendor. The firms whose business depends on that dependence know it, and some of them are racing to sign agencies into long, closed, all-in commitments before agencies realize how much they can now do on their own.
You'll recognize the pattern. A platform that owns the data model. A proprietary format with an export that technically exists but doesn't work. A "strategic partnership" with a five-year term and a ten-year roadmap. Source code that stays with the vendor. Pricing that rewards consolidating everything in one place.
None of that requires bad people. It only requires incentives. A long lock-in protects the vendor's revenue. It does nothing for the agency, and nothing for the public that pays for both.
The good news is that the government doesn't have to accept it.
The rules are already on your side
Federal policy has been moving in this direction for a decade, and the current rules are explicit.
OMB M-25-22 (2025) tells agencies to pay "careful attention to vendor sourcing, data portability, and long-term interoperability to avoid significant and costly dependencies on a single vendor." It lists protections that include "data and model portability, providing agencies with rights to code and models produced in performance of a contract, and transparency in licensing and pricing." It even warns against arrangements that consolidate spending with one vendor "through fixed contract lengths."
The SHARE IT Act (2024) requires agencies to "acquire and exercise rights sufficient to enable the governmentwide access to" custom-developed code, including code produced under contract.
OMB M-16-21 (2016), the Federal Source Code Policy, already required agencies to acquire rights to custom code and to "ensure delivery of the custom-developed code, documentation, and other associated materials... throughout the development process," explicitly to reduce vendor lock-in.
FAR 52.227-14 gives the government unlimited rights in data, including software, first produced under the contract, unless the contract says otherwise.
DFARS 252.227-7014 defines government purpose rights: the government can use, modify, and release the software within the government and authorize others to do so for government purposes, but not for commercial use.
That last one describes what I think is the right default for most custom work. The contractor can keep a non-exclusive license and its reusable background IP. The government gets the right to modify the code for its own needs and let other contractors work on it. Nobody gets to resell what the public paid for. Everybody keeps the ability to switch.
Analysts give opinions, not guarantees
Agencies often lean on industry analysts when choosing a platform. That's reasonable, as long as everyone remembers what an analyst report is.
Vendors have sued Gartner over its rankings. ZL Technologies sued over its Magic Quadrant placement. NetScout alleged pay-to-play. Both lost. A federal appeals court affirmed dismissal of ZL's claims under a First Amendment test, and the Connecticut Supreme Court held that Gartner's statements about NetScout "were expressions of opinion."
That is the right legal result. It is also a useful reminder. An analyst ranking is protected opinion, not a warranty. It isn't a basis for betting a decade of an agency's mission on one provider.
And on lock-in, Gartner's current opinion is worth hearing. In September 2026, it predicted that by 2028, 70 percent of enterprises will abandon agentic AI built by vendor forward-deployed engineers, "trapped by soaring costs and unable to evolve it on their own." Its advice includes writing deliverables and knowledge transfer into the contract, and to "execute the exit plan established at the start." Another Gartner analyst put the model risk simply: "If you are relying on a single provider with a single model, there's risk there."
Intelligrit's engineers are forward deployed, so that warning applies to firms like ours too. The answer isn't to avoid embedded engineers. It is to make sure that whatever they build, the agency can keep running, change, and hand to someone else.
What to ask for
Think of it in three levels.
The minimum: you can leave.
Your data is exportable at any time, in documented, open formats, with the schema and definitions, at no extra charge.
There is a written exit plan from day one, including how data and knowledge transfer at closeout, and someone has actually tested it.
No single contract term outlasts your ability to predict the technology. Use short base periods with options.
Better: you control the data.
The data lives in an environment the agency controls, and the contractor works there with authorized access.
Business rules, data definitions, and integration contracts are written down as deliverables, not left as tribal knowledge.
Models, prompts, and agent configurations built for you are portable across AI providers.
Ideal: you can change the code.
Custom code is delivered continuously, with the build pipeline and documentation, not at the end.
The government has the right to modify it for its own needs and let other contractors work on it, such as government purpose rights or broader, with the contractor keeping a non-exclusive license and its background IP.
Agency staff can point their own AI tools at the source to understand, extend, and fix the system.
And work with contractors who welcome all of this. A contractor confident in its value doesn't need to trap you. At Intelligrit, the source, data, and transition rights for each engagement are set in the agreement and the applicable clauses, and we show our work as we go: working software, tests, decisions, blockers, and named owners. That's what to expect from anyone you hire.
Control your own future
Governments are already acting on this. In 2025, the German state of Schleswig-Holstein moved more than 40,000 accounts and over 100 million emails and calendar entries off Microsoft Exchange and Outlook onto open-source alternatives, and acknowledged some problems along the way. The International Criminal Court announced it was replacing Microsoft Office with an open-source suite after concluding it had become too dependent on US technology. Those are not easy moves. They are what it looks like when an organization decides control is worth the effort.
US agencies have a better opportunity than that. They don't need to unwind a decade of dependency all at once. They need to stop signing new ones, starting with the next contract.
AI is giving government a real chance to stop being trapped by its contractors. Keep the data. Get the right to change the code. Keep the exit open. Work with people who are glad you have those rights. Then let the best option win, every year.
Sources
Ben Cottier and others, Epoch AI, "LLM inference prices have fallen rapidly but unequally across tasks" (March 2025).
Stanford HAI, 2025 AI Index Report (2025).
Mike Acton, "Data-Oriented Design and C++", CppCon (2014).
Martin Kleppmann, Designing Data-Intensive Applications (O'Reilly, 2017), chapter 4, "data outlives code".
GAO, "Weapon System Sustainment: DOD Can Improve Planning and Management of Data Rights", GAO-25-107468 (September 2025).
GAO, "F-35 Aircraft: DOD and the Military Services Need to Reassess the Future Sustainment Strategy", GAO-23-105341 (September 2023).
GAO, "Information Technology: Agencies Need to Plan for Modernizing Critical Decades-Old Legacy Systems", GAO-25-107795 (July 2025).
U.S. Army, "Army enacts new policy on intellectual property" (January 2019), on Army Directive 2018-26.
OMB, "M-25-22: Driving Efficient Acquisition of Artificial Intelligence in Government" (April 2025).
SHARE IT Act, Public Law 118-187 (December 2024).
OMB, "M-16-21: Federal Source Code Policy" (August 2016).
DFARS 252.227-7014, Rights in Other Than Commercial Computer Software.
U.S. Court of Appeals for the Ninth Circuit, ZL Technologies, Inc. v. Gartner Group, Inc., No. 10-16061 (May 2011).
Connecticut Supreme Court, NetScout Systems, Inc. v. Gartner, Inc., 334 Conn. 396 (2020); see also The Register (January 2020).
Gartner, "Gartner Predicts 70% of Enterprises Will Abandon Agentic AI Built by Vendor Forward-Deployed Engineering by 2028" (September 29, 2026), as reprinted by IT Voice.
Agam Shah, "Too good to be true? Avoid free AI token offers — or risk vendor lock-in", Computerworld (June 2026).
Liam Proven, "Schleswig-Holstein waves auf Wiedersehen to Microsoft stack", The Register (October 2025).
"The International Criminal Court is ditching Microsoft software for an open source alternative", TechRadar (October 2025).