The email will be short, and it will use the word “sunset.” Your team will spend the afternoon scrambling to figure out what actually runs on the retired model, and by evening, someone will say the sentence you’ll remember forever: we should have gotten the export terms in writing.
You’ll ask, reasonably, why nobody did.
The “Plumbing Meeting” Trap
Fourteen months earlier, there was a meeting you skipped. It was the “plumbing meeting”—integration architecture, auth, data flows, the technical annex. Your engineers and the vendor’s engineers were there. The invitation promised “technical alignment,” the kind of session executives avoid because it’s where the nerds handle the pipes.
Four items were on the agenda. Item three was the exit language: who keeps your fine-tunes and embeddings when the relationship ends? A vendor representative casually mentioned that if American companies slow down, “someone else wins.” Nobody argued; it felt true, and none of them felt authorized to make the call. The item was dropped. It never reached your desk, because nothing that gets deleted ever does. You only saw a green checkmark and a start date.
Eleven seconds of silence in a technical meeting cost you the entire quarter.
The Illusion of Software Ownership
You didn’t buy software. Software you install; it sits on your machines and functions even when the vendor stops returning calls. You bought access to a mind living on someone else’s servers in a building you’ll never visit. That mind is a statistical object fed on a diet of text and images no one has fully audited.
Your team spent a year teaching this model your business. They built prompts, converted contracts into embeddings, and fine-tuned the model against your institutional memory. This is arguably the most valuable asset your company created last year, yet most executives have no idea where it is stored—or who has the right to walk out the door with it.
The Economics of “Hard to Leave”
The four largest tech companies have committed roughly $725 billion in capital expenditure this year—a 77% surge from 2025. They are spending at the scale of national infrastructure programs, yet a study by DigitalRoute found that only 8% of organizations are confident they understand the true cost of their AI features. When you sell a product whose unit economics are a mystery, the oldest trick in the book is to make leaving as difficult as possible.
Consider the “SaaSpocalypse.” Between January and April, $2 trillion evaporated from software stocks. Atlassian, Salesforce, and Workday took massive hits. But when Atlassian reported earnings on August 6, their seats were expanding again, credited to AI adoption. The pricing models underlying thirty years of enterprise software are now unstable in both directions.
Data Sovereignty is Not a Luxury
Aaron Gibson, CEO of the analytics platform Hurree, faced a choice when a third-party tool couldn’t handle his company’s scale. Instead of paying for a broken dependency, he spent six weeks building his own. “Six weeks of disruption was nothing compared to being permanently dependent on infrastructure we didn’t control,” he says.
This is the real argument for data sovereignty. It isn’t about lofty principles; it’s about what happens when the thing you depend on breaks, disappears, or changes the rules. When a vendor deprecates a tool, they don’t just kill a feature—they kill the “posture” you spent months teaching the machine to adopt.
The “Context” Moat
Atlassian CEO Mike Cannon-Brookes argues that in the AI era, “context is the edge”—it’s hard to build and cannot be hired. He believes that twenty-five years of connecting teams creates a moat that competitors can’t touch. But there is a dangerous distinction here: the context you build together is an asset; context you cannot export is a hostage. If you can’t walk away from your own memory because someone else is keeping it for you, you aren’t a customer—you’re a prisoner.
Lessons from the Factory Floor
In 1911, Frederick Winslow Taylor sought to extract traditional knowledge from workmen and reduce it to “rules, laws, and formulae.” He framed it as a favor to the workers, but the result was that the worker became replaceable because their skill was no longer in their hands—it was in the building, and the building had an owner. Today, your institutional context is being moved into a data center you’ve never seen, governed by a terms-of-service agreement you clicked through at 2:00 PM on a Tuesday.
The Path Forward: Transparency and Disclosure
Medicine offers a rare, successful precedent. When surgeons realized their primary anatomical atlas, the Pernkopf-Atlas, was built on the bodies of Nazi victims, they didn’t just burn the books. They developed the “Vienna Protocol.” They used the atlas only under strict conditions: formal disclosure of its origins, bioethicist review, and a memorial to the victims. They made the information available to the user, and they let the user decide.
The AI industry needs a similar “model card” approach. If a vendor wants the power of a landlord over your access, they must accept the obligations of one regarding their inputs. This doesn’t require a philosopher; it requires explicit guarantees: indemnity for model outputs, notice before deprecation, and a promise that your institutional knowledge never enters the vendor’s training set.
As Gibson puts it, “Trust is the only moat left. If your business model only works because customers cannot leave, it is not a moat. It is captivity with nicer branding.”
The solution isn’t to stop progress; it’s to ensure someone senior is in the room during the plumbing meeting, with a standing rule that the technical annex is not where your company’s future is traded away. One week of extra development is a small price to pay to ensure you still own your own brain.
