A tool you built your work on can disappear overnight, and almost everything written about it is written for developers: migration endpoints, deprecation schedules, version notes. The founder question is plainer and more frightening. What happens to my work, my records, my running operation, the day the thing I built on goes away. This is the evidence base for that question, dated because these facts age fast, and then the answer the rest of this pillar carries.

When a tool is deprecated, your work is only as safe as the exit you built before it happened, and most founders build none. Start with the base rate, because it is the part people underestimate. US startup shutdowns rose to 966 in 2024 from 769 the year before, a 25.6% jump, and enterprise software accounted for roughly a third of them. Tools do not just occasionally fail. They fail at a rising rate, and most die young.

The concrete version is the one that stays with you. Bench, an accounting and tax tool that served more than 35,000 business customers, went dark the same day in late December 2024 and locked customers out at the very start of tax season, telling them to file extensions while they scrambled for their own records. That is what “deprecated” means to a founder. Not a changelog entry. A door that no longer opens, with your data behind it.

966 US startups shut down in 2024, up 25.6% from 769 the year before, with enterprise software roughly a third of them.
TechCrunch, on Carta data, 2025

A model retirement is routine, so the answer is to make it a part swap rather than a rebuild. Vendors sunset models on a schedule, and a widely used one being pulled is now a normal event in the cycle, not a freak occurrence. The mistake is treating the specific model as the foundation, so that retiring it means re-engineering everything built on top.

The durable move is to keep the model at arm’s length, where the thing you swap is the engine and the thing you keep is the business setup around it. The how of swapping cleanly, the open, model-agnostic architecture that makes it routine, is its own subject, and we treat it in how to avoid vendor lock-in. The point here is the outcome: a model retirement should cost you an afternoon of confirming the swap, not a quarter of rebuilding.

It breaks if the logic lived in the tool. It survives if the logic lived in your foundation. The difference is visible in how the well-funded names died. Pedigree is not insurance.

Olive AI, once valued at four billion dollars, wound down in October 2023. Artifact, the AI news app from the founders of Instagram, was killed less than a year after launch when the math did not work. Adept, an enterprise AI startup that had raised 350 million dollars at roughly a billion-dollar valuation, lost its founders and most senior staff to a larger company under a licensing arrangement, leaving around twenty of a hundred employees behind. The acquihire-then-sunset pattern hits enterprise tools, not just consumer apps, which means a strong logo on the invoice tells you nothing about whether the tool will still be there in eighteen months. If your business logic is encoded inside one of those tools, its death is your reset. If the logic lives in your own foundation, the tool’s death is a supplier change.

Because the model stopped covering for you, not because it got worse. A tool does not have to shut down to break your work. A model can update and quietly change how it behaves. When a frontier model updated in April 2026, it stopped silently compensating for incomplete instructions and began reading prompts more literally, so setups that had run fine for months suddenly produced different results. The model did not regress and it did not learn anything new. It stopped filling in the gaps your prompts had been leaning on.

To a founder who built a process on the old behavior, the distinction is academic. The work changed under you, and it took real time to find out why nothing was broken and everything was different. This silent drift is the version of the kill that does not make the news, and it is the strongest argument for not hard-wiring your operation to one model’s exact quirks. The deeper, dated walk-through of that specific behavior shift is in the operator’s account of it.

Build so the model and the tool are replaceable components, and the foundation is the asset. That is the whole answer, and it is the line the rest of this pillar carries. The model is a part. The foundation, your business context, your workflows, the record of what got done, is the thing that must not reset when a vendor pulls the plug. Parts get replaced. The asset does not reset.

In practice that means the platform you run on should treat models as disposable on purpose, swapping and absorbing new ones underneath you without touching the setup on top. Works is built that way: the user’s setup is durable and the capability underneath is replaceable, so a model retired upstream becomes a part swap, not a migration, and a tool that dies is a supplier you change rather than an operation you rebuild. In a market where a tool’s survival is closer to a coin flip than a guarantee, that vendor-agnostic stance is not a feature. It is survival insurance.

This is not fear-mongering, it is a base rate. Shutdowns rose a quarter in a single year and even the billion-dollar names got hollowed out. The point is not to panic about which tool dies next. It is to build so that when one does, your business does not.

Survive the next AI sunset. Get early access. Or DM us for the Compounding brief, a short written breakdown of how to build so a shutdown does not reset you.

What happens to my work when an AI tool is shut down?

The outcome depends on where your business logic and records live. Logic and context held inside the tool disappear when the tool does, which is what Bench’s 35,000-plus customers discovered in December 2024: locked out at the start of tax season with records inaccessible and no export ready. Logic held in a foundation you control survives the shutdown as a supplier change with no data loss. US startup shutdowns rose 25.6 percent in a single year, making this a planning requirement rather than a remote scenario. More in how to build so a shutdown does not reset you.

My AI model is being retired. What should I do?

A model retirement is a routine part of the vendor cycle, and the setup determines whether it costs an afternoon or a quarter. When the model is held at arm’s length in a replaceable slot, swapping in a successor is a confirmation step. When business logic was written around one model’s exact quirks, the retirement triggers a re-engineering project. The architecture that makes it routine is covered in avoiding vendor lock-in.

Are tools from big, well-funded vendors safe from shutdown?

No. Pedigree is not insurance. A four-billion-dollar valuation did not save Olive AI, the founders of Instagram killed Artifact inside a year, and a 350-million-dollar enterprise startup was hollowed out by an acquihire. The base rate of product death is high even at the largest companies. Plan for it rather than betting against it. The wider evidence is in the graveyard is the norm.

Will my automations break when the model behind them changes?

They can, even without a shutdown. A model update can change behavior, as one did in April 2026 when it stopped compensating for loose instructions, and setups wired to the old behavior shifted. The fix is to not hard-wire your operation to one model’s exact quirks, which is the maintenance and switching cost we cover in the maintenance tax and the switching tax.

How do I build so a vendor shutdown does not reset me?

Separate what can die from what must survive. The model and the tool are both replaceable parts; the business context, workflows, and records are the asset that must not reset when a vendor pulls the plug. A platform that treats models as disposable by design, swapping and absorbing them underneath your setup without touching what sits on top, converts any shutdown from a migration into a supplier change. That model-agnostic architecture is detailed in open architecture and vendor lock-in.

1. TechCrunch (Mary Ann Azevedo), 2025 will likely be another brutal year of failed startups, data suggests
2. GeekWire (Taylor Soper), Amazon hires founders from well-funded enterprise AI startup Adept, 2024
3. TechCrunch (Charles Rollet), Bench shuts down, leaving thousands of businesses without access to accounting and tax docs, 2024
4. Healthcare Dive (Emily Olsen), Health AI startup Olive to shut down, 2023
5. Engadget (Karissa Bell), Instagram’s founders are shutting down Artifact, their year-old news app, 2024
   
   

Keep reading:

• The maintenance tax a kill triggers
• The switching tax a kill forces
• Vendor lock-in and the model-agnostic how
• Compounding AI, the whole picture