Most AI tools a small business pays for sit on top of something it didn’t choose and can’t control: a model, an API, a connector, a pricing tier. When that layer changes, the tool changes with it. That is platform dependency risk. A custom build doesn’t remove it. It moves it somewhere you can see it and plan for it.
We build custom AI tools through our AI integration service, so weigh that. Every fact below about Anthropic, OpenAI, Meta and Zapier was checked against their own documentation and announcements in October 2026.
Short answer: Every platform you build on will change: models get retired and endpoints removed. What matters is how much notice you get and who does the work. Anthropic commits to at least 60 days’ notice before retiring a publicly released Claude model, and OpenAI to at least six months for its generally available models. A subscription inherits every change upstream of it, on someone else’s schedule. A custom build on documented, versioned APIs gets those notices directly, but someone on your side has to act on them. Build when the workflow is specific and stable and you can fund maintenance. Otherwise subscribe, and audit what each tool depends on.
What the Vendors’ Own Notices Say
As of October 2026, from Anthropic, OpenAI, Meta and Zapier:
| Change | What the vendor documents |
|---|---|
| Claude model retirement | At least 60 days’ notice for publicly released models; requests to a retired model fail |
| Claude model IDs | Each ID is a pinned snapshot; updates ship under a new ID |
| Claude Sonnet 4.5 | Deprecated September 30, 2026; retires November 30, 2026 |
| OpenAI GA models | At least 6 months’ notice before retirement |
| OpenAI preview models | Can be retired with much shorter notice, such as 2 weeks |
| OpenAI Assistants API | Removed August 26, 2026, a year after notice |
| Instagram Basic Display API | Announced September 4, 2024; requests fail from December 4, 2024 |
| Zapier Instagram app | Zaps using it stopped working after December 4, 2024 |
Sources: Claude model deprecations, Claude model IDs and versioning, OpenAI deprecations, Meta on the Instagram Basic Display API, Zapier on its Instagram app.
What Platform Dependency Actually Looks Like
Most SMBs think of vendor lock-in as a pricing problem: switching costs, data portability. Those are real. The operational risk is different. The platform changes what it offers, and the part of your process that relied on it stops.
The dependency is often several layers deep. You subscribe to a tool. The tool calls another company’s model or API. Your contract is only with the first layer.
The API Cliff, When Platforms Change the Rules
Meta’s Instagram Basic Display API is a documented case. On September 4, 2024, Meta announced that the API would no longer be available from December 4, 2024, and that all requests to it would fail from that day. Partners had 90 days to adjust (source).
The knock-on effect showed up one layer down. Zapier’s Instagram integration ran on that API, and Zapier told users that Zaps built with it would no longer work after December 4, 2024. The replacement only works with Business or Creator accounts (source). A business that automated posting through one of those Zaps never signed anything with Meta. It still had to rebuild.
That is a normal platform lifecycle, announced and dated. The question is whether anyone on your side was watching.
The Shutdown Pattern Has a Second Version
A platform doesn’t have to close to affect you. It can replace one product with another. OpenAI notified developers in August 2025 that its Assistants API would be removed on August 26, 2026, with the Responses and Conversations APIs as the replacement (source). OpenAI kept operating. The product the integrations were built on didn’t.
Model retirement works the same way. Anthropic lists Claude Sonnet 4.5 as deprecated on September 30, 2026, retiring on November 30, 2026, with Claude Sonnet 5.5 as the recommended replacement (source). Any tool calling that model has to move, and the new model needs testing on your real inputs first.
The Notice Periods SMBs Should Know
The numbers worth planning around are the notice periods vendors publish. They tell you how long you have between the email and the failure.
- Claude models: at least 60 days, by email to customers with active deployments and in the docs (source).
- OpenAI models: at least 6 months for generally available models, 3 months for specialized variants, and much shorter notice, such as 2 weeks, for previews. Safety or compliance concerns can shorten these (source).
- Claude on partner clouds: Amazon Bedrock and Google Cloud set their own retirement schedules, so dates can differ (source).
What Those Numbers Mean for Point AI Tools
A point AI tool, one subscription per job, usually runs on someone else’s model. When that model is retired, the tool vendor gets the notice, not you. You rely on them to migrate in time and to check that outputs still look right.
The most exposed tools are thin layers over a single model, with no data or workflow of their own. If a general model can now do the job directly, the product has little reason to last.
Preview features carry the most risk. OpenAI doesn’t recommend preview models for business-critical production workloads unless you can migrate on short notice (source). If your tool is built on one, that exposure is yours too.
The Subscription Stack Problem
Tool sprawl compounds the risk. Most SMBs buy several AI tools, each solving a slice of the same workflow. Every subscription is a dependency with its own upstream vendors and notice periods, and usually nobody tracks them in one place.
The Custom Build Trap, It’s Not a Safe Default Either
The case for custom is real, but its risks are often understated.
Why In-House AI Initiatives Stall
We don’t have a reliable failure rate to give you. The pattern behind stalled builds is easier to describe than to count. The model usually works. The scoping, the infrastructure and the knowledge needed to keep it working get underestimated. A working demo is not a maintained system.
Most SMBs don’t have an engineering team on call. A custom build that nobody maintains degrades. Models are retired on the vendor’s schedule. Libraries and CRM fields change. It stops working quietly, unless someone built the logging that tells you.
The Troubleshooting Window After Launch
Custom builds have a cost that rarely appears in the first quote: the debugging and iteration after launch. Real data brings edge cases that test data didn’t show. How long that phase lasts depends on the workflow and the data, so plan for it rather than pick a number.
The vendor side moves too. Anthropic notes that the serving infrastructure around a model, such as request routing, safety classifiers and sampling logic, can change and occasionally cause minor differences in behavior on the same model ID (source). A custom tool needs a set of test inputs that catches that.
The real choice isn’t custom vs off-the-shelf. It’s ownership without support vs dependency without control.
What Ownership Actually Means in Practice
Ownership has specific technical and legal parts.
Documented APIs vs Unofficial Side-Channels
Build only on what the vendor documents. A documented, versioned API comes with a deprecation policy, notices and a migration path. On the Claude API, each model ID is a pinned snapshot: Anthropic doesn’t update the weights or configuration of an existing ID, and a new version ships under a new ID (source). Your tool calls a fixed model until you decide to move it.
Undocumented behavior has none of that. Scraped interface data or response fields that aren’t in the reference can disappear without notice, because nothing promised them. The rule applies at every layer: the model, the integrations (official OAuth flows and published endpoints) and the data (storage your business controls).
Defined Inputs, Defined Outputs, Full Client Control
A tool built for ownership has a clear spec: this input goes in, this output comes out, and you hold the code, the prompts and the keys. The vendors you still depend on are named, with their notice periods.
It’s the difference between a custom WordPress development engagement where you own the codebase, and a website builder subscription where the platform holds the files. Both give you a working site. The ownership is different.
When we build a custom AI tool, the aim is that you hold the code, the documentation and the architecture. You still depend on a model vendor, and models will be retired. But the dependency is known, dated and listed, not hidden two layers down.
How to Audit Your Current AI Tool Stack for Obsolescence Risk
Before deciding to build or buy, map what you already depend on.
Five Questions to Ask Before You Build On Any Platform
- Is this feature in a published, versioned API? If you can’t find the API docs, you’re building on informal behavior.
- Does the vendor publish a deprecation policy? Look for minimum notice periods, like the Anthropic and OpenAI pages above.
- Is it a preview or beta feature? Those can go with much shorter notice.
- Who owns what you put in and get out? Read the terms for output ownership, training on your data and retention.
- Can you export everything cleanly? A full, structured export another system could load, not an unusable file.
Red Flags in Your Existing Tool Stack
- Tools that have moved features from free to paid tiers over time.
- Tools recently acquired, whose future depends on the new owner’s plans.
- AI tools that are a thin layer over one model, with no data of their own.
- Automations built on a connector to someone else’s API, like the Zapier case above.
- Any workflow where one tool’s outage stops your team working.
Map each tool against the five questions and note which dependencies have no published deprecation policy. That’s where your risk sits.
Related Posts
- AI vendor lock-in risks: what to check before you sign
- AI model updates and your integration: what happens when the model changes
- Build vs subscribe: when owning your AI tooling pays
- Custom AI tool maintenance: what ongoing ownership requires
Frequently Asked Questions
What is platform dependency risk and why does it matter for SMBs?
It’s the exposure you take on when a process relies on a third-party platform continuing to offer a feature, pricing tier or API as it is today. SMBs rarely have engineering time to migrate quickly. When a dependency goes, the cost is the disrupted workflow, the rushed migration and the staff time to rebuild.
Are there real examples of businesses losing critical tools due to platform changes?
Yes, and the vendors documented them. Meta shut down the Instagram Basic Display API on December 4, 2024, 90 days after announcing it, and Zapier’s Instagram app stopped working the same day. OpenAI removed its Assistants API on August 26, 2026, a year after notifying developers. These are normal lifecycle events, not exceptions.
What’s the actual cost difference between custom AI builds and SaaS subscriptions over three years?
It depends on your stack, so we don’t quote a standard figure. For subscriptions: seats times price for 36 months, plus any price changes the terms allow. For a custom build: the build, model usage at the vendor’s published prices, hosting and maintenance, including at least one model migration. Anthropic’s current models list retirement dates of “not sooner than” September or October 2027, so a three-year plan should assume a move (source).
How do I know if a custom build is the right call for my business?
Three conditions point toward custom: no off-the-shelf tool fits your workflow without workarounds; your data can’t safely or legally go to a third-party platform; or you pay for several tools that solve overlapping problems. For commodity work like email summaries or basic scheduling, a subscription is fine.
What should a custom AI build include to protect against future obsolescence?
Documented, versioned APIs at every layer. You hold the code, prompts, keys and documentation. The model is pinned by its exact ID and treated as a replaceable part, so a retirement means swapping and retesting one component, not rebuilding the system. Keep real test inputs with known-good outputs, and logging that shows failures.
What’s the difference between building on a documented API versus an unofficial feature?
A documented API has versions, a changelog and a deprecation policy, so you get notice: at least 60 days for a publicly released Claude model, at least six months for an OpenAI generally available model. An unofficial feature is behavior the platform never committed to, such as undocumented response fields or scraped interface data. When it changes, there’s no notice, because there was never a promise.