Everyone has seen a list of dead AI tools. What nobody publishes is the denominator.
A count of shutdowns tells you nothing on its own. Two hundred dead tools out of a tracked universe of two thousand is a stable market. Two hundred out of four hundred is a bloodbath. The number that matters is the rate, and the rate requires you to have been watching the tools before they died.
We have catalogued 2,653 AI tools since 2023, with an editor reviewing and dating every entry. 232 of them have since shut down or been acquired. That is a mortality rate of 8.7%, and because we only count tools we actually tracked while they were alive, it is a floor rather than an estimate.
Here is what three years of that data looks like.
Failure rate is not evenly distributed
The catalogue-wide rate is about 9%. The spread underneath it is 8x.
| Category | Failure rate | Dead / tracked |
|---|---|---|
| Automotive | 16.1% | 5 / 31 |
| NSFW | 12.1% | 4 / 33 |
| E-Commerce | 10.6% | 5 / 47 |
| Coding Assistants | 10.0% | 4 / 40 |
| Customer Support | 8.8% | 12 / 137 |
| Developer Tools | 7.9% | 25 / 318 |
| AI Infrastructure | 6.0% | 18 / 301 |
| AI Agents | 3.8% | 11 / 288 |
| Vector DBs and RAG | 2.9% | 1 / 34 |
| Security and Governance | 2.1% | 3 / 144 |
If you are picking a tool to build on, the category it sits in predicts its survival better than the market-wide average does. A coding assistant is roughly five times likelier to be gone than a security and governance tool.
Two honest caveats. Small categories post loud rates off a handful of deaths, which is why we exclude anything under 12 tracked tools and print the sample beside every rate. And a low rate in a young category can mean nothing has had time to die yet.
The most common way an AI tool dies is silence
Of the 167 retirements where we recorded a cause:
| Cause | Count | Share |
|---|---|---|
| Acquired and sunset | 60 | 35.9% |
| Shut down | 56 | 33.5% |
| Domain lapsed | 50 | 29.9% |
| Abandoned | 1 | 0.6% |
Almost 30% of dead tools never announced anything. No blog post, no email, no sunset notice. The domain simply stopped resolving. If you have a dependency on a small AI vendor, an expired certificate is a more reliable early warning than their changelog.
Acquisition is not the same as death
125 of the 232 retirements were acquisitions, and 65 of those tools are still shipping under their own brand — 52% of everything acquired. Lists that fold acquisitions into a "dead" bucket roughly double the apparent carnage.
NVIDIA has absorbed 8 tools from our catalogue, OpenAI 7, then Cursor (Anysphere), Google, Meta and Salesforce at 3 each.
Take the data
The dataset is open under CC BY 4.0, in formats meant for machines rather than screenshots:
-
https://tooldirectory.ai/feed/ai-tool-mortality.md— the headline figures as markdown -
https://tooldirectory.ai/feed/graveyard.json— the full record, one object per retired tool
Methodology, per-year cohorts and the full 34-category table: AI tool failure rate by category.
If you use it, we would rather you cite the rate than the count. The count is the part that is easy to inflate.
Top comments (0)