I Timed 20 AI Shutdowns. The Median Warning Was 15 Days. Not One Contract Promised a Number.

TL;DR: I timed 20 AI tools that shut down or were acquihired between January and August 2026. Eleven had a notice window I could compute from public evidence, and the median was 15 days between the public notice and the day the product stopped working. The range ran from 0 days (three products died with no announcement at all) to 64. The contractual answer is worse: of the seven Terms of Service documents I could still read on 2026-08-15, not one names a number of days of notice before the vendor may discontinue the service, and eight of the twelve independent vendors in my sample no longer serve a Terms of Service at any working URL.

Clockwise is the clean case. Salesforce acquihired the team; The Register reported it on 20 March 2026; and the entire getclockwise.com domain now serves one goodbye letter that reads, "Our product will no longer be available starting on March 27, 2026" and "Your data will be deleted shortly after you lose access to the product." Eight days, then deletion. Uber, Netflix and Atlassian were customers.

So I went and timed nineteen more.

What is the actual notice window?

I took 20 tools that shut down or were acquihired January to August 2026, and for each one I tried to compute days between the public notice and the day access ended. Eleven of twenty produced a number.

Tool	Notice	Access ends	Days
NextSlide (OpenAI)	none	already gone	0
Google Interview Warmup	none	April 2026	0
Zeda.io	none	July 2026	0
Hiro (OpenAI)	Apr 13	Apr 20	7
Clockwise (Salesforce)	Mar 20	Mar 27	8
Yupp	Mar 31	Apr 15	15
Agency (Klaviyo)	Aug 5	Aug 31	26
GitHub Models	Jul 1	Jul 30	29
ChatGPT Atlas	Jul 9	Aug 9	31
Claude Opus 4.1	Jun 5	Aug 5	61
Google Whisk	Feb 25	Apr 30	64

Median 15 days. Seven of those eleven notice dates come from a primary artifact I fetched (the vendor's page, an official changelog, a docs table). Four (Clockwise, Yupp, Atlas, Whisk) I dated from press coverage, because the vendor's own page does not say. Across only the seven primary-dated ones, the median drops to 7 days.

Deletion runs behind loss of access, sometimes generously. Hiro stopped working on April 20 and kept the export path open until May 13, a 30-day tail with a real self-service export at app.hirofinance.com/settings/export. Youper's notice page ends service September 30 and starts deleting October 1. Agency's homepage names August 5 and August 31 and then says only that it "will be reaching out directly about what it means for your account and your data": no export procedure, no format, no retention commitment, no word on prepaid refunds.

Why can't you compute this for the other nine?

Because the vendor does not date its own notice. I fetched nine independent vendors' wind-down pages on 2026-08-15 and extracted every date string. Only two, Hiro and Agency, state when the notice went up. Clockwise, Delibr and Youper give you the end date and never the start date. Mindy, NextSlide, Zeda and Yupp carry no date at all: nextslide.ai is 29 characters long and reads, in full, "NextSlide is joining OpenAI."

That is the mechanism behind the whole problem. The notice window is the one number a buyer needs, and it is the one number the notice omits. You cannot compute your own warning period from the page that is warning you.

What does the contract actually promise?

Less than nothing, and often less than that. Of the twelve independent vendors in my sample, only four still serve a Terms of Service at a working URL. Clockwise and Mindy now return the identical goodbye page at every path on the domain, including /terms: I fetched /terms and a deliberately nonsensical path and got byte-identical 61,452-byte and 10,730-byte responses. Delibr, Youper and Agency 404 on /terms. The contract is deleted before the product is.

Of the seven contracts I could read, zero name a number of days of notice before discontinuing a service. The numbers those documents do contain are cure periods for breach and price-change windows, which protect a different thing. OpenAI's terms are the best of them and still unquantified: "We may decide to discontinue our Services, but if we do, we will give you advance notice and a refund for any prepaid, unused Services." Advance notice of what length is left to OpenAI. Anthropic's consumer terms go the other way and reserve the right to "modify, suspend, or discontinue the Services or your access to the Services, in whole or in part, at any time without notice." GitHub's reserve the right "at any time and from time to time to modify or discontinue."

The one real commitment I found is not in a contract at all. Anthropic's model-deprecation docs promise "at least 60 days' notice before model retirement for publicly released models," and the Opus 4.1 row proves it: deprecated June 5, retired August 5, 61 days. The same company's terms say it owes you nothing. The promise that held was the one in the documentation.

Does the graveyard data back this up?

Partly, and it has to be used carefully. ToolDirectory's graveyard report covers 208 entries as of July 2026 and finds that "51% of outright deaths were silent," with a lapsed domain the single most common ending. That matches my three zeros. But the authors say plainly that they "maintain the dataset by hand" and that it is "a large hand-maintained sample, not a census of every AI startup," so 51% is a property of their sample, not of the industry, and I am not going to present it as one.

Registries are leads, not evidence. I checked each one at the vendor's own domain, which is how I caught that tinq.ai, listed as a 2026 death, actually 301s to trychirp.com. It got renamed, not killed. Kumo.ai, listed as acquired by NVIDIA, does redirect: straight to docs.nvidia.com/sdgm.

What does this not prove?

Eleven windows is a small n, and it is not a random sample. I picked tools that were documented enough to time, which biases toward vendors organized enough to publish something. The true median across all AI shutdowns is almost certainly lower than 15 days, because the ones with no page at all cannot be sampled by anyone, including me. Nine of my twenty had no computable window, and I am reporting them as unmeasurable rather than as zeros, which is the conservative choice.

The other honest note: a short window is not the same as bad behavior. Clockwise gave eight days and prorated refunds and a migration partner. Agency gave twenty-six days and no export path. Longer is not automatically better, and the export procedure matters more than the calendar.

What follows from this is not "avoid AI vendors." It is that the notice period is not a thing you are owed, so it is not a thing to plan around. Your leverage is entirely upstream: does the product have an export you can run yourself today, without asking anyone. Hiro's was on a settings page. Delibr's produces "a single JSON file containing your teams, collections, and all documents with their full content and comments." Agency's does not exist. That difference was visible on the pricing page long before the goodbye letter, which is the subject of the companion checklist to this piece. If you are counting vendor risk across a stack, the body count from auditing eight outreach tools is the same exercise pointed at claims instead of calendars, and the consolidation math explains why the acquihire is now the default ending.