When an AI lab buys the app you depend on

An app you open every day puts a banner across the top: it has been acquired by an AI lab you have heard of. The announcement is short and warm and tells you that you can keep using the product just as before. Nothing breaks that day. Nothing breaks that month either. When something does change, it arrives as a new price, a new effective date on the terms, or a menu item that is quietly no longer there, and by then nobody is writing announcements about it.

This guide is about reading that banner correctly and doing the small amount of work it justifies, which is more than nothing and much less than panic. It is written for solo operators, freelancers and small teams who run on a handful of AI tools and have no leverage over any of them. It is not for people sizing up the deal as investors. It is also not much use if you hold a negotiated enterprise contract with a named account manager and a lawyer, because you already have protections in writing that self-serve customers do not.

What the acquirer is actually buying

The announcement usually tells you, if you read it as an itemised receipt rather than a press release. Cognition, whose AI software engineer Devin is “used by some of the world’s top companies for software engineering”, announced on 23 July 2026 that it had acquired The Interaction Company of California, maker of Poke, “a personal agent that lives in your texts” [8]. The post makes three points about what it got. People had exchanged “more than 100 million messages with it in just the last three months”. Poke was “the only AI agent approved to text natively on Apple Messages”. And “soon, Cognition’s models and infrastructure will make Poke even faster and more reliable” [8].

Two of those three are about a channel and a platform permission. The operative word in the second is “approved”: native access to a messaging platform is granted by the platform, not built by the team that wants it [8]. The third says what the new owner intends to change, which is the model and the infrastructure underneath rather than the interface on top. Read as a receipt, the post itemises the things the acquirer did not already have: a consumer surface, an audience standing on it, and a permission it could not issue to itself.

So the useful question when the next one lands is not why a lab wants an app in that category. It is which capability the acquired team has that the acquirer’s core product was missing, because that capability is the part with a reason to survive. If what was bought was the team, the product is a souvenir. If it was the audience, the product lasts about as long as the audience does. If it was a platform relationship, the product lasts as long as the relationship needs a product to exist. None of that is a forecast you can bank on. It is the only part of an announcement that carries information about intent rather than tone.

Your terms already said yes on your behalf

Nobody will ask your permission, because you gave it when you signed up. OpenAI’s Terms of Use, effective 1 January 2026, state that “You may not assign or transfer any rights or obligations under these Terms and any attempt to do so will be void”, while reserving OpenAI’s own right to “assign our rights or obligations under these Terms to any affiliate, subsidiary, or successor in interest of any business associated with our Services” [1]. Anthropic’s Consumer Terms, effective 8 October 2025, put it more plainly still: “These Terms may not be transferred or assigned by you without our prior written consent, but may be assigned by us without restriction” [2].

The privacy policies say the same about your data. OpenAI’s, updated 6 February 2026, says that in a strategic transaction, reorganisation or bankruptcy your personal data “may be disclosed in the diligence process with counterparties and others assisting with the Transaction and transferred to a successor or affiliate as part of that Transaction along with other assets” [3]. Anthropic’s, effective 8 July 2026, says that where there is “a merger, corporate transaction, bankruptcy, or other situation involving the transfer of business assets”, Anthropic “will receive or disclose your personal data as part of these corporate transactions” [4]. These are not sharp practice. A company that could not transfer its customer relationships would be unsellable, which is why both policies reserve the right in nearly the same words [3][4].

The consequence is worth being clear about. The acquisition itself is not the event. Your account moved, your data moved and your contract moved, and none of it needed your input or gave you a say. The events worth reacting to are the decisions the new owner makes afterwards, and those surface in three predictable places: the pricing page, the effective date on the terms, and the deprecation notice.

The deprecation calendar is the only clock that binds

Neither vendor promises to keep anything alive. Anthropic reserves 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 to you” [2], and OpenAI says it “may decide to discontinue our Services” [1]. What both do publish is how much warning you get before a specific model is switched off, and those numbers are the most durable facts in this subject. OpenAI commits to at least 6 months’ notice before retiring a generally available model, at least 3 months for specialised variants such as chat or Codex variants, and warns that preview models “may be retired with much shorter notice, such as 2 weeks” [5]. Anthropic gives “at least 60 days’ notice before model retirement for publicly released models”, and defines a retired model as one where “requests to retired models will fail” [6].

Watch how those commitments cash out. Anthropic deprecated claude-opus-4-1-20250805 on 5 June 2026 and retired it on 5 August 2026, two months later [6]. OpenAI notified developers of the Assistants API deprecation on 26 August 2025 and removed it from the API on 26 August 2026, a full year [5]. Both dates cleared the published minimum. Both were also set unilaterally, on the vendor’s schedule, and neither had anything to do with an acquisition. This is the ordinary metabolism of the industry, and an acquisition mostly changes who is holding the calendar.

Which means you should take “Poke users can continue using the product just as before” completely literally [8]. It is an accurate statement about the present with roughly one notice period of implied runway behind it. And if the thing you depend on is a consumer app rather than an API model, you may get less than that, because the pages carrying those numbers cover models and API endpoints rather than features, apps and integrations [5][6].

Export in the week you read the announcement, not the week of the shutdown

The export is slower than the click suggests, and the gap is where the damage happens. In ChatGPT you sign in, open the profile menu, select Settings, then Data controls, then Export, then Confirm export [7]. OpenAI’s own help page says exports “can take up to 7 days to arrive” and that “the download link expires 24 hours after you receive it” [7]. It also notes that the feature is “not available for ChatGPT Business, Enterprise, or ChatGPT for Healthcare workspaces”, and that “an export cannot restore chats that have already been deleted” [7].

Put those two numbers side by side. From the click to a file on your disk, the worst case is 7 days plus a 24-hour window you have to be awake and paying attention for. That is comfortable on a quiet Tuesday in September and a bad bet in the last fortnight of a shutdown notice. Run one now, unzip it, and open the conversations you would actually be upset to lose. An export you have never opened is a belief rather than a backup.

Then do the equivalent for everything else the tool touches. Export the Notion pages you generated into it. Screenshot or copy out the configuration of the Zapier automations that call it, since one tool’s data export will not contain another tool’s settings. Write down which accounts are connected through OAuth. None of this takes an afternoon, and all of it is the same work you would have to do later under time pressure.

Price your exit before somebody prices it for you

Prompts are the portable part. A prompt is plain text you already hold a copy of, and it moves wherever you do. The lock-in lives everywhere else: the automations, the integrations, the saved views, the connected accounts, the naming conventions your collaborators learned, the exports other systems expect in a particular shape.

Counting those is the whole exercise. Write down every distinct thing you would have to rebuild if the tool disappeared, estimate the hours honestly, and multiply by what an hour of your time is worth. The number lands in one of two places. Small enough to stop worrying about, which is a genuine result. Or large enough that the mistake was single-sourcing that capability rather than anything the acquirer has done, which is also a genuine result. The number is not there to make you switch. It is there so that when the shutdown email arrives you already know whether it is a Saturday morning or a fortnight of lost billable work.

The signals worth watching for the next twelve months

Four things tell you where the new owner is heading, and all four are public. The pricing page tells you whether the product is being repositioned, folded into a parent plan, or quietly starved. The effective date on the terms and the privacy policy tells you when the paperwork was last rewritten, and you can only notice a change if you know the date you started from: OpenAI’s terms currently read 1 January 2026 and Anthropic’s 8 October 2025 [1][2]. The deprecation or lifecycle page tells you whether the acquired product now sits inside the acquirer’s formal notice regime, which is better news than sitting outside it [5][6]. And whether the acquired team’s work shows up under the parent brand tells you what the acquirer actually bought, which was the question from the first section.

Read the refund position now, too, while nothing is happening, because vendors differ and the difference is real money. OpenAI’s terms say that if it discontinues the Services it “will give you advance notice and a refund for any prepaid, unused Services”, and promise “at least 30 days advance notice of changes to these Terms that materially adversely impact you” [1]. Anthropic’s consumer terms say that “except as expressly provided in these Terms or where required by law, all payments are non-refundable”, but that where Anthropic terminates a subscription for a reason other than your breach and you bought it through the website, it “will refund you, on a pro rata basis, the fees you paid for the remaining portion of your Subscription after termination” [2]. Both positions are normal and neither is a scandal. They are not the same, and if you are twelve months prepaid on an annual plan, the difference is the size of your annual plan.

What still goes wrong

The notice periods you can cite are for the wrong object. OpenAI’s 6 months and Anthropic’s 60 days apply to models and API endpoints [5][6]. The consumer app that just got acquired may have no published lifecycle page at all, in which case the reassuring numbers in this guide describe a neighbouring category and not the thing you are worried about. Anthropic’s own consumer terms reserve the right to discontinue the Services “at any time without notice to you”, and that clause sits alongside the 60-day model policy rather than under it [2][6].

Exports return content, not capability. A ZIP of conversations does not reproduce the integrations, the connected accounts, the model behind the product or the permissions it held on other platforms. Poke’s value in the announcement was partly a platform approval [8], and no export contains one of those. If what you valued was the way the tool reached into your calendar or your messages, the export gives you a transcript of a thing that no longer works. That is worth having and it is not a migration path.

And overreacting has a price. Moving early costs the same hours as moving late while buying you less information, because the decisions that would inform the move have not been made yet. The honest position is that an announcement cannot tell you whether this deal kills the product. It can tell you which capability was bought, and therefore which part the new owner has a reason to protect. Do the export, count the rebuild hours, set the reminder, and then go back to work.