What If Your AI Vendor Shuts Down? Checks Before You Commit

It depends on the vendor's terms and how it closes. An orderly shutdown usually gives you an export window before data is deleted; a sudden one may leave nothing: Clockwise deleted user data when it shut down on 27 March 2026. Assume you'll get weeks at most, so keep your own exports and a replacement plan from day one.

Shutdowns come in more shapes than a company closing. OpenAI discontinued Sora in two stages, closing the app on 26 April 2026 and the API on 24 September 2026. Shopify's Stocky inventory app stopped working on 31 August 2026, with purchase orders moving into Shopify's own admin. Microsoft Lens left the app stores on 9 February 2026 and stopped scanning after 9 March. Even when the company survives, the product you built a process around can end, so the checks below cover products and features as well as vendors.

Follow me on Instagram@sagnikteaches

Recent AI shutdowns and retirements, from abrupt to phased

The pattern is clear enough to plan for. Big platforms retire products and features on a schedule, usually with notice and a migration path. Small vendors can close abruptly, and what happens to your data then depends on terms you agreed at sign-up and on whether anyone is left to honour them.

Connect on LinkedInSagnik Bhattacharya

Before you sign: seven checks

1. A working export. Why: an export you've never run is a hope, not a plan. Verify: run one during the trial, open the files, and count the records against what's in the tool. Note what's excluded. Claude's organisation export, for instance, leaves out deleted messages, files and projects, so project files need keeping elsewhere.
2. Data return and deletion terms. Why: they decide your window. Verify: find the termination section of the terms and the data processing agreement, and note the days. For comparison, Microsoft says data left behind after a business subscription is cancelled might be deleted after 30 days and is deleted no later than 180 days.
3. Notice of discontinuation. Why: weeks of notice turn a crisis into a project. Verify: look for a notice period for ending the service or removing features. If there's none, ask for one in writing.
4. Configuration you own. Why: prompts, workflows and knowledge files are the hard part to rebuild. Verify: confirm you can view and copy every instruction, automation and uploaded file, and that API keys and automation accounts sit in your name.
5. Dependencies disclosed. Why: a vendor built on another company's model inherits that company's retirements, as Sora's API closure showed. Verify: read the sub-processor list and ask what the vendor does when a provider retires a model.
6. Signs of staying power. Why: you're betting on the vendor for as long as you use it. Verify: look for a public pricing page with ordinary subscriptions, a changelog with recent releases, named people running the company and quick answers from support during the trial.
7. A named replacement. Why: knowing where you'd go halves the time to get there. Verify: write down one alternative, even if it's a manual process, and a rough estimate of the time to switch.

For a custom-built tool rather than a subscription, add an eighth: agree who holds the code and the documentation if the supplier disappears. What to check in an AI vendor's data processing agreement covers the contract clauses in more depth.

Subscribe on YouTube@codingliquids

If the terms say nothing about discontinuation, a short clause is worth requesting. An example of wording to discuss with the vendor and, for anything significant, your legal adviser:

Discontinuation. The Supplier will give the Customer at least 60 days'
written notice before ending the Service or removing a feature the
Customer uses. During that period the Supplier will provide a complete
export of Customer Data in CSV or JSON format within 14 days of a
request, and will confirm deletion of Customer Data in writing after
the Customer has confirmed receipt of the export.

Small vendors often agree to the notice period and push back on the rest. Even the notice alone is worth having.

While you use it: keep a copy that doesn't depend on the vendor

The single most useful habit is a scheduled export into storage you control. Here's how the two assistants most small firms use handle it:

• ChatGPT: on individual plans, export from Settings, Data Controls, Export data; the download link arrives by email and expires 24 hours after you receive it. On Business and Enterprise workspaces, self-service export isn't available, so ask your workspace owner how organisation-managed data access works before you need it.
• Claude: individuals export from Settings, Privacy. On Team and Enterprise plans only the Primary Owner can export organisation data, from the organisation settings under Data and privacy, and Claude's help page on organisation exports lists what's left out.

Beyond exports, keep the things that are expensive to recreate in your own files: the instructions behind each assistant or automation, the reference documents it uses, a one-page diagram of what connects to what, and a list of every account and API key involved. Keeping your data and prompts portable goes further on this, and the backup checklist before connecting AI tools covers the underlying business data.

A copy register makes the habit visible. The storage facility in the worked example below keeps one like this:

The "not yet" is honest and useful: it's the next thing to test.

Partial shutdowns: the feature ends, the vendor stays

Most small firms meet vendor risk in this quieter form. The company is fine and your data is safe, but something a process depends on stops working. Because nothing dramatic happens, the break can go unnoticed for weeks.

• Excel's =COPILOT() function was retired on 14 September 2026. Existing results stay in the cells, but no new formulas can be created. A monthly report template that added a row of =COPILOT() formulas each month kept last month's answers and silently stopped producing new ones. Microsoft points users to the Copilot side pane instead.
• Hubdoc ended automatic bill collection from supplier portals in 2022. Firms that relied on it went back to downloading PDFs by hand, and some only noticed when supplier statements showed missing bills.
• ChatGPT agent was retired on 9 July 2026 and replaced by ChatGPT Work at existing plan prices. The replacement is more capable, but any written procedure that told staff to "use agent mode" needed rewriting.
• Amazon Lookout for Vision, an image-inspection service, ended on 31 October 2025, so anyone using it for quality checks had to move their inspection models elsewhere.

The defence is the same register, with one more column: for each tool, list the specific features your processes use, not just the product name. When a notice mentions one of those features, you'll recognise it.

What to watch for after signing

Retirements are usually announced before they bite, in emails that look like routine product news. Assign one person to read vendor notices and set a mail rule that flags words such as retire, discontinue, sunset, end of life, migration and "changes to your plan". Things worth acting on:

• End-of-life dates for a product, an API or a model the vendor depends on.
• Feature freezes, such as HubSpot's move to stop new installs of five agents while old installs carried on.
• Renames and replacements, such as ChatGPT agent becoming ChatGPT Work in July 2026, which usually mean retesting.
• Pricing restructures and acquisitions, which often come before product changes.

When a long notice arrives, an assistant can pull out what matters. Paste the notice and the relevant section of the terms, and ask:

Below is a notice from a software vendor and the termination section of
its terms. List: what is ending, the exact dates, what data we can
export and in what format, how long data is kept after the end date,
any action we must take and by when. Quote the source sentence for each
item. If something isn't stated, write "not stated".

For the storage facility's vendor, the output looked like this (illustrative):

Ending: legacy booking widget (v1)            "v1 will be retired on..."
Date: 30 November                               quoted
Export: conversation history, CSV               quoted
Retention after end: not stated
Action: switch to widget v2 before 30 November  quoted

The reply missed one thing: the terms allowed the vendor to give notice "by posting on our website", which meant the next notice might never arrive by email. That line went into the register as a monthly check of the vendor's changelog page.

Exit day: a ten-day runbook

1. Day 1: confirm the end date and what ends; freeze changes to the tool; tell staff and name the person in charge.
2. Days 1-2: run a full export, open the files and count records against the live system. Export twice, to two places.
3. Days 2-3: switch on the fallback, even if it's manual, so customers aren't affected while you rebuild.
4. Days 4-6: set up the replacement from your own documented instructions and reference files, and import the data.
5. Day 7: run old and new side by side on real work where the old tool still runs.
6. Day 8: move integrations, website widgets and inbox rules to the replacement.
7. Day 9: revoke the old tool's access to your email, calendars and systems; request written confirmation of data deletion.
8. Day 10: review what was missing from your copies and update the register.

The checklist on a storage facility's booking assistant

Here's an illustrative self-storage business with two sites that uses a small vendor's AI booking assistant on its website. It handles about 600 conversations a month, holds about 90 open quotes at any time, and keeps a waitlist of 140 people for sold-out unit sizes. A Zapier workflow passes confirmed bookings into the facility's storage management software.

The owner ran the seven checks during the trial:

The waitlist was the find that mattered. Without the Zapier copy, a sudden closure would have lost 140 warm leads for the facility's scarcest units. With it, the owner's estimate of switching time fell from about two weeks of reconstructing records to about two days.

The owner also put a rough price on a sudden closure, to decide how much effort the checks deserved. Rebuilding without documents: about 40 hours of staff time at an illustrative $30 an hour, $1,200, plus two weeks of enquiries handled only by a contact form. With the documents and exports: about 12 hours, $360, and two days on the fallback. The difference, roughly $840 in time plus ten days of weaker enquiry handling, paid for the half-day the checks took several times over. Money prepaid to the vendor was the smallest line, because the facility paid monthly.

Big vendors need the same thinking, just for different reasons. A courier firm had its drivers photograph signed delivery notes with Microsoft Lens. When scanning stopped after 9 March 2026, the fix, OneDrive's scan feature, took twenty minutes to show the drivers, but only because the office manager had spotted the retirement notice a month earlier. Firms that hadn't were collecting paper again. And anyone still running a custom GPT built in 2024 or 2025 has a hard date: they stop running on 11 December 2026, with OpenAI's migration turning each GPT's instructions into a skill and its knowledge files into reference files inside a plugin.

Running the checks on the tools you rely on today

You don't need to wait for a new purchase. List every AI tool your business uses, mark the ones whose loss would stop work for more than a day, and run checks 1, 4 and 7 on those first: export, owned configuration and a named replacement. Put first the tools that hold customer data, talk to customers directly, or feed other automations, because those are the ones where a closure spreads. For the rest, a monthly export and a line in the register is enough.

Once a quarter, test one copy properly: pick an item from the register and rebuild it from your copy alone, without logging in to the vendor. A copy you've never restored from is as untested as an export you've never run, and the first rebuild always turns up a missing file or an undocumented setting.

Pair the list with your billing choices, because prepaying a year to a young vendor raises the stakes of every item above; annual versus monthly billing covers that trade-off, and the lifetime-deal tutorial covers the riskiest version of it.

Vendor shutdown questions

Can I get my money back if an AI vendor shuts down?

Sometimes, but don't count on it. A company closing because it has run out of money may have none left to refund prepaid fees, and some closures happen with little notice. Limit the risk by paying monthly for young vendors, avoiding long prepayments, and keeping the receipts and contract so you can make a claim if there's a formal process for creditors.

Does data-protection law make a vendor return or delete my data?

Where a vendor processes personal data for you, data-protection law such as the GDPR generally expects the contract to say whether that data is returned or deleted when the service ends. Check your data processing agreement for those terms and the timescale. It won't guarantee an export if a company collapses, so ask your data-protection adviser about anything sensitive.

Is source-code escrow worth it for a small business?

Only for bespoke software that runs a critical part of your business and that someone else could maintain if the supplier failed. Escrow means a third party holds the code and releases it to you under agreed conditions. For off-the-shelf AI subscriptions, regular data exports and a written exit plan protect you far more cheaply than escrow would.

Should small businesses avoid AI startups altogether?

No. Young vendors often build the most useful tools for small firms. Manage the risk instead: pay monthly, start with jobs that wouldn't hurt to lose, test the export before relying on the tool, keep your prompts and settings in your own files, and name a replacement. Those habits make a startup's closure an inconvenience rather than a crisis.

Further reads

• AI Software Contracts: Auto-Renewals, Price Rises, Notice Periods — Notice periods and renewal terms to negotiate up front.
• How to Evaluate an AI Software Vendor: A Small Business Scorecard — Score a vendor's stability alongside its features.
• What Uptime and Support Should an AI Vendor Promise You? — The uptime and support terms to ask for.
• Should a Custom GPT Answer Your Customers? Limits and Safer Options — Safer options now that custom GPTs are retiring.
• Who Owns the AI Workflows a Consultant Builds for You? — Make sure workflows built for you are yours to move.
• AI Consultant Handover Checklist: What You Need Before They Leave — What to collect before an outside builder leaves.
• Questions to Ask an AI Vendor Before You Sign Anything — The questions to put to an AI vendor before, during and after the demo, with what good and walk-away answers sound like and a scoring sheet.
• What to Do Before You Buy Any AI Tool: A 10-Point Checklist — Ten checks to run before paying for any AI tool, from pricing the job it replaces to reading the exit terms, with a kitchen-fitter worked example.
• Salon and Spa Software With Built-In AI: What to Compare — Five tests that separate useful salon and spa AI from a feature list, with September 2026 prices for Vagaro, GlossGenius, Boulevard, Mangomint and more.
• Are AI Tools Built for Coaches Worth Paying For Over ChatGPT? — What coaching platforms, AI clones and note-takers add over a $20 ChatGPT plan, a year of each costed, and the platform-closure risk to plan for.
• Questions to Ask Before Buying AI That Touches Client Data — Nineteen questions to put to any AI vendor before client files go in, with what a good answer looks like and the replies that should stop a purchase.
• AI Governance Checklist for Small Financial Advice Firms — Eight groups of checks, each with the evidence to keep, so a small advice firm can show how AI is approved, supervised and recorded.
• Should Coaches Build an AI Version of Their Method for Clients? — When an AI version of a coaching method helps clients and when it undermines you, what to build it on now custom GPTs are retiring, and the guardrails it needs.
• Choosing Nursery Software With AI: What to Compare — What to compare when nursery software offers AI: children's data handling, role controls, a demo test with one observation, and a scored choice.
• AI Video Survey Tools for Movers Compared: Cost and Accuracy (2026) — Six video survey tools for removal firms compared on published price, pricing model and what their accuracy claims actually measure.
• AI Note-Taking Tools for Therapists: A Confidentiality Checklist — A confidentiality checklist for therapists choosing an AI note-taker, with what SimplePractice, Upheal, Mentalyc and Heidi say they do with session data.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: vendor notices on the Clockwise shutdown, the Excel =COPILOT() retirement (Microsoft Support), ChatGPT agent's replacement by ChatGPT Work, the Sora discontinuation (OpenAI help centre), Shopify Stocky, Microsoft Lens, Kajabi Creator Studio and HubSpot's retired Breeze agents; OpenAI help centre on custom GPT retirement and on exporting ChatGPT data; Claude help centre on exporting organisation data; Microsoft Learn on cancelling business subscriptions. Checked September 2026.