How to Move Your Work From One AI Tool to Another Without Losing It
Six phases, in order — plus the legal right that forces an export when there's no export button.
Most people migrating between AI tools export the wrong thing.
They download their outputs — the images, the generated text, the finished files — feel prepared, and then discover the painful part months later. Because outputs are the cheapest thing to replace. What actually took you time was the inputs: the prompts you refined over dozens of attempts, the knowledge of which phrasings work, the test cases you use to check quality.
One analysis of production migrations found 40 to 70% of the total effort sat in data preparation and rebuilding evaluations — not in the connection change, which takes an afternoon.
So here's the playbook in the order it actually needs doing, including what to do when a tool has no export button and you're inside a shutdown window.
📋 The Six Phases
Phase 1 — Inventory: What You Actually Have
Before you touch an export button, write down what's in the tool. Four categories, and people usually only think of the first.
| Category | Examples | Replaceable? |
|---|---|---|
| Outputs | Generated images, audio, drafts | Easily — regenerate them |
| Inputs | Prompts, uploaded documents, settings | Expensively — weeks of work |
| Wiring | Automations, webhooks, integrations | Must be rebuilt entirely |
| Knowledge | What works, what breaks, workarounds | Only if you wrote it down |
That last row is the one that hurts. If your knowledge of the tool lives only in your head, you're about to spend weeks rebuilding it against a new tool while doing real work. Spend an hour now writing it down — the failure modes, the prompt patterns that fixed them, the settings you changed and why.
Phase 2 — Extract: Getting It Out
Three rules, learned from watching this go wrong.
Export the source, not just the processed version
If you uploaded documents to a chatbot, you need the original documents — not the tool's indexed version of them, which won't transfer anyway. Anything the tool processed gets rebuilt; anything you provided can move.
Download files, never keep URLs
If your work references files by link rather than storing them, you have nothing. Signed URLs from any provider expire, and they certainly won't outlive the service. Pull the actual files down.
Treat announcement dates as deadlines, not shutdown dates
This is where people get burned. Notice periods cover access, not always data. One developer tool shut down abruptly in January 2026 and deleted user data two weeks later. Another vendor's final export window was described conditionally — it said it would notify users if one were offered.
Never plan around a grace period you've been promised only in the conditional tense.
⚡ And If There Simply Isn't an Export Button?
Plenty of tools don't have one. Some make it deliberately awkward.
Depending on where you are, you may have a legal right to compel one — in a machine-readable format, for free.
Most migration guides never mention this.
Phase 3 — When There's No Export Button
Two routes, one legal and one technical.
The legal route
If you're in the EU or UK, GDPR Article 20 gives you the right to receive personal data you provided in a structured, commonly used and machine-readable format, and to transmit it to another provider without hindrance from the original one. The provision exists specifically to break lock-in.
Details worth knowing when you make the request:
- The first request is free. Providers may charge a reasonable fee only for additional copies or genuinely excessive requests.
- Format standards are enforceable. A PDF scan doesn't comply because it isn't machine-readable. A proprietary format requiring their software doesn't comply because it isn't commonly used. An unlabelled text dump doesn't comply because it isn't structured. CSV or JSON is the expected answer.
- You can ask for direct transfer. Where technically feasible, you have the right to have data sent from one provider straight to another rather than downloading and re-uploading it yourself.
Separately, the EU Data Act places switching obligations on cloud and SaaS providers — including removing contractual barriers to migration, meaning no exit fees or penalties for switching, and enabling export in interoperable formats.
⚠️ Three real limits on that right.
It applies only where processing rests on consent or a contract, not other legal bases. It covers only data you provided — typed in or observed — and not anything the provider derived or inferred. A useful rule of thumb: if you typed it, it's portable; if their code produced it, it probably isn't. And these are EU and UK regimes, so if you're elsewhere you may have narrower access rights or none. I'm not a lawyer — for anything commercially significant, get advice for your jurisdiction.
The technical route
Many AI tools are built on standard foundations, which means your data often sits in a standard place even when the interface offers no way out.
When one AI browser shut down, its bookmarks didn't transfer automatically — but because it was Chromium-based, they sat in an ordinary profile folder as readable JSON. Anyone who copied that file before deleting the app kept everything.
The general lesson: before uninstalling a desktop tool, look in your system's application support directory for its folder and copy the whole thing somewhere safe. It costs nothing and it's frequently the only remaining copy.
Phase 4 — Rebuild: The Unavoidable List
Accept upfront that some things cannot be moved and must be recreated. Knowing which ones stops you searching for an export that doesn't exist.
Indexed knowledge. Every document you fed the old tool has to be fed to the new one. The source moves; the processed index doesn't. Budget roughly the same time it took originally.
Prompt calibration. You can paste prompts across immediately. Getting comparable output requires rewriting and testing, and with image or video tools the rewriting is substantial — motion and camera vocabulary are interpreted differently by every model.
All wiring. Every automation, webhook and integration. Note that output formats usually change too, so the downstream steps need adjusting as well as the connection itself.
Your quality instinct. The unglamorous one. You know your current tool's failure modes because you hit them repeatedly. That's rebuilt through use — but a written test set turns weeks of vague re-learning into an afternoon of checking.
Phase 5 — Parallel Run: Never Cut Over Cold
Run both tools simultaneously for at least two weeks. Yes, you're paying twice. That's the cost of not discovering a gap after the old tool is gone.
The parallel period is your only chance to compare against a live reference. Once the old tool is cancelled or shut down, you're comparing new output against memory — and memory flatters whatever you used to have.
Do this during the overlap: put the same real work through both tools. Not test prompts — actual jobs. Then note where they diverge. That divergence list is what you'll spend the following month fixing, and finding it now is much cheaper than finding it in front of a client.
If you're migrating because of a shutdown, this window closes on a fixed date. Start it as early as the announcement, not as the deadline approaches.
Phase 6 — Verify and Clean Up
The phase everybody skips, and the one that causes the mysterious failure three months later.
- Open your exports. Actually open them. An archive that downloaded successfully but contains an error page is a real and common outcome.
- Search everywhere for the old tool's name. Automations, config files, saved settings — and any database or spreadsheet where you stored a model name or ID as text. Code searches miss those, and they fail silently later.
- Revoke access. Remove the old tool's permissions from your accounts, and delete API keys. An abandoned integration with live credentials is a security problem, not just clutter.
- Cancel deliberately. Confirm the cancellation went through and note the date. Don't assume.
- Write the migration note. Half a page: what you moved, what you rebuilt, what surprised you. Your next migration will be much faster for it.
One habit worth adopting permanently while you're in here: route future automations through one intermediate step rather than calling a tool directly. Then the next change is a config edit rather than a rebuild.
Frequently Asked Questions
Can I force an AI tool to give me my data?
In the EU and UK, GDPR Article 20 gives you the right to receive personal data you provided in a structured, commonly used and machine-readable format, and to transmit it elsewhere without hindrance. The first request is free. It applies only where processing rests on consent or contract, and only to data you provided rather than anything the provider derived. Rights differ elsewhere — check your jurisdiction.
Can a provider charge me an exit fee for switching?
The EU Data Act places switching obligations on cloud and SaaS providers, including removing contractual barriers to migration such as exit fees or penalties, and enabling export in interoperable formats. If you're outside that regime, check your contract before assuming.
What should I export first if a tool is shutting down?
Your inputs, not your outputs. Prompts, uploaded source documents, and anything you configured. Outputs can usually be regenerated; prompts refined over months cannot. Then download actual files rather than keeping links, since URLs expire and won't outlive the service.
How long should I run both tools in parallel?
At least two weeks, putting real work through both rather than test prompts. Once the old tool is gone you're comparing against memory, which flatters what you used to have. If a shutdown date is forcing the migration, start the overlap at the announcement rather than near the deadline.
What can't be migrated at all?
Indexed knowledge has to be rebuilt from source. Prompt calibration has to be re-earned through testing. Every automation and integration has to be rewired, including downstream steps since output formats change. And your instinct for where a tool fails is rebuilt through use, though a written test set shortens that considerably.
The Takeaway
Migrations go wrong in predictable ways. People export outputs and lose inputs. They trust a grace period described in the conditional. They cut over cold and discover the gap when the reference is gone.
The sequence that works: inventory all four categories, extract your inputs and source files, use the legal route if there's no export button, accept the rebuild list, run parallel on real work, then verify and clean up properly.
And write down what you learned. Migrations get cheap once you've built portability into how you work — but only if you treat the first one as the investment rather than the emergency.
🚀 Stay Connected With Simple AI Tools
Practical AI guides — the playbooks, not the hype.
👇 💬 Drop your comment below and let us know your thoughts! ✨
Comments
Post a Comment